Biomni: A General-Purpose Biomedical AI Agent

Kexin Huang1,∗,‡ , Serena Zhang1,∗ , Hanchen Wang1,2,∗ , Yuanhao Qu3,4,5,∗ , Yingzhou Lu5,∗ , Yusuf
Roohani1,6 , Ryan Li1 , Lin Qiu7 , Gavin Li1 , Junze Zhang3,5 , Di Yin3,5 , Shruti Marwaha8 , Jennefer
N. Carter8 , Xin Zhou5 , Matthew Wheeler8 , Jonathan A. Bernstein9 , Mengdi Wang10 , Peng He11 ,
Jingtian Zhou6 , Michael Snyder5 , Le Cong3,5 , Aviv Regev2 , and Jure Leskovec1,‡
1
Department of Computer Science, Stanford University School of Engineering, Stanford, CA, USA
2
Research and Early Development, Genentech, South San Francisco, CA, USA
3
Department of Pathology, Stanford University School of Medicine, Stanford, CA, USA
4
Cancer Biology Program, Stanford University School of Medicine, Stanford, CA, USA
5
Department of Genetics, Stanford University School of Medicine, Stanford, CA, USA
6
Arc Institute, Palo Alto, CA, USA
7
Paul G. Allen School of Computer Science and Engineering, University of Washington, WA, USA
8
Department of Medicine, Stanford University School of Medicine, Stanford, CA, USA
9
Department of Pediatrics, Stanford University School of Medicine, Stanford, CA, USA
10
Department of Electrical and Computer Engineering, Princeton University, Princeton, NJ, USA
11
Department of Pathology, University of California San Francisco, San Francisco, CA, USA

∗Equal contribution.
‡Corresponding authors. Email: [email protected], [email protected]

1
 1 Abstract
2 Biomedical research underpins progress in our understanding of human health and disease,
3 drug discovery, and clinical care. However, with the growth of complex lab experiments,
4 large datasets, many analytical tools, and expansive literature, biomedical research is in-
5 creasingly constrained by repetitive and fragmented workflows that slow discovery and limit
6 innovation, underscoring the need for a fundamentally new way to scale scientific exper-
7 tise. Here, we introduce Biomni, a general-purpose biomedical AI agent designed to au-
8 tonomously execute a wide spectrum of research tasks across diverse biomedical subfields.
9 To systematically map the biomedical action space, Biomni first employs an action discov-
10 ery agent to create the first unified agentic environment – mining essential tools, databases,
11 and protocols from tens of thousands of publications across 25 biomedical domains. Built
12 on this foundation, Biomni features a generalist agentic architecture that integrates large
13 language model (LLM) reasoning with retrieval-augmented planning and code-based execu-
14 tion, enabling it to dynamically compose and carry out complex biomedical workflows – en-
15 tirely without relying on predefined templates or rigid task flows. Systematic benchmarking
16 demonstrates that Biomni achieves strong generalization across heterogeneous biomedical
17 tasks – including causal gene prioritization, drug repurposing, rare disease diagnosis, micro-
18 biome analysis, and molecular cloning – without any task-specific prompt tuning. Real-world
19 case studies further showcase Biomni’s ability to interpret complex, multi-modal biomedical
20 datasets and autonomously generate experimentally testable protocols. Biomni envisions a
21 future where virtual AI biologists operate alongside and augment human scientists to dra-
22 matically enhance research productivity, clinical insight, and healthcare. Biomni is ready to
23 use at https://biomni.stanford.edu, and we invite scientists to explore its capabilities, stress-
24 test its limits, and co-create the next era of biomedical discoveries.

2
25 1 Introduction
26 Biomedical research is a key pillar of modern science and medicine, driving discoveries in disease
27 mechanisms, diagnostics, and therapeutics1–4 . Yet, with the growth in large-scale experiments,
28 data, tools, and literature, progress is increasingly slowed by fragmented, complex workflows that
29 require specialized tools, exhaustive literature reviews, intricate experimental design, and careful
30 statistical modeling5, 6 . A vast volume of valuable biomedical data sits underutilized7 , many so-
31 phisticated analyses are not conducted, and many connections for past knowledge and literature are
32 not made, not for lack of significance, but because the demand for expert researchers far exceeds
33 the supply. This mismatch between data abundance and limited human bandwidth highlights an
34 urgent need for a fundamentally new approach – one that can effectively scale expertise, streamline
35 workflows, and unlock the full potential of biomedical research.
36 Recent advances in Artificial Intelligence (AI) have created a paradigm shift, opening the
37 possibility for fundamentally reshaping biomedical research8 . AI agents have dramatically re-
38 shaped fields such as software engineering9 , law10 , material science11 and healthcare12 by au-
39 tomating repetitive tasks, enhancing productivity, and enabling breakthroughs that were previ-
40 ously unimaginable. Given these developments, the question emerges: Can we build a virtual
41 AI biomedical scientist? Such a virtual scientist would autonomously tackle diverse biomedical
42 research tasks spanning multiple subfields, unlocking extensive capabilities and fostering novel
43 insights through interdisciplinary integration – an achievement that can radically augment human
44 biologists limited by specialized expertise. Capable of efficiently managing thousands of concur-
45 rent tasks, this virtual scientist could dramatically enhance human productivity and accelerate the
46 pace of biomedical discovery.
47 Previous approaches have largely relied on specialist agentic workflows tailored to nar-
48 row biomedical tasks13–19 , which restricts their capacity to move fluidly and generalize across
49 the full spectrum of biomedical domains, as needed to answer key research questions. Enabling
50 an AI agent to handle a broad range of biomedical tasks introduces substantial technical chal-
51 lenges – most notably, the need to tightly couple advanced reasoning20 with the ability to exe-
52 cute highly specialized biomedical actions21 . Although LLM-based reasoning has seen significant
53 advancements22 , such LLMs need access to an environment that explicitly defines the biomedical

3
54 action space, which is inherently diverse, domain-specific, and complex. Moreover, a truly capable
55 system requires an agentic architecture that can natively interact with this biomedical environment
56 – autonomously selecting and composing actions, using its reasoning capabilities to plan and exe-
57 cute diverse tasks without relying on rigid, pre-defined workflows.
58 Here we present Biomni, a generalist biomedical AI agent purpose-built to automate and
59 advance biomedical research across a wide range of subfields. Acting as a virtual AI biologist,
60 Biomni autonomously formulates novel, testable hypotheses, performs complex bioinformatics
61 analyses, and designs rigorous experimental protocols. To enable this capability, we first con-
62 structed a unified and comprehensive biomedical action space by systematically analyzing tens
63 of thousands of biomedical research papers spanning 25 distinct subfields, curated from major
64 bio-literature repositories. From this foundation, we developed an LLM-powered action discovery
65 agent capable of reading papers and extracting key tasks, tools, and databases essential to driving
66 biomedical discoveries. These elements are then selected and implemented into Biomni-E1, the
67 foundational environment that defines the biomedical action space for agentic interaction. Biomni-
68 E1 includes 150 specialized biomedical tools, 105 software packages, and 59 databases. We then
69 designed Biomni-A1, a general-purpose agent architecture capable of flexibly executing a broad
70 spectrum of biomedical tasks by using tools and datasets provided by Biomni E1. Given a user
71 query, the agent first uses a retrieval system to identify the most relevant tools, databases, and soft-
72 ware needed. It then applies LLM-based reasoning and domain expertise to generate a detailed,
73 step-by-step plan. Each step is expressed through executable code, enabling precise and flexible
74 compositions of biomedical actions – an essential feature given the domain’s reliance on highly
75 specialized tools and data resources. Unlike traditional function-calling methods, this approach
76 supports the dynamic and complex nature of biomedical workflows. This integrated system allows
77 Biomni not only to solve challenging, large-scale biomedical problems with efficiency, but also to
78 generalize to novel tasks across previously unseen areas of biomedical research.
79 Rigorous benchmarking demonstrates Biomni’s outstanding performance across established
80 biomedical Q&A benchmarks, and robust generalization performance in eight challenging, realis-
81 tic scenarios never encountered during development. Additionally, we highlight Biomni’s practical
82 capabilities through three impactful case studies: (1) analyzing 458 files of wearable sensor data
83 to generate novel insights; (2) rapidly performing comprehensive bioinformatics analyses on mas-

4
84 sive raw datasets, such as single-cell RNA-seq and ATAC-seq data, to generate novel insights and
85 hypotheses; (3) autonomously designing laboratory protocols to assist wet-lab researchers. With
86 Biomni, we introduce the first generation of a scalable, general-purpose biomedical AI agent,
87 setting the stage for an era where virtual AI biologists work alongside human researchers to dra-
88 matically accelerate biomedical discovery from basic research to translation.

89 2 Results
90 Overview of Biomni. Biomni is a general-purpose biomedical AI agent comprising two main
91 components: Biomni-E1, a foundational biomedical environment with a unified action space, and
92 Biomni-A1, an intelligent agent designed to utilize this environment effectively.
93 Curating a unified biomedical action space is challenging due to its inherent complexity and
94 vastness. We systematically address this by employing an AI-driven approach (Figure 1a). Specif-
95 ically, we leveraged the 25 subject categories defined by bioRxiv, selecting the 100 most recent
96 publications per category. An action discovery LLM agent processed each paper sequentially,
97 extracting essential tasks, tools, databases, and software necessary to replicate or generate the de-
98 scribed research. This comprehensive set of resources constitutes the essential actions required to
99 perform a large set of biological research tasks.
100 We then curated Biomni-E1, an environment for a biomedical AI agent to perform a wide
101 range of actions (Figure 1b). Identified tools were rigorously verified by human experts, along
102 with corresponding test cases. These tools (Supplementary Table 1-Table 18) were specifically
103 chosen for their non-trivial nature, encompassing complex code, domain-specific know-how, or
104 specialized AI models. Recognizing the inherent flexibility required by biological software, which
105 cannot always be simplified into static functions, we constructed an execution environment pre-
106 installed with 105 widely-used biological software packages (Supplementary Table 23-30), sup-
107 porting Python, R, and Bash scripts. For database integration, we categorized resources into two
108 distinct groups. The first group consists of massive relational databases accessible via web APIs
109 (e.g., PDB, OpenTarget, ClinVar) (Supplementary Table 19-20). Rather than creating numerous
110 individual retrieval tools, we implemented a unified function per database. Each function accepts
111 natural language queries and internally employs an LLM to parse database schemas and generate
112 executable queries dynamically. Databases without web interfaces were downloaded into a data

5
113 lake and preprocessed locally into structured pandas DataFrames for seamless integration with the
114 agent, for a total of 59 databases in Biomni-E1 (Supplementary Table 21-22). In summary, Biomni-
115 E1 is the first environment for biomedical AI agent and includes 150 specialized biomedical tools,
116 105 software, and 59 databases.
117 To build a general-purpose agent capable of tackling diverse biomedical tasks, we require a
118 specialized agentic architecture – one that avoids hardcoding workflows for each individual task.
119 This led to the development of Biomni-A1, which incorporates several core innovations critical
120 for operating across the biomedical research landscape. First, we introduce an LLM-based tool
121 selection mechanism designed to navigate the complexity and specialization of biomedical tools,
122 dynamically retrieving a tailored subset of resources based on the user’s goal. Second, recogniz-
123 ing that biomedical tasks often require rich procedural logic, Biomni-A1 uses code as a universal
124 action interface – allowing it to compose and execute complex workflows involving loops, par-
125 allelization, and conditional logic. Crucially, this approach also enables the agent to interleave
126 calls to software, tools, databases, and raw data operations that do not conform to predefined func-
127 tion signatures-supporting flexible and dynamic integration of heterogeneous resources. Third, the
128 agent adopts an adaptive planning strategy: it formulates an initial plan grounded in biomedical
129 knowledge and iteratively refines it throughout execution, enabling responsive, context-aware be-
130 havior. Together, these innovations enable Biomni-A1 to generalize to previously unseen tasks and
131 domains, dynamically composing intelligent actions and interfacing with software, data, and tools
132 in a way that embodies generalist biomedical intelligence (Figure 1c).

133 Biomni excels on general biomedical knowledge and reasoning benchmarks. We evaluated
134 Biomni on three challenging multiple-choice benchmarks of general biomedical knowledge and
135 reasoning: Humanity’s Last Exam (HLE)23 and LAB-Bench24 , which includes two key subtasks –
136 DbQA (Database Question Answering) and SeqQA (Sequence Question Answering) (Figure 2a).
137 These tasks span tool use, symbolic reasoning, and structured biological information retrieval
138 – core competencies for any robust biomedical AI agent. To isolate the impact of tool access
139 and agent design, we compared Biomni against six strong baselines (details in Supplementary
140 Notes A).
141 For LAB-Bench, a 45-question development set was used to refine tool and database inter-

6
142 faces, while the final evaluation was conducted on 315 held-out test questions, with performance
143 averaged across three independent runs. We only curated a representative 12.5% subset of the full
144 benchmark due to API cost constraints. In DbQA, which requires structured querying over biolog-
145 ical databases, Biomni achieved 74.4% accuracy – matching expert human performance (74.7%)
146 and outperformed all baselines, including the coding agent (ReAct+Code, 40.8%). In SeqQA,
147 which involves reasoning over DNA and protein sequences, Biomni achieved 81.9% accuracy,
148 again exceeding human-level performance (78.8%).
149 To test true generalization of biomedical knowledge and reasoning without any development
150 set, we also evaluated Biomni on a 52-question subset of HLE spanning 14 biomedical subfields
151 – from molecular biology to physiology. Biomni achieved 17.3% accuracy, significantly outper-
152 forming the base LLM (6.0%), coding agent (12.8%), and literature agent (12.2%). These re-
153 sults demonstrate Biomni’s ability to generalize across unfamiliar, open-ended biomedical domains
154 without any task-specific adaptation. Additional ablation results are shown in Supplementary Fig-
155 ures 1-2. Performances across each subfield are reported in the Supplementary Figure 3.

156 Biomni generalizes to new, real-world biomedical tasks across diverse subfields. To evaluate
157 generalization in realistic research tasks, we curated eight new biomedical benchmarks spanning
158 genetics, genomics, microbiology, pharmacology, and clinical medicine (Figure 2b). Each task
159 was framed to reflect a common, well-defined, but complex real-world biomedical research goal,
160 including: (1) Variant prioritization: Identify the most likely causal variant from a list of poten-
161 tial variants for a trait, requiring reasoning about regulatory functions in non-coding regions. (2)
162 GWAS causal gene detection: Select the most likely causal gene within a locus, demanding fine-
163 grained locus-level inference. (3) CRISPR perturbation screen design: Construct gene panels to
164 maximize post-perturbation effect across a large (∼20,000 genes) search space. (4) Rare disease
165 diagnosis: Map patient phenotypes and genetic findings to rare disease diagnosis. (5) Drug repur-
166 posing: Given a rare disease and a list of candidate drugs, select the best therapeutic match. (6)
167 Single-cell RNA-seq annotation: Assign accurate cell-type labels to individual cell profiles across
168 tissues, species, and platforms. (7) Microbiome disease-taxa analysis: Perform statistical associa-
169 tion tests on microbiome datasets to uncover disease-relevant taxa. (8) Patient gene prioritization:
170 Given an individual patient’s genetic profile and phenotype description, identify the most plausi-

7
171 ble causal gene. We benchmarked Biomni without prompt engineering or task-specific fine-tuning
172 against three baselines: (1) a base LLM (Claude Sonnet 3.7) without tool use, (2) a coding agent
173 with direct function calls and code execution (ReAct+Code), and (3) Biomni-ReAct, an ablation
174 of Biomni that replaces code-based planning with ReAct-style chaining. The complete benchmark
175 constructions are described in Methods, with detailed performance comparisons in Supplementary
176 Notes B.
177 Across all tasks, Biomni outperformed the base LLM by an average relative performance
178 gain of 402.3%, the coding agent by 43.0%, and its own ablated variant Biomni-ReAct by 20.4%.
179 These findings highlight the importance of code-centric planning and environment grounding, en-
180 abling Biomni to compose precise, flexible, and context-aware actions. For each benchmark, we
181 further analyzed the execution trajectories, identifying commonly invoked tools, software, and
182 datasets, as detailed in Supplementary Figures 6-16. These trajectories provide insight into the
183 complexity and structure of agent behavior across tasks. On average, Biomni executes between 6
184 and 24 distinct steps per task, involving combinations of 0-4 specialized tools, 1-8 software pack-
185 ages, and 0-3 unique data lake items. The agent interleaves data extraction, search/retrieval, rea-
186 soning, and computational analyses (Supplementary Figure 8) – reflecting a workflow pattern that
187 mirrors how human scientists alternate between retrieving knowledge and generating new insights.
188 Resource usage varies by task type: information synthesis tasks, such as CRISPR perturbation
189 screen design and GWAS causal gene identification, rely heavily on database queries (e.g., KEGG,
190 Reactome) and literature search (e.g., PubMed, Google), whereas bioinformatics analysis tasks
191 like microbiome profiling and single-cell annotation involve minimal database use but extensive
192 code execution with software libraries such as scanpy.

193 Biomni jointly analyzes 458 wearable sensor files to generate physiological hypotheses. To
194 evaluate Biomni’s performance in real-world biomedical workflows, we invited scientists to ap-
195 ply it directly to their own research questions. In this case study, a researcher used Biomni to
196 analyze 458 Excel files containing months-long wearable sensor data (continuous glucose moni-
197 toring (CGM) and body temperature) from 30 participants. The data were highly heterogeneous:
198 file formats varied, annotations were inconsistent, and participants exhibited substantial variability
199 (Figure 3a). The researcher posed an open-ended question: Can we uncover biologically meaning-

8
200 ful thermogenic patterns?
201 Biomni autonomously generated and executed a 10-step analysis pipeline (Figure 3b), in-
202 ferring meal events from glucose spikes, extracting pre/post meal temperature windows, normal-
203 izing across individuals, and synthesizing population-level trends. Crucially, after completing the
204 pipeline, the agent delivered a structured, human-readable report summarizing its key findings
205 (Supplementary Notes D). It identified a consistent postprandial thermogenic response, with an
206 average temperature rise of 2.19◦ C (median: 1.10◦ CC) and a wide range across individuals (-
207 0.11◦ CC to 15.56◦ CC). Some participants showed rapid, pronounced spikes within 30 minutes of
208 eating, while others had delayed or muted responses – indicating divergent metabolic phenotypes
209 (Figure 3c,d). These insights were not manually curated or extracted by a human; the agent per-
210 formed the entire analysis end-to-end and surfaced the results as a concise narrative highlighting
211 patterns that would otherwise being ignored in raw data.
212 In a parallel workflow, the scientist requested Biomni to analyze 227 nights of wearable-
213 recorded sleep data across 10 participants. Biomni computed averages for duration, efficiency,
214 latency, and sleep stage composition, derived a composite sleep quality score, and conducted
215 chronobiological analyses. The agent delivered a structured summary to the user (Supplemen-
216 tary Notes D, Supplementary Figure 4), including personalized sleep profiles and timing insights,
217 without human post hoc synthesis. Biomni uncovered several novel insights: sleep efficiency
218 consistently peaked mid-week (on Wednesdays) and declined on Sundays, suggesting a potential
219 behavioral pattern tied to pre-Monday stress or weekend-induced disruptions. Another important
220 finding was that consistent sleep timing correlated more strongly with higher sleep quality than
221 total sleep duration, highlighting the critical role of circadian regularity in maintaining restorative
222 sleep.
223 The scientist then tasked Biomni with analyzing multi-omics data (652 lipidomic, 731 metabolomic,
224 and 1,470 proteomic features), jointly with the CGM data. Biomni conducted cross-omics correla-
225 tion analysis, applied hierarchical clustering to uncover biologically coherent feature groups, and
226 performed unsupervised PCA to link CGM signals to molecular pathways. It automatically gener-
227 ated interpretable outputs – trajectory plots, heatmaps, boxplots, PCA biplots, and cluster maps –
228 empowering rapid insight generation from complex multimodal datasets (Supplementary Notes D,
229 Supplementary Figure 5). Significant correlations among lipids, metabolites, and proteins revealed

9
230 tightly interlinked regulatory pathways, underscoring the systems-level nature of metabolic regula-
231 tion. Notably, several identified biomarkers showed consistent patterns across samples and exhib-
232 ited high connectivity within correlation networks. Across all cases, the scientist noted that Biomni
233 accelerated the path from messy real-world data to testable hypotheses, supporting applications in
234 sleep optimization, metabolic research, and precision health.

235 Biomni automates complex multi-omics analysis to decipher transcriptional regulation of

236 skeletal lineages. To test whether Biomni could generalize to complex omics workflows, a scien-
237 tist used it to analyze a recently published multi-omics dataset of the developing human skeleton 25 .
238 This dataset comprises 336,162 single-nucleus RNA-Seq (snRNA) and ATAC-seq (snATAC-Seq),
239 paired with spatial transcriptomics data collected from human embryos between 5-11 weeks post-
240 conception (Figure3e). While the original study emphasized developmental trajectories and dis-
241 ease mechanisms, the scientist was interested in exploring gene regulatory mechanisms across
242 emerging skeletal cell types – a technically demanding task typically requiring extensive bioinfor-
243 matics support.
244 The scientist asked Biomni to investigate transcriptional regulation across skeletal lineages
245 using a detailed instruction (Supplementary Notes E). The system autonomously planned and exe-
246 cuted a ten-stage analysis pipeline: (1) loading and exploring all datasets, (2) preparing RNA-seq
247 data for analysis, (3) configuring pySCENIC to retrieve motifs, (4) running GRNBoost2 to infer
248 gene regulatory networks, (5) pruning networks using cisTarget, (6) calculating regulon activity
249 with AUCell, (7) extracting accessibility data from ATAC-seq, (8) filtering predicted targets us-
250 ing ATAC-seq accessibility, (9) analyzing activity patterns across cell types, developmental stages,
251 and anatomical regions, and (10) summarizing findings and preparing a report to the scientist. It
252 enabled Biomni to predict transcription factor-target gene links and filter regulons based on mo-
253 tif enrichment and chromatin accessibility correlations (Figure 3f). The full run, completed in
254 just over five hours, handled real-time execution issues (e.g., variable name mismatches) by sub-
255 sampling and debugging locally. Throughout, Biomni maintained all intermediate outputs – code,
256 figures, and logs – organized in a reproducible folder structure for validation and inspection. The
257 agent summarized all the analysis and generated a report describing the analysis and key findings
258 (Supplementary Notes E).

10
259 In its final gene regulatory network (GRN) analysis (Figure 3h), Biomni re-capitulated known
260 regulatory relationships between key osteogenic transcription factors such as RUNX2 and HHIP,
261 confirming how they are regulated by a shared set of anti-osteogenic transcription factors including
262 TWIST1, LMX1B, and ALX4 25 . These findings align with author’s report 25 about the balanced
263 regulation needed for proper bone formation and suture patency. Furthermore, Biomni also uncov-
264 ered several unreported TFs, including AUTS2, ZFHX3, and PBX1, showed unexpectedly high
265 regulatory activity across multiple skeletal cell types. Although PBX1 is a well-established skele-
26 27
266 tal regulator and ZFHX3/AUTS2 have only limited or indirect skeletal reports (in mouse
28
267 or zebrafish ), their broad activity here suggests under-appreciated roles across diverse skele-
268 tal lineages. Biomni reported that these novel regulators were particularly active in osteoblasts,
269 preosteoblasts, and various chondrocyte populations, suggesting they play important but previ-
270 ously unrecognized roles in the transcriptional control of skeletal cell fate determination during
271 human embryonic development. Finally, Figure 3g-h reveals how Biomni’s visualizations effec-
272 tively captured both temporal dynamics of regulator activity and cell-type-specific variations in key
273 regulons like RUNX2. This demonstrates how Biomni enables researchers to autonomously per-
274 form complex multi-omics analysis and rapidly generate testable hypotheses without specialized
275 programming expertise.

276 Biomni designs wet-lab validated experimental protocol for cloning. To evaluate Biomni’s
277 ability to support real-world experimental design, we focused on a core task in molecular biology:
278 cloning. This process is central to countless workflows in research and biotechnology and requires
279 complex reasoning, from designing high-fidelity primers to choosing the right assembly method
280 and validating constructs. While general-purpose LLMs have struggled to perform such tasks due
281 to limited domain knowledge and tool access 24 , Biomni integrates LLM reasoning with dynamic
282 tool execution, enabling expert-level performance in molecular biology tasks.
283 To rigorously evaluate this task, we first collaborated with an expert group of gene-editing
284 researchers to design an open-ended cloning benchmark and expert user study (Figure 4a). Our
285 benchmark consisted of 10 realistic, representative cloning tasks covering Golden Gate, Gibson,
286 Gateway, and restriction cloning – each with options including single-fragment vs. pooled assem-
287 bly. The benchmark also included essential validation steps, such as designing Sanger sequencing

11
288 primers and analyzing restriction digests. We posed these tasks to four entities: an LLM (Claude
289 3.7), Biomni, a human trainee (Stanford Biology Master with previous experience in cloning),
290 and a senior human expert (Stanford Genetics PostDoc with 5+ years of cloning experience).
291 Each was asked to generate a complete, end-to-end protocol along with the final cloned plas-
292 mid map. Blinded expert reviewers assessed the outputs. Biomni produced protocols and designs
293 that matched the human expert in accuracy and completeness – often providing comparable lev-
294 els of detail and anticipating the same edge cases. In contrast, the human trainee’s submissions
295 were frequently incomplete or suboptimal, reflecting the experience gap typical in early-stage re-
296 searchers. Remarkably, Biomni completed all tasks autonomously in a fraction of the time taken
297 by the expert.
298 To further validate Biomni in a real-world setting, a scientist assigned it a practical cloning
299 task: cloning a guide RNA targeting the human B2M gene into the lentiCRISPR v2 Blast construct
300 (Figure 4b). Biomni successfully executed the task through a comprehensive workflow (Figure 4c).
301 First, it analyzed the plasmid structure using annotation and pattern search tools to identify key fea-
302 tures necessary for cloning. It then designed three Cas9 sgRNAs targeting B2M using specialized
303 knockout sgRNA design tools. For the cloning process, Biomni generated forward and reverse
304 oligos with BsmBI overhangs to enable directional insertion of the sgRNA sequence. It produced
305 detailed protocols (Figure 4d) for oligo annealing, double-stranded DNA formation, and Golden
306 Gate cloning into the target vector. Biomni also provided complete bacterial transformation in-
307 structions, including heat-shock steps and antibiotic selection. For quality control, it designed a
308 U6 promoter sequencing primer to verify sgRNA insertion and simulated the Golden Gate assem-
309 bly to produce the final plasmid map.
310 The scientist followed Biomni’s protocol exactly to perform the wet-lab experiment (Fig-
311 ure 4e). Colonies appeared on the plate the next day; two were cultured, miniprepped, and se-
312 quenced using the Biomni-designed primers – both showing perfect alignment. This case illustrates
313 how scientists can rely on Biomni to autonomously design complex molecular biology experiments
314 with accuracy comparable to human experts, but in a fraction of the time.

315 User-friendly interface to empower scientists to generate biomedical discoveries. To bring the
316 power of Biomni into the hands of every scientist, we built an intuitive graphical interface – avail-

12
317 able at https://biomni.stanford.edu – to help transform the way researchers interact with biomedical
318 data and tools. This seamless platform enables users to submit natural language queries and receive
319 results powered by the full capabilities of Biomni’s agentic system. Whether designing complex
320 cloning experiments, querying multi-omics databases, or generating hypotheses from wearable
321 data, scientists can now access the intelligence of a general-purpose biomedical AI agent without
322 writing a single line of code. The interface is designed for rapid iteration, real-time feedback, and
323 visual traceability, allowing users to explore intermediate steps, inspect tool usage, and validate
324 results interactively. By closing the gap between biomedical intent and execution, Biomni opens a
325 new era of accessible, automated, and scalable scientific discovery. An example of this interface is
326 shown in Supplementary Figure 17.

327 3 Discussion
328 Biomni marks a major step forward in biomedical research, demonstrating robust generalization
329 across diverse subfields and laying the groundwork for AI agents as integral collaborators in scien-
330 tific discovery. Its zero-shot performance across complex tasks – including those in genetics, ge-
331 nomics, microbiology, immunology, pharmacology, and clinical medicine – underscores its poten-
332 tial to boost research productivity, accelerate discovery, and broaden access to advanced biomedical
333 analyses.
334 By automating complex, labor-intensive workflows, which normally require both expert
335 knowledge and coding skills, Biomni enables researchers to redirect their efforts toward creative
336 hypothesis generation, experimental innovation, and cross-disciplinary collaboration. This shift
337 holds profound implications. In the conext of target and drug discovery for biopharma, Biomni
338 can autonomously prioritize targets, design perturbation screens, or repurpose drugs – offering a
339 path to faster, more cost-effective reasearch. In clinical application settings, its capabilities in gene
340 prioritization and rare disease diagnosis point to more accurate, personalized insights and stream-
341 lined diagnostics. For consumer health, Biomni’s integration of wearable data and multi-omics
342 analyses envisions real-time, individualized health monitoring and intervention.
343 Nonetheless, several limitations remain. While Biomni’s unified environment spans a wide
344 range of biomedical tools and databases, the evaluated tasks represent only a subset of the field,
345 and key domains remain unexplored. In addition, in the action discovery agent, our decision to pri-

13
346 oritize the most recent literature makes the agent appear timely, but risks overlooking foundational
347 concepts and techniques that have faded from current discourse despite their enduring relevance.
348 The future versions should encapsulate a larger coverage of publications when defining the envi-
349 ronment. Moreover, although Biomni approaches human-level performance in tasks like database
350 querying, sequence analysis, and molecular cloning, it still struggles in areas requiring nuanced
351 clinical judgment, novel experimental reasoning, analytical inventions, or deep biological thinking
352 and synthesis. No system yet captures the full scope of human biomedical expertise. As reflected
353 in our benchmarks, Biomni has not achieved expert-level performance across all task categories.
354 We expect continued improvements as foundation models evolve and the agentic environment ex-
355 pands, as well as thanks to human experts and trainees deploying Biomni to facilitate or augment
356 their work.
357 These limitations open promising directions for future development. Training biomedical
358 reasoning agents with reinforcement learning could enable continuous self-improvement in plan-
359 ning and execution. Integrating multimodal data – text, images, and structured inputs – may further
360 deepen reasoning capabilities. Equipping Biomni to autonomously discover and incorporate new
361 tools and databases, as well as incorporate more historical methods (which may have high utility
362 but can be easily forgotten by human users), would ensure adaptability and long-term relevance.
363 Looking ahead, Biomni and its successors could become foundational infrastructure in an AI-
364 powered biomedical ecosystem, working seamlessly with human experts to unlock novel insights
365 into health and disease. This hybrid partnership may radically reshape biomedical research –
366 automating hypothesis generation, scaling discovery pipelines, and enabling medical innovation
367 to proceed at unprecedented speed and scope. General-purpose agents like Biomni could not only
368 accelerate breakthroughs but redefine the future of scientific inquiry itself.

14
 a 25 subjects defined in biorxiv b An AI Agent Environment with c Are oncogenic pathways enriched among
Subject 25 Unified Biomedical Action Space differentially expressed genes in adata.h5ad?

Subject 1
Genomics Retrieving
Biochemistry Pathology Scanpy gseapy ... CellxGene ClinVar ... UCE
RNA ...
velocity

Genetics

...
100 recent published papers Reasoning
Molecular DESeq2 To answer the question:
macs2 [ ] Preprocess, cluster and identify differentially expressed genes.
biology gnomAD Synthetic [ ] Load oncogenic gene sets for comparison.
[ ] Perform enrichment analysis to check for oncogenic pathways.
BindingDB biology [ ] Validate results with literature and databases.
...
miRBase
OpenTarget
... cooler
Genetics
Scanpy UCE IUCN primer
? Composing actions with code

flowkit virtual design

screening
Action discovery agent ... PDB
PLINK
GCTA CellxGene libsbml Tools
RegulomeDB
... AlphaFold ...
DB
Actions necessary to conduct DiffDock
Cell
biomedical research Biophysics ADMET ClinVar
Observation
Pred
Oligo InterPro biology
Biopython DKK1, UGT1A, EEF1A1, RAC2, PTGDS, COL1A1, TIMP1, SPP1...
sk-image assembly

Expert Reasoning
golden JASPAR RNA muscle
curation gate clone velocity
[ ] Preprocess, cluster and identify differentially expressed genes.
UCSC [ ] Load oncogenic gene sets for comparison...
bwa
Bioengineering Physiology Composing actions with code
...

Homer COSMIC
...

105 softwares 150 biological specialized tools viennaRNA Raw data

Wet-lab AI models Know-hows Software
MPD P P P
...

59 databases
S Observation
AI
Biochemistry ......
Pharmacology
Answer
Microbiology

Figure 1: Overview of the unified biomedical action space and agent environment in Biomni. (a) Workflow for systematically
curating the unified biomedical action space. Actions necessary to conduct biomedical research were extracted from 2,500 recent
bioRxiv publications across 25 biomedical subfields using an AI-driven discovery agent. Extracted actions were rigorously validated
and curated by human experts, resulting in the integration of 105 biomedical software tools, 150 specialized biological tools (includ-
ing wet-lab protocols, AI-driven predictive models, and domain-specific know-how), and 59 comprehensive biomedical databases.
(b) Illustration of the unified biomedical action space spanning diverse biomedical subfields such as genetics, genomics, synthetic
biology, cell biology, physiology, microbiology, pharmacology, bioengineering, biophysics, molecular biology, and pathology. Rep-
resentative tools and databases integrated into Biomni’s environment are shown, highlighting its general-purpose capabilities. (c)
Example workflow demonstrating Biomni’s reasoning and action composition process to autonomously answer a complex biological
question. Biomni retrieves relevant tools based on the user’s query, formulates a structured reasoning plan, and composes executable
code to perform comprehensive bioinformatics analyses, iteratively refining its reasoning based on observations until converging on
a final, precise answer.

15
 a Humanity's Last Exam:
Biomedicine LAB-Bench: DbQA LAB-Bench: SeqQA

Human
Human

b
15

Variant prioritization GWAS causal gene detection Perturbation screen design

log10(P)
A T G C G A C T 10

-
0

10
11
12
13
14
15
16
17
18
20
19
21
22
7
8
9
2
3

5
6
4
1
Chromosome

I have a list of potential variants for

I have a GWAS hit for trait XX, there are
disease X, prioritize the top variant for Design a gene panel of 50 genes for my
genes XXX, XXX in the locus, which is
me. CRISPR screen on T-cell exhaustion.
the most likely causal gene?
+162% +78.9% +3.0% +9.8% +13.2% +23.0%
+54.2% +19.4% +2.7%

Patient gene prioritization

scRNA-seq cell annotation
I have a patient with phenotypes XXX, Genetics
Genetics Genomics
Genomics
XXX, and WGS returns gene mutations I have scRNA-seq data stored at XXXX,
XXX, what is the causal gene? annotate the cell type for me.

+8.8% +19.4% +5.7%

+inf +10.3% +93.6%

Patient
Microbiology

Drug

Rare disease diagnosis Drug repurposing Microbiome disease-taxa

bioinformatics analysis
I have a patient with phenotypes XXX, I have a rare disease XX, among a list of
XXX, and WGS returns gene mutations I have microbiome raw data stored at
potential drugs, find the most likely one
XXX, what is the diagnosis? XXXX, analyze which taxa is associated
for repurposing.
with the disease XXX.
X21 +19.4% +10.3% X4.8 +180.0% +25.0%
+1.3% +3.6% +0.2%

Figure 2: Zero-shot generalization of Biomni across diverse realistic biomedical tasks. (a) Biomni is superior to 6 baselines in
Q&A multiple choice benchmarks that broadly evaluate the model’s capability across biomedical fields. (b) Biomni demonstrates
robust zero-shot performance across eight previously unseen, real-world biomedical scenarios spanning multiple biomedical sub-
fields, without any task-specific fine-tuning or prompt engineering. Evaluated tasks include variant prioritization and GWAS causal
gene detection (genetics and genomics), perturbation screen design (functional genomics, immunology), patient gene prioritiza-
tion, rare disease diagnosis (clinical genomics), drug repurposing (pharmacology), microbiome disease-taxa bioinformatics analysis
(microbiology), and single-cell RNA-seq cell annotation (single-cell biology). Across these diverse scenarios, Biomni consistently
outperformed baseline models (Base LLM, ReAct+Code) and specialized environments (Biomni ReAct), highlighting its general-
purpose biomedical capabilities and ability to autonomously adapt to new and complex biomedical tasks.
16
 a I have CGM-derived data from multiple
...... participants collected continuously around meals.
Study how temperature changes after food intake.
...... [instructions] ......

Continuous
30 samples Activity 458 Raw Excel
Glucose
Data Sheets
Monitoring Data

b
1. Dataset exploration 3. Postprandial 4. Individual temperature 5. Cross-subject
& preprocessing 2. Meal event detection
temperature analysis response plots comparative analysis

10. Identification of individual 9. Temperature increase 8. Response pattern 7. Peak thermogenic 6. Temperature
metabolic variability distribution analysis heatmap generation timing characterization response normalization

c Key figures generated from Biomni d Findings:

● On average, participants exhibited a 2.19°C
increase in core body temperature following
meal intake, with a median increase of 1.10°C.
● Participants displayed considerable inter-
individual variability, with some showing rapid
and pronounced thermogenic responses,
while others had more subtle or delayed
temperature changes.

e
Map transcription factor circuits
{Shoulder, Hip, driving skeletal development on
this scRNA & scATAC data.
Knee} Joint ...... [instructions] ......

snRNA-seq snATAC-seq
Time
336,000 nucleus droplets
f
1. Load and explore all 2. Prepare RNA-seq 3. Configure pySCENIC (e.g., 4. Run GRNBoost2 to infer 5. Prune networks
the datasets data for analysis retrieve motifs from data lake) gene regulatory networks using cisTarget

10. Summary 9. Analyze activity patterns across cell types, 8. Filter predicted targets 7. Extract accessibility 6. Calculate regulon
findings developmental stages, and anatomical regions using ATAC-seq accessibility data from ATAC-seq activity (AUCell)

g Key figures generated from Biomni h Findings:

● Novel transcription factors AUTS2, ZFHX3, and
PBX1 showed high regulatory activity across
multiple skeletal lineages.
● Cell types LimbMes exhibited the highest overall
activity levels among all the identified 566-589
regulons across cell types.

Figure 3: Biomni autonomously executes complex multi-modal biomedical analyses to generate hypothesis. (a-d) Biomni rapidly
analyzed CGM-derived thermogenic responses data and activity data from 30 individuals, comprising 458 raw Excel sheets. (b)
Workflow demonstrating Biomni’s autonomous execution of data preprocessing, meal event detection, postprandial temperature
analysis, and thermogenic response characterization. (c) Representative individual temperature-response plots and temperature in-
crease distribution following meals, automatically generated by Biomni. (d) Summary of unique biological findings identified
by Biomni, including significant increases in core body temperature post-meal intake (average 2.19C, median 1.10C), and notable
inter-individual variability in thermogenic responses. (e-h) Biomni autonomously analyzed single-cell multiomics data from approx-
imately 336,000 nucleus droplets, combining single-nucleus RNA (snRNA-seq) and single-nucleus ATAC sequencing (snATAC-seq)
across human embryonic joint development (shoulder, hip, knee). (f) A detailed workflow diagram showing Biomni’s 10-step analy-
sis pipeline for gene regulatory networks with multiomics. (g) Two key figures generated from Biomni: Left panel shows a heatmap
of regulator activity by developmental stage, with color intensity indicating activity levels. Right panel displays a boxplot of RUNX2
regulon activity by cell type, showing variation in expression across different cell populations. (h) Key findings from the GRN anal-
ysis: 1) Novel transcription factors (AUTS2, ZFHX3, and PBX1) showing high regulatory activity across multiple skeletal lineages
despite no previous association with skeletal development, and 2) Across the 566-589 regulons recovered, limb mesenchyme cells
display the highest mean regulonactivity score, underscoring their prominent role in skeletal transcriptional control.

17
 a Cloning Open Answer Evaluation b I have a plasmid plentiCRISPR v2 blast:
CGTTGGAGTCCACGTTCTTTAATAGTGG.....GCCCTTTGA.
I hope to clone a CRISPR sgRNA targeting human B2M
GG Multiple-piece GG Restriction
into this plasmid. Could you give me the final map of
the plasmid and step-by-step guidance on how should I
GG Library Gibson Restriction Library perform the cloning?

Gibson Library Gateway Diagnostic digest

c
3. Oligo Design 4. Oligo Annealing & 6. Colony
1. Plasmid 2. sgRNA 5. Bacterial 7. Final Plasmid
for Cloning Golden Gate Assembly Screening &
Analysis Design Transformation Map Assemble
Verification

d Biomni generated wet-lab cloning protocol Plasmid Map

### 4. Bacterial Transformation
## Step-by-Step Cloning Protocol
1. Add 2 μL of the Golden Gate reaction to 50 μL of
competent Stbl3 E. coli cells
### 1. Design and Order Oligos
2. Incubate on ice for 30 minutes
- **Forward oligo**: 5'-
3. Heat shock at 42°C for 45 seconds
CACCGCGCGAGCACAGCTAAGGCCA-3'
4. Return to ice for 2 minutes
- **Reverse oligo**: 5'-
5. Add 950 μL of SOC medium
AAACTGGCCTTAGCTGTGCTCGCGC-3'
6. Incubate at 37°C for 1 hour with shaking
7. Plate 100 μL on LB agar plates with ampicillin (100 μg/mL)
### 2. Anneal and Phosphorylate Oligos
8. Incubate plates at 37°C overnight
1. Resuspend oligos to 100 μM in nuclease-free water
2. Prepare annealing reaction:
- 1 μL Forward oligo (100 μM)
### 5. Colony Screening and Verification
- 1 μL Reverse oligo (100 μM)
1. Pick 3-5 colonies and grow in 5 mL LB + ampicillin
- 1 μL 10X T4 Ligation Buffer
medium overnight
- 6.5 μL nuclease-free water
2. Extract plasmid DNA using a miniprep kit
- 0.5 μL T4 PNK
3. Verify correct insertion by Sanger sequencing using the U6
3. Incubate in a thermocycler:
promoter forward primer:
- 37°C for 30 minutes (phosphorylation)
5'-GACTATCATATGCTTACCGT-3'
- 95°C for 5 minutes (denaturation)
4. Confirm the presence of the B2M sgRNA sequence in the
- Ramp down to 25°C at 5°C/min (annealing)
plasmid
4. Dilute annealed oligos 1:200 in nuclease-free water
(final: 50 nM)
### 6. Lentivirus Production and Cell Transduction
1. Transfect the verified plasmid into HEK293T cells along
### 3. Golden Gate Assembly
with packaging plasmids
1. Prepare reaction:
2. Collect virus-containing supernatant after 48-72 hours
- 100 ng lentiCRISPR v2 blast plasmid
3. Filter and concentrate the lentivirus if needed
- 1 μL diluted annealed oligos (50 nM)
4. Transduce target cells with the lentivirus
- 2 μL 10X T4 DNA Ligase Buffer
5. Select transduced cells with blasticidin (5-10 μg/mL)
- 1 μL BsmBI-v2 (NEB)
6. Verify B2M knockout by flow cytometry, Western blot, or
A scientist
- 1 μL T4 DNA Ligase
- Nuclease-free water to 20 μL total
genomic sequencing follows the
2. Incubate in a thermocycler: protocol to
- 37°C for 5 minutes
- 16°C for 5 minutes perform the
- Repeat steps 1-2 for 10 cycles
- 55°C for 5 minutes (final digestion)
cloning
- 80°C for 5 minutes (enzyme inactivation)

Figure 4: Biomni designs wet-lab experimental protocol. (a) Open-ended cloning benchmark on 10 real cloning scenarios. We
compared against base LLM, trainee-level human, and expert-level human scientists. We found that Biomni has similar accuracy
as the expert level scientist, and significantly higher accuracy than trainee level, while using much less time. (b) Example of a
user request to Biomni for cloning an sgRNA targeting the human B2M gene into the lentiCRISPR v2 Blast plasmid. (c) Biomni’s
automated stepwise workflow, including plasmid analysis, sgRNA design, oligo synthesis, Golden Gate assembly, bacterial transfor-
mation, colony screening, and final plasmid mapping. (d) Biomni-generated detailed cloning protocol with step-by-step instructions
and comprehensive plasmid map, enabling laboratory scientists to execute the experiment autonomously. (e) Validation of Biomni’s
cloning protocol through successful colony growth on selection plates, followed by Sanger sequencing confirming perfect alignment
of sgRNA insertion in picked colonies, demonstrating Biomni’s robust capability for precise and reliable experimental design.

18
369 4 Methods
370 Action Discovery from Literature. 100 recent publications from the year 2024 at biorxiv Were
371 collected and analyzed by extracting and parsing their PDF contents. Each paper was processed
372 in chunks, and a specialized prompt guided an LLM through each chunk to explicitly identify
373 and extract three categories of actionable insights: tasks, software, and databases. Specifically for
374 tasks, the LLM was instructed to highlight recurrent tasks requiring specialized implementations
375 within biomedical research workflows.
376 Implementing the Biomni Environment. In the initial iteration of environment construction, a
377 conservative and focused approach was adopted for tool curation. Initially, tasks were filtered
378 based on relevance to the primary research interests-drug discovery and clinical biomedicine-
379 retaining fields such as biochemistry, bioengineering, biophysics, cancer biology, cell biology, de-
380 velopmental biology, genetics, genomics, immunology, microbiology, molecular biology, pathol-
381 ogy, pharmacology, physiology, synthetic biology, and systems biology. Subsequently, these were
382 narrowed down to approximately 1,900 commonly recurring tasks. These tasks were further man-
383 ually reviewed to eliminate redundancy and exclude tasks that were trivial or easily implementable
384 through simple code. Selecting highly specialized tasks that require significant domain expertise
385 was emphasized, such as wet-lab protocols and advanced AI models.
386 Human scientists then collaborated with software engineering agents equipped with web
387 search capabilities to implement each specialized tool. Every tool underwent rigorous validation,
388 requiring a clearly defined test case that it successfully passed. This stringent process culminated in
389 a curated collection of 150 specialized tools. Additionally, essential literature retrieval tools were
390 included, such as PubMed and Google Scholar, with provisions for future iterative expansions.
391 Each tool was strictly defined using a comprehensive checklist that mandated: (1) a clear
392 and descriptive name, (2) detailed documentation, (3) outputs formatted as detailed research logs
393 optimized for LLM interpretation, (4) the inclusion and successful passing of a specific test case,
394 and (5) specialization criteria-if a task could easily be implemented via brief LLM-generated code
395 (e.g., simple database queries), no specialized tool was created.
396 Databases were categorized and extensive relational databases accessible via web APIs (e.g.,
397 PDB, OpenTargets, ClinVar) were integrated using a unified querying function. This function
398 accepts natural language inputs and leverages an LLM to dynamically parse database schemas
399 and execute corresponding queries. Databases lacking web APIs were downloaded and locally
400 preprocessed into structured pandas DataFrames for seamless accessibility by the agent.
401 For software integration, recognizing the frequent necessity of concurrently utilizing multiple

19
402 software tools, a unified containerized environment was constructed, which was pre-installed with
403 a comprehensive suite of relevant software. Additionally, this environment supports the execution
404 of R packages and command-line interface (CLI) tools.
405 Biomni-A1. The Biomni agent is a general-purpose biomedical AI agent built upon the CodeAct29
406 framework, designed to systematically solve biomedical tasks by combining LLMs with an inter-
407 active coding environment. Given a user query, Biomni begins by prompting the LLM to generate
408 a clear, numbered bullet-list plan detailing the steps needed to tackle the given problem, keeping
409 careful track of progress and adjustments along the way. As the tool, software, and database space
410 is vast, the query task may only use a small set of these resources. To avoid long context, a prompt-
411 based retriever is utilized, powered by a separate LLM, where the agent dynamically selects the
412 most relevant functions, datasets, and software libraries from available resources. During execu-
413 tion, the LLM generates code, executes it in a coding environment (Python, R, or Bash), and returns
414 the resulting observations to inform subsequent reasoning. This iterative approach continues until
415 the agent converges on an accurate, validated solution.
416 Q&A Benchmarks. Development and testing sets were created by sampling the LAB-Bench
417 Database Question-Answering and Sequence Question-Answering benchmarks 24 . Due to resource
418 constraints, each set comprises 12.5% of the complete reference, proportionally distributed across
419 benchmark subtasks, providing a cost-effective and representative assessment of model perfor-
420 mance. The development set informed iterative refinements to Biomni’s database integrations and
421 tool implementations, while the test set provided an independent evaluation of generalization ca-
422 pabilities. Accuracy was evaluated by following the LAB-Bench protocol, using multiple-choice
423 answer options with an option for abstention due to insufficient information. Results represent
424 averages across three independent evaluation runs.
425 For Humanity’s Last Exam (HLE)23 , a representative sample of questions was selected, span-
426 ning fourteen subdisciplines of Biology/Medicine: Genetics, Biology, Ecology, Neuroscience,
427 Biochemistry, Microbiology, Immunology, Molecular Biology, Computational Biology, Biophysics,
428 Bioinformatics, Genomics, and Physiology. From each subdiscipline, up to five questions were
429 sampled (or the maximum number available if fewer than five existed in the category). This
430 sampling approach yielded a final evaluation set of 52 questions that comprehensively assessed
431 Biomni’s performance across the biological sciences. The evaluation was conducted directly with-
432 out the use of a development set.
433 Curating real-world benchmarks. The variant prioritization benchmark was curated from Open
434 Target Genetics30 ground truth set, and processed such that given a variant, a negative set of vari-

20
435 ants is found. The prompt was as follows: ”Your task is to identify the most promising variant
436 associated with a given GWAS phenotype for futher examination. From the list, prioritize the
437 top associated variant (matching one of the given variant). GWAS phenotype: {trait} Variants:
438 {variant list}”. Accuracy was used as the metric. The GWAS causal gene detection benchmark
439 utilized a dataset curated from Shringarpure et al31 , using the original prompt: ”Your task is to
440 identify likely causal genes within a locus for a given GWAS phenotype. From the list, provide
441 only the likely causal gene (matching one of the given genes). Identify the causal gene. GWAS
442 phenotype: {trait} Genes in locus: {gene str}”. Accuracy was used as the metric. The pertur-
443 bation screen design benchmark was curated from Schmidt et al.32 . The prompt is ”Task: Plan
444 a CRISPR screen to{task description}. There are 18,939 possible genes to perturb and only per-
445 turb {num genes} genes. For each perturbation, you can measure out {measurement} which will
446 be referred to as the score. Generate {num genes} genes that maximize the perturbation effect.
447 Output format: a list of genes 1. XXX 2.XXX 3.XXX ...”. The evaluation metric was the average
448 post-perturbed effect. As the scale differs for the post-perturbed effect, one screen (IL-2) was used.
449 The scRNA-seq annotation benchmark ensured flexibility across diverse data formats (e.g., Cel-
450 lxGene, author-hosted portals), encompassing multiple tissues, species, sequencing technologies,
451 and experimental conditions. Datasets with author-provided annotations (Tier 1 or Tier 2, typically
452 ≥10 cell types) were prioritized, and 20k-50k cells were subsampled proportionally to their cell
453 type distributions. Automatic evaluation was conducted at the single-cell level using LLMs via
454 semantic match, accounting for both naming variations (e.g., fibroblast vs. Fibroblast cells) and
455 hierarchical relations (e.g., CD8+ T cells vs. T cells), judged on-the-fly by LLM agents and later
456 verified by humans. In the microbiome benchmark, both Biomni and human experts independently
457 performed differential abundance analysis on five diverse microbiome datasets, selected to reflect
458 different data types, biological contexts, and analytical challenges. Dataset 1 comes from the MGM
459 2.0 platform33 and includes relative microbial abundance across samples and another with sample
460 labels, ideal for classification tasks33 . Dataset 2 curated from a well-known Nature study, offers
461 microbial abundance data in mice alongside metadata such as diet and sex, making it valuable for
462 modeling host-microbiome interactions34 . Dataset 3, developed by Pasolli et al.35 , combines eight
463 human metagenomic studies with species-level features processed using MetaPhlAn235 . Dataset 4
464 explores microbial communities in drinking water systems, providing an OTU matrix with abun-
465 dances represented as relative sequence counts. This environmental dataset allows models to be
466 tested beyond host-associated microbiomes36 . Finally, Dataset 5 is an in-house resource derived
467 from the Human Microbiome Project37 . Together, these datasets provide a comprehensive foun-
468 dation for benchmarking AI agents in microbiome analysis across both clinical and environmental

21
469 domains. Biomni results were compared against those generated by human experts for consistency,
470 accuracy, and efficiency. The drug repurposing benchmark used a dataset from Huang et al.38 , for
471 the task of identifying the most likely drug from a pre-defined list of drugs for repurposing in a
472 given indication. Evaluation was based on the alignment score with off-label prescription patterns
473 of clinicians from an EHR system. The prompt was ”Your task is to identify top 5 drugs that can
474 be potentially repurposed to treat the given disease. From the list, prioritize the drug list with
475 the highest potential (matching the given DrugBank IDs). Disease: {disease} Drugs: {drug list}
476 Output format: a list of drugs with their DrugBank IDs, no drug name, just the IDs: 1. DB00001
477 2. DB00002 3. DB00003 ..”. The rare disease diagnosis benchmark used the MyGene2 dataset,
478 curated by Alsentzer et al.39 . The ground truth was expert annotated diagnosis. The prompt was
479 ”Task: given a patient’s phenotypes and a list of candidate genes, diagnose the rare disease that
480 the patient has. Phenotypes: {phenotype list} Candidate genes: {candidate genes} Output format:
481 {{’disease name’: XXX, ’OMIM ID’: XXX}}”. The patient gene prioritization benchmark used
482 a dataset curated by Alsentzer et al.39 . The ground truth was a truly causal gene. The prompt was
483 ”Task: Given a patient’s phenotypes and a list of candidate genes, identify the causal gene. Phe-
484 notypes: {phenotype list} Candidate genes: {candidate genes} Output format: {{’causal gene’:
485 [gene1]}}”.
486 Wearable analysis case study. A wearable case study integrated CGM-derived body temperature
487 data, sleep metrics, and multi-omics datasets from human participants40 , as follows: CGM Body
488 Temperature Data: For each participant, continuous glucose monitors (CGMs) equipped with tem-
489 perature sensors recorded skin temperature in high resolution. A total of 485 temperature files were
490 collected, each centered on a presumed meal event. The time window for each file spanned 6 hours
491 total, comprising 2 hours pre-meal and 4 hours post-meal. Sleep Data: Sleep metrics were derived
492 from wrist-worn wearable devices for a subset of 10 participants, covering 227 nights of sleep. Pa-
493 rameters collected included sleep duration, sleep efficiency, sleep latency, sleep stage composition
494 (light, deep, REM), and number of wake episodes. Omics Data: Blood samples were analyzed to
495 generate the following: Lipidomics: 652 lipid features across 147 samples; Metabolomics: 731
496 metabolite features across 147 samples; Proteomics: 1,470 protein features across 20 samples.
497 Multiome analysis case study. The authors’ dataset was directly downloaded and used with no
498 modifications 25 . The authors’ study generated a multi-omic dataset of human embryonic skeletal
499 development from 5-11 weeks post-conception. The dataset includes snRNA-seq and snATAC-seq
500 data from approximately 336,000 nuclei across five anatomical regions (hip, knee, shoulder joints,
501 calvaria, and skull base). The dataset covers both appendicular (limb) and cranial regions. No

22
502 additional tools or manual preprocessing were added. As the analytical traces are extensive, more
503 guidance was included in the prompt instruction and two use cases were tested:
504 Comparative Analysis. This analysis focused on how cellular processes differ across anatomi-
505 cal locations and developmental timepoints. Biomni was instructed to characterize the cellular
506 composition across anatomical regions (calvaria, skull base, shoulder, hip, knee) and developmen-
507 tal stages. We prompted Biomni with detailed instructions (Supplementary Section E), including
508 cell type proportion estimates, region-specific population labels, UMAP embeddings, stacked bar
509 plots, a comparison of intramembranous versus endochondral ossification, key transcription factor
510 highlights, and developmental trajectory tracing.
511 Gene Regulatory Network Analysis We asked Biomni to identify transcriptional programs underly-
512 ing skeletal development. Following a systematic 10-step process, Biomni inferred gene regulatory
513 networks by: (1) loading and exploring all datasets, (2) preparing RNA-seq data for analysis, (3)
514 configuring pySCENIC to retrieve motifs, (4) running GRNBoost2 to infer gene regulatory net-
515 works, (5) pruning networks using cisTarget, (6) calculating regulon activity with AUCell, (7)
516 extracting accessibility data from ATAC-seq, (8) filtering predicted targets using ATAC-seq ac-
517 cessibility, (9) analyzing activity patterns across cell types, developmental stages, and anatomical
518 regions, and (10) summarizing findings.
519 Manual verification To evaluate whether the aggregated findings are truly reflected by the data or
520 merely simulated or hallucinated by the LLM, manual (human) verification was conducted follow-
521 ing the traces and codes generate by Biomni.
522 Wetlab Benchmark Development and Evaluation. A comprehensive benchmark was developed
523 consisting of 20 open-ended cloning questions curated from real-world applications to represent
524 the diversity and complexity of molecular cloning tasks across four major categories: Golden
525 Gate assembly, Gibson assembly, restriction enzyme cloning, and Gateway cloning. Each cate-
526 gory included both single-construct and pooled cloning scenarios. Additionally, the benchmark
527 incorporated common validation methods, including diagnostic restriction digestion, Sanger se-
528 quencing primer design, and sequence alignment analysis. For establishing baseline performance,
529 three human experts with extensive experience in molecular cloning were recruited. These experts
530 were instructed to complete each task without utilizing language models but were permitted to use
531 standard molecular biology tools, search engines, and publicly available online resources such as
532 plasmid repositories and primer design platforms. The time required for each expert to complete
533 each task was recorded, from initial task understanding to the final protocol and plasmid map gen-
534 eration. In parallel, Biomni and general LLM models were evaluated on identical tasks. Each
535 system was provided with the same task descriptions and required to generate detailed end-to-end

23
536 experimental protocols and final cloned plasmid maps. For general LLMs, Claude 3.7 was used as
537 one of the most capable publicly-available models at the time of testing, providing it with the same
538 information but without access to specialized molecular biology tools. For evaluation, an indepen-
539 dent senior researcher with experience in molecular cloning technologies was recruited and blinded
540 to the source of each protocol (human expert, Biomni, or general LLM). The evaluator assessed
541 each protocol and plasmid map based on two primary criteria: (1) Accuracy: The correctness of
542 the proposed methodology, including appropriate enzyme selection, reaction conditions, primer
543 design parameters, and plasmid construction strategy. (2) Completeness: The thoroughness of the
544 protocol, including all necessary steps, reagents, concentrations, incubation times, and verification
545 methods. Each criterion was scored on a scale of 1-5 according to a detailed rubric (Supplementary
546 Table S31-32). The average scores across all 20 tasks were calculated for each system and human
547 expert to enable direct comparison.
548 Wetlab Validation. A practical cloning task was selected for validation: the insertion of a guide
549 RNA targeting the human B2M gene into the lentiCRISPR v2 Blast construct. This task was
550 chosen for its relevance to CRISPR-based gene editing applications and its moderate complexity,
551 involving multiple molecular biology techniques. The experiment was conducted in a standard
552 molecular biology laboratory setting using commercially available reagents and materials. The
553 lentiCRISPR v2 Blast plasmid was obtained from Addgene. All protocols for the experiment were
554 generated entirely by Biomni without modification (Supplementary Notes F), including plasmid
555 analysis, sgRNA design, oligo design with appropriate overhangs, detailed Golden Gate assembly
556 conditions, bacterial transformation parameters, and verification strategies. For validation of the
557 cloning results, standard molecular biology practices were followed, selecting colonies for cul-
558 ture and miniprep, followed by Sanger sequencing using the Biomni-designed primers. Sequence
559 alignment analysis was performed to verify the correct insertion of the sgRNA sequence. The
560 success of the cloning process was determined by the presence of bacterial colonies on selective
561 media and subsequent sequence verification confirming the accurate incorporation of the designed
562 sgRNA construct into the lentiCRISPR v2 Blast backbone.

24
563 Data availability. All data used in Biomni are publicly available at Harvard Dataverse under
564 https://doi.org/10.7910/DVN/CE4ZYG.

565 Code availability. Biomni is open-sourced at https://github.com/snap-stanford/biomni. A web-

566 based user interface is available at https://biomni.stanford.edu. Note that the public tool is not for
567 protected health information.

568 Acknowledgements. We thank Emily Alsentzer, Andrew Lee, members of Jure Leskovec’s lab,
569 and members of Euan Ashley’s lab, for providing helpful feedbacks. K.H. and J.L. also gratefully
570 acknowledge the support of NSF under Nos. OAC-1835598 (CINES), CCF-1918940 (Expedi-
571 tions), DMS-2327709 (IHBEM), IIS-2403318 (III); Stanford Data Applications Initiative, Wu Tsai
572 Neurosciences Institute, Stanford Institute for Human-Centered AI, Chan Zuckerberg Initiative,
573 Amazon, Genentech, GSK, Hitachi, SAP, and UCB. K.H. acknowledge the support of Stanford
574 Bio-X fellowship. Research reported in this publication was supported by the National Institute
575 of Neurological Disorders and Stroke of the National Institutes of Health under Award Number
576 U01NS134358. The content is solely the responsibility of the authors and does not necessarily
577 represent the official views of the National Institutes of Health.

578 Authors contribution. K.H., Y.R., J.L. conceived the study. K.H. and J.L. supervised the project.
579 K.H. designed and developed the framework. K.H., S.Z., H.W., Y.Q., Y.L. implemented tools
580 and databases. K.H. designed and implemented the generalist agent architecture. K.H. and R.L.
581 designed the action discovery agent. S.Z. performed benchmarks on Q&A tasks. K.H., H.W., Y.L.
582 collected and implemented benchmarks on realistic tasks. X.Z. provided advice on microbiome
583 benchmark. H.W., J.Z., P.H., K.H. performed multi-omics integration case study. Y.L., K.H.
584 performed wearable data analysis case study. Y.Q., J.Z., D.Y., S.Z., Y.L., K.H. performed wet-
585 lab case study. K.H., S.M., J.C., M.W., J.B. performed rare disease diagnosis case study. R.L.
586 performed qualitative trace analysis. R.L., L.Q., G.L., provided support for software. K.H., S.Z.,
587 H.W., Y.Q., A.R., Y.L. wrote the draft paper. All authors discussed the results and contributed to
588 the final manuscript.

589 Competing interests. A.R. and H.W. are employees of Genentech and A.R. has equity in Roche.
590 All other authors declare no competing interests.

25
591 References
592

593 1. Cong, L. et al. Multiplex genome engineering using crispr/cas systems. Science 339, 819–823
594 (2013).
595 2. Jumper, J. et al. Highly accurate protein structure prediction with alphafold. nature 596,
596 583–589 (2021).
597 3. Van Dyck, C. H. et al. Lecanemab in early alzheimers disease. New England Journal of
598 Medicine 388, 9–21 (2023).
599 4. López-Otı́n, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging:
600 An expanding universe. Cell 186, 243–278 (2023).
601 5. Botvinik-Nezer, R. et al. Variability in the analysis of a single neuroimaging dataset by many
602 teams. Nature 582, 84–88 (2020).
603 6. Thiele, I. & Palsson, B. Ø. A protocol for generating a high-quality genome-scale metabolic
604 reconstruction. Nature protocols 5, 93–121 (2010).
605 7. Gibney, E. & Van Noorden, R. Scientists losing data at a rapid rate. Nature 10 (2013).
606 8. Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60
607 (2023).
608 9. Ridnik, T., Kredo, D. & Friedman, I. Code generation with alphacodium: From prompt engi-
609 neering to flow engineering. arXiv preprint arXiv:2401.08500 (2024).
610 10. Cui, J., Li, Z., Yan, Y., Chen, B. & Yuan, L. Chatlaw: Open-source legal large language model
611 with integrated external knowledge bases. CoRR (2023).
612 11. Tom, G. et al. Self-driving laboratories for chemistry and materials science. Chemical Reviews
613 124, 9633–9732 (2024).
614 12. Peng, C. et al. A study of generative large language model for medical research and healthcare.
615 NPJ digital medicine 6, 210 (2023).
616 13. Qu, Y. et al. Crispr-gpt: An llm agent for automated design of gene-editing experiments.
617 bioRxiv 2024–04 (2024).
618 14. Swanson, K., Wu, W., Bulaong, N. L., Pak, J. E. & Zou, J. The virtual lab: Ai agents design
619 new sars-cov-2 nanobodies with experimental validation. bioRxiv 2024–11 (2024).
620 15. Roohani, Y. et al. Biodiscoveryagent: An ai agent for designing genetic perturbation experi-
621 ments. International Conference on Learning Representations (2025).
622 16. Wang, E. et al. Txgemma: Efficient and agentic llms for therapeutics. arXiv preprint
623 arXiv:2504.06196 (2025).
624 17. Xiao, Y. et al. Cellagent: An llm-driven multi-agent framework for automated single-cell data
625 analysis. BioRxiv 2024–05 (2024).
626 18. Youngblut, N. D. et al. scbasecamp: an ai agent-curated, uniformly processed, and continually
627 expanding single cell data repository. bioRxiv 2025–02 (2025).

26
628 19. Hu, M. et al. Evaluation of large language models for discovery of gene set function. Nature
629 methods 1–10 (2024).
630 20. Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Advances
631 in neural information processing systems 35, 24824–24837 (2022).
632 21. Yao, S. et al. React: Synergizing reasoning and acting in language models. In International
633 Conference on Learning Representations (ICLR) (2023).
634 22. Guo, D. et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement
635 learning. arXiv preprint arXiv:2501.12948 (2025).
636 23. Phan, L. et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249 (2025).
637 24. Laurent, J. M. et al. Lab-bench: Measuring capabilities of language models for biology re-
638 search. arXiv preprint arXiv:2407.10362 (2024).
639 25. To, K. et al. A multi-omic atlas of human embryonic skeletal development. Nature 635,
640 657–667 (2024).
641 26. Gordon, J. A. et al. Pbx1 represses osteoblastogenesis by blocking hoxa10-mediated recruit-
642 ment of chromatin remodeling factors. Molecular and cellular biology (2010).
643 27. Gomez, G. A. et al. Evaluation of potential roles of zinc finger homeobox 3 (zfhx3) expressed
644 in chondrocytes and osteoblasts on skeletal growth in mice. Calcified Tissue International
645 115, 445–454 (2024).
646 28. Geng, Z., Tai, Y. T., Wang, Q. & Gao, Z. Auts2 disruption causes neuronal differentiation
647 defects in human cerebral organoids through hyperactivation of the wnt/β-catenin pathway.
648 Scientific reports 14, 19522 (2024).
649 29. Wang, X. et al. Executable code actions elicit better llm agents. In Forty-first International
650 Conference on Machine Learning (2024).
651 30. Ghoussaini, M. et al. Open targets genetics: systematic identification of trait-associated genes
652 using large-scale genetics and functional genomics. Nucleic acids research 49, D1311–D1320
653 (2021).
654 31. Shringarpure, S. S. et al. Large language models identify causal genes in complex trait gwas.
655 medRxiv 2024–05 (2024).
656 32. Schmidt, R. et al. Crispr activation and interference screens decode stimulation responses in
657 primary human t cells. Science 375, eabj4008 (2022).
658 33. Zhang, H., Kang, Z., Zhang, Y., Yang, R. & Ning, K. Towards a generative paradigm for
659 large-scale microbiome analysis by generative language model. bioRxiv 2025–01 (2025).
660 34. Turnbaugh, P. J. et al. A core gut microbiome in obese and lean twins. nature 457, 480–484
661 (2009).
662 35. Pasolli, E., Truong, D. T., Malik, F., Waldron, L. & Segata, N. Machine learning meta-analysis
663 of large metagenomic datasets: tools and biological insights. PLoS computational biology 12,
664 e1004977 (2016).

27
665 36. Gomez-Alvarez, V. & Revetta, R. P. Monitoring of nitrification in chloraminated drinking
666 water distribution systems with microbiome bioindicators using supervised machine learning.
667 Frontiers in Microbiology 11, 571009 (2020).
668 37. Turnbaugh, P. J. et al. The human microbiome project. Nature 449, 804–810 (2007).
669 38. Huang, K. et al. A foundation model for clinician-centered drug repurposing. Nature Medicine
670 30, 3601–3613 (2024).
671 39. Alsentzer, E. et al. Few shot learning for phenotype-driven diagnosis of patients with rare
672 genetic diseases. medRxiv 2022–12 (2022).
673 40. Metwally, A. A. et al. Prediction of metabolic subphenotypes of type 2 diabetes via continuous
674 glucose monitoring and machine learning. Nature Biomedical Engineering 1–18 (2024).
675 41. Katz, D. et al. The multi-omic, multi-tissue response to acute endurance and resistance exer-
676 cise: Results from the molecular transducers of physical activity consortium. Circulation 150,
677 A4143199–A4143199 (2024).

28
678 Supplementary materials
679 Supplementary Notes
680 A Details about baselines
681 We use the following baselines for benchmark comparison: (1) a base LLM without tools, (2) Re-
682 Act, using function-calling via chain-of-thought reasoning, (3) ReAct+Code, which adds Python
683 code execution capabilities, (4) ReAct+Literature, which includes access to PubMed, web search,
684 PDF extraction from URLs, and ArXiv papers, (5) ReAct+Code+Literature, combining both struc-
685 tured and unstructured tool resources, and (6) Biomni-ReAct, an ablation of Biomni that shares the
686 full environment but replaces Biomni-A1s code-based planning with standard ReAct-style tool
687 chaining.

688 B Details about real-world benchmark

689 In the variant prioritization benchmark, curated from Open Targets Genetics30 , the agent must
690 identify the top disease-relevant variant from a GWAS-linked candidate set. Biomni achieved
691 a 78.9% gain over the base LLM and outperformed ReAct+Code (+162%) and Biomni-ReAct
692 (+3.0%), highlighting its capacity to reason over regulatory variant relevance in noncoding regions.
693 In GWAS causal gene detection, adapted from Shringarpure et al.31 , Biomni was asked to
694 select the most likely causal gene from among candidates within a locus. It achieved a 19.4% gain
695 over base LLMs and modest improvements over both ReAct+Code (+54.2%) and Biomni-ReAct
696 (+2.7%), showing superior locus-level inference and granular reasoning.
697 In the CRISPR perturbation screen design task, adatped from Roohani et al.15 using data from
698 Schmidt et al.32 , Biomni generated gene panels to maximize the experimental effect. It improved
699 post-perturbation performance by 13.2% over ReAct+Code and 23.0% over Biomni-ReAct, de-
700 spite the large gene space (∼19k genes), showcasing its ability to perform experimental planning
701 strategically under strong constraints.
702 In patient gene prioritization and rare disease diagnosis, using datasets from Alsentzer et
703 al.39 , Biomni mapped patient phenotypes and genetic findings to causal genes and diagnoses,
704 achieving 19.4% and 10.3% gains over base LLMs, and outperforming Biomni-ReAct by 5.7%
705 and 10.3%, respectively-demonstrating robustness in clinically grounded reasoning.
706 In drug repurposing, adapted from TxGNN38 , Biomni selected candidate drugs aligned with
707 EHR-based off-label prescription patterns. Though the alignment gains over base LLMs were mod-
708 est (+3.6%), Biomni remained competitive with both ReAct+Code and Biomni-ReAct, indicating
709 its ability to handle subtle pharmacological signal amid high uncertainty.
710 In single-cell RNA-seq cell annotation, where the task involved assigning cell types across

29
711 species, tissues, and platforms, Biomni achieved a 93.6% improvement over Biomni-ReAct and
712 10.3% over ReAct+Code, based on semantic matching verified by human adjudication-demonstrating
713 highly accurate biological labeling in diverse contexts.
714 Finally, in microbiome disease-taxa analysis, Biomni autonomously performed statistical
715 testing and visualization on five metagenomic datasets from public datasets, like Human Micro-
716 biome Project37 and literature. It outperformed the base LLM by 180.0%, ReAct+Code by 48.8%,
717 and Biomni-ReAct by 25.0%, underscoring its strength in executing complex, noisy workflows
718 end to end 41 .
719 C Biomni prompt
System prompt for the generalist agent

You are a helpful biomedical assistant assigned with the task of

problem-solving.
To achieve this, you will be using an interactive coding
environment equipped with a variety of tool functions, data,
and softwares to assist you throughout the process.

Given a task, make a plan first. The plan should be a numbered

list of steps that you will take to solve the task. Be specific
and detailed.
Format your plan as a checklist with empty checkboxes like this:
1. [ ] First step
2. [ ] Second step
3. [ ] Third step

Follow the plan step by step. After completing each step, update
the checklist by replacing the empty checkbox with a checkmark:
1. [ ] First step (completed)
2. [ ] Second step
3. [ ] Third step

If a step fails or needs modification, mark it with an X and

explain why:
1. [ ] First step (completed)
2. [ ] Second step (failed because...)
3. [ ] Modified second step
4. [ ] Third step

At each turn, you should first provide your detailed thinking and
reasoning given the conversation history, along with the
updated plan (Always show the updated plan after each step so
720

30
 the user can track progress).
After that, you have two options:

1) Interact with a programming environment and receive the

corresponding output within <observe></observe>. Your code
should be enclosed using "<execute>" tag, for example: <execute
> print("Hello World!") </execute>. IMPORTANT: You must end the
code block with </execute> tag.
- For Python code (default): <execute> print("Hello World!") </
execute>
- For R code: <execute> #!R\nlibrary(ggplot2)\nprint("Hello
from R") </execute>
- For Bash scripts and commands: <execute> #!BASH\necho "Hello
from Bash"\nls -la </execute>
- For CLI softwares, use Bash scripts.

2) When you think it is ready, directly provide a solution that

adheres to the required format for the given task to the user.
Your solution should be enclosed using "<solution>" tag, for
example: The answer is <solution> A </solution>. IMPORTANT: You
must end the solution block with </solution> tag.
- If user does not specify the format, use a report format. In
the report, include the result and also a summary of how
you solved the problem. Make it concise and to the point.
Be rigorous.
- Use numbered references like [1], [2] in the summary if
applicable. Provide brief footnotes for each reference at
the end of the report, explaining the rationale or evidence
.

You have many chances to interact with the environment to receive

the observation. So you can decompose your code into multiple
steps.
Don’t overcomplicate the code. Keep it simple and easy to
understand.
When writing the code, please print out the steps and results in a
clear and concise manner, like a research log.
When calling the existing python functions in the function
dictionary, YOU MUST SAVE THE OUTPUT and PRINT OUT the result.
For example, result = understand_scRNA(XXX) print(result)
Otherwise the system will not be able to know what has been done.
Don’t overdo it. Stop when the plan is finished or the task is
already solved. Be relatively simple and concise and
understandable to the user.
Also, avoid faking or simulating code/data. Your user is a
721

31
 biomedical researcher. Thus, stay true and rigorous.
For the thinking process, put before the execute code block. Do
not use print statement in the execute code block for the
thinking process.
If you draw figures, make publication-ready and beautiful figures.

For R code, use the #!R marker at the beginning of your code block
to indicate it’s R code.
For Bash scripts and commands, use the #!BASH marker at the
beginning of your code block. This allows for both simple
commands and multi-line scripts with variables, conditionals,
loops, and other Bash features.

In each response, you must include EITHER <execute> or <solution>

tag. Not both at the same time. Do not respond with messages
without any tags. No empty messages. In each response, there
could ONLY BE ONE TAG. Not even mention the other tag in your
response, since it will cause error. In each response, for the
tag, also just use once, not multiple times.

Try to save all generated files or images to the ’/tmp/

agent_outputs/’ directory.
IMPORTANT: Report and print the exact absolute path in the <
observation> block so the system can find and potentially
display it (e.g. print(f"Image saved at: path/to/image.png")).

Environment Resources:

- Function Dictionary:
{function_intro}
---
{tool_desc}
---

{import_instruction}

- Biological Data Lake (Amazon S3):

The biological data lake is stored in an Amazon S3 bucket: {
s3_datalake_uri}
{data_lake_intro}
You need to use appropriate tools/libraries to access files from
this S3 bucket within your code. Assume necessary AWS
credentials are configured in the execution environment.
----
Available Files (use these paths relative to the bucket URI):
722

32
 {data_lake_content}
----
- Software Library:
{library_intro}
Each library is listed with its description to help you understand
its functionality.
----
{library_content_formatted}
----
- Note on using R packages and Bash scripts:
- R packages: Use subprocess.run([’Rscript’, ’-e’, ’your R code
here’]) in Python, or use the #!R marker in your execute
block.
- Bash scripts and commands: Use the #!BASH marker in your
execute block for both simple commands and complex shell
scripts with variables, loops, conditionals, etc.
723

System prompt for the tool retriever LLM

You are an expert biomedical research assistant. Your task is to

select the relevant resources to help answer a user’s query.

USER QUERY: {query}

Below are the available resources. For each category, select items
that are directly or indirectly relevant to answering the
query.
Be generous in your selection - include resources that might be
useful for the task, even if they’re not explicitly mentioned
in the query.
It’s better to include slightly more resources than to miss
potentially useful ones.

AVAILABLE TOOLS:
{self._format_resources_for_prompt(resources.get(’tools’, []))}

AVAILABLE DATA LAKE ITEMS:

{self._format_resources_for_prompt(resources.get(’data_lake’, []))
}

AVAILABLE SOFTWARE LIBRARIES:

{self._format_resources_for_prompt(resources.get(’libraries’, []))
}

724

33
 For each category, respond with ONLY the indices of the relevant
items in the following format:
TOOLS: [list of indices]
DATA_LAKE: [list of indices]
LIBRARIES: [list of indices]

For example:
TOOLS: [0, 3, 5, 7, 9]
DATA_LAKE: [1, 2, 4]
LIBRARIES: [0, 2, 4, 5, 8]

If a category has no relevant items, use an empty list, e.g.,

DATA_LAKE: []

IMPORTANT GUIDELINES:
1. Be generous but not excessive - aim to include all potentially
relevant resources
2. ALWAYS prioritize database tools for general queries - include
as many database tools as possible
3. Include all literature search tools
4. For wet lab sequence type of queries, ALWAYS include molecular
biology tools
5. For data lake items, include datasets that could provide useful
information
6. For libraries, include those that provide functions needed for
analysis
7. Don’t exclude resources just because they’re not explicitly
mentioned in the query
8. When in doubt about a database tool or molecular biology tool,
include it rather than exclude it
725

System prompt for the action discovery agent

You are a research methodology expert specializing in identifying

computational tasks and data analysis procedures in academic
papers.

Your job is to analyze chunks of academic papers and identify ONLY

the most common, generalizable computational tasks that are
widely used across biomedical research and can be implemented
with Python or Linux code.

STRICT GUIDELINES:
1. ONLY extract tasks that are extremely common and standard in
726

34
 computational biomedical research
2. Each task MUST have clear, well-defined inputs and outputs
3. Tasks MUST be generalizable across many different papers and
research questions
4. Be VERY selective - only include tasks that appear in hundreds
of papers
5. If a task is specific to this paper, unclear, or not widely
used, DO NOT include it
6. Focus on computational tasks that can be implemented with
Python or Linux code
7. Each task should be something that could be implemented as a
function with clear inputs/outputs
8. Also identify commonly used databases and software packages
mentioned in the text
9. Tasks MUST be CONCRETE and SPECIFIC - include exact
methodological details
10. Avoid vague task names like "Statistical Analysis" - instead
use specific protocol names like "Two-way ANOVA with Tukey’s
Post-hoc Test using SciPy"
11. DO NOT include wet lab procedures that cannot be implemented
computationally
12. ONLY include tasks that could be automated with code

For the following chunk of text from an academic paper, provide:

1. A list of ONLY the most common, generalizable COMPUTATIONAL
tasks identified (be extremely selective)
2. For each task, clearly define:
- Task name: A SPECIFIC and CONCRETE name with methodological
details (e.g., "RNA-seq Differential Expression Analysis
with DESeq2" instead of just "Gene Expression Analysis")
- Input: What SPECIFIC data or parameters the task requires
- Output: What SPECIFIC data or results the task produces
- Code implementation: How this task could be implemented with
Python or Linux code, including key libraries/packages
- Frequency: How common this computational task is in
biomedical research
- Standard methods: The established computational techniques
used to perform this task
- Example: A brief description of how THIS specific paper uses
this task (with specific details from the paper)
3. A list of commonly used databases mentioned in the text (if any
)
4. A list of commonly used software packages/tools mentioned in
the text (if any)

727

35
 PAPER CHUNK:
{chunk_text}

Remember, it’s better to return NO tasks than to include tasks

that aren’t extremely common, generalizable, and implementable
with code. Quality over quantity is essential. Tasks MUST be
CONCRETE with SPECIFIC methodological details and MUST be
implementable with Python or Linux code.
728

729 D Biomni generated report for the wearable case study

User Prompt for CGM Wearable Temperature Data Analysis

Task: Here I Have some CGM thermo data from participants, I

recorded their body temperature after they took certain type
food, and used the device to record the data
Dataset: CGM thermo data in the format of csv
Path: {data_path}
Objective: Identify how each participants body temperature change
after they took food, and also see individual difference after
taking the food
Expected Output: plot the graphs please
Output Format:save figures in {save_path}
730

Agent Output for CGM Wearable Temperature Data Analysis

# CGM Thermo Data Analysis Summary

## Analysis Overview:
This analysis examined temperature data from CGM devices to
identify how body temperature changes after food intake and to
analyze individual differences in temperature responses.

## Methodology:
1. Temperature data was analyzed to identify potential meal times
based on significant temperature increases.
2. For each potential meal, a 6-hour window (2 hours before to 4
hours after) was analyzed to characterize the temperature
response.
3. Temperature changes were normalized relative to the pre-meal
baseline to enable comparison across subjects.

## Key Findings:
731

36
 - Average temperature increase after meals: 2.19 C
- Median temperature increase: 1.10 C
- Range of temperature increases: -0.11 C to 15.56 C

Individual Differences:
- Substantial variation was observed in temperature responses
across subjects.
- Some subjects consistently showed larger temperature increases
after meals than others.
- The timing of peak temperature also varied across subjects.

## Limitations:
- Without actual meal timing data, potential meal times were
inferred from temperature patterns.
- Some identified ’meals’ may be temperature increases due to
other factors (exercise, environment, etc.).
- The analysis does not account for different meal types (high
carb, high fat, high protein).

## Generated Visualizations:
1. Daily temperature analysis plots
2. Meal window plots
3. Box plot comparing temperature increases across subjects
4. Bar chart of average temperature increases by subject
5. Average temperature response pattern after meals
6. Individual differences in temperature response patterns
7. Heatmap of temperature changes by subject
8. Distribution of temperature increases
732

User Prompt for CGM Wearable Multi-Omics Data Analysis

Task: Here I Have some CGM omics from participants,can you help me
fully explore the omics data and generate some new insights"
Dataset: CGM omics data in the format of csv
Path: {data_path}
Objective: fully explore omics data and generate biological
insights
Expected Output: plot the graphs please and write a insightful
summary
Output Format:save figures in {save_path}
733

37
 Agent Output for CGM Wearable Multi-Omics Data Analysis

# CGM Omics Data Analysis: Key Biological Insights

## Dataset Overview
The analysis explored three distinct omics datasets from CGM
participants:
- **Lipidomics**: 652 features across 147 samples
- **Metabolomics**: 731 features across 147 samples
- **Proteomics**: Various proteins measured across multiple
samples

## Key Findings

### 1. Multi-omics Integration

The integrated analysis revealed coordinated patterns across
lipids, metabolites, and proteins, suggesting complex
biological processes that span multiple molecular layers.
Principal Component Analysis (PCA) showed distinct clustering
patterns in each dataset, with the first two components
explaining significant variance:
- Lipidomics: PC1 (24.15%), PC2 (15.34%)
- Metabolomics: PC1 (33.68%), PC2 (11.12%)
- Proteomics: PC1 (21.82%), PC2 (11.99%)

### 2. Potential Biomarkers

Several molecular features were identified as potential biomarkers
based on their high variance and correlation patterns. These
features likely represent key regulatory nodes in molecular
networks and could serve as targets for future studies.

### 3. Temporal and Seasonal Patterns

Temporal analysis revealed seasonal variations in molecular
profiles, suggesting environmental or lifestyle influences on
human physiology. These patterns may reflect adaptations to
changing conditions, dietary shifts, or activity patterns
throughout the year.

### 4. Molecular Networks

Correlation network analysis identified modules of co-regulated
features within each omics layer, likely representing
functional units or biological pathways. These networks provide
a systems-level view of molecular relationships and highlight
potential regulatory mechanisms.

### 5. Cross-omics Relationships

734

38
 Strong correlations were observed between features across
different omics layers, indicating coordinated regulation and
biological pathways that span multiple molecular types. These
relationships help understand the interplay between different
molecular classes in health and disease.

## Future Directions
1. **Functional Validation**: Validate identified biomarkers in
targeted studies
2. **Pathway Analysis**: Map features to specific biological
pathways
3. **Clinical Integration**: Integrate omics data with clinical
measurements
4. **Longitudinal Studies**: Follow individuals over time to
understand dynamic changes
5. **Machine Learning Models**: Develop predictive models for
early disease markers

The comprehensive analysis provides a foundation for understanding

the complex molecular landscape of CGM participants and offers
insights into potential biomarkers and biological mechanisms
that could inform future clinical applications.
735

User Prompt for CGM Wearable Sleep Data Analysis

Task: Here I have some sleep data from wearables from participants
,can you help me fully explore the omics data and generate some
new insights"
Dataset: sleep data from wearables in the format of csv
Path: {data_path}
Objective: delve deep into sleep data and generate biological
insights
Expected Output: plot the graphs please and write a insightful
summary
Output Format:save figures in {save_path}
736

Agent Output for CGM Wearable Sleep Data Analysis

# Sleep Data Analysis: Insights and Findings

## Overview
This analysis explored sleep data from wearable devices across 10
participants, covering 227 nights of sleep. The dataset
737

39
 captured detailed sleep metrics including duration, efficiency,
sleep stages (light, deep, REM), and various other parameters.
Through comprehensive analysis, we’ve identified several key
patterns and biological insights.

## Key Findings

### 1. Sleep Duration and Efficiency

- **Average sleep duration**: 7.82 hours, which falls within the
recommended 7-9 hours for adults
- **Sleep efficiency**: Mean of 86.86%, indicating generally good
sleep quality (>85% is considered efficient)
- **Individual variation**: Significant differences between
subjects, with average sleep duration ranging from 6.13 to 9.78
hours
- **Consistency**: 33.48% of nights had less than 7 hours of sleep
, while 49.78% had more than 8 hours

### 2. Sleep Architecture

- **Sleep stage composition**: On average, participants spent
59.09% in light sleep, 14.88% in deep sleep, and 19.02% in REM
sleep
- **Deep sleep patterns**: Deep sleep percentage ranged from
11.15% to 20.88% across subjects, with subject 43883-083
showing notably high deep sleep
- **REM sleep variation**: REM sleep percentage varied from 14.79%
to 23.26% across subjects, with subject 43883-087 showing the
highest REM percentage
- **Sleep latency**: Average time to fall asleep was remarkably
low at 4.13 minutes (below 10 minutes may indicate sleep
deprivation)

### 3. Sleep Quality Patterns

- **Quality scoring**: Our custom sleep quality score (
incorporating efficiency, sleep stages, latency, and wake-ups)
showed consistent patterns across subjects
- **Wake episodes**: Participants experienced an average of 24.23
wake episodes per night, which is relatively high
- **Sleep stage correlations**: Deep sleep percentage showed the
strongest positive correlation with overall sleep quality
- **Duration-quality relationship**: Moderate positive correlation
between sleep duration and quality score, with optimal quality
occurring around 7.5-8.5 hours

### 4. Chronobiological Insights

738

40
 - **Weekday vs. weekend**: Surprisingly, no statistically
significant difference between weekday and weekend sleep
duration (p=0.52)
- **Social jetlag**: Most subjects showed mild social jetlag (
shift in sleep timing between weekdays and weekends), with an
average of 0.49 hours
- **Sleep timing consistency**: Subjects with more consistent
sleep onset times tended to have better sleep quality
- **Day-of-week patterns**: Sleep efficiency was highest on
Wednesdays and lowest on Sundays, suggesting potential end-of-
weekend disruption

### 5. Individual Sleep Profiles

- **Subject 43883-079**: Longest average sleep duration (9.78
hours) but not the highest efficiency, suggesting potential
oversleeping
- **Subject 43883-074**: Shortest average sleep duration (6.13
hours) and lower efficiency (83.35%), indicating potential
sleep deficiency
- **Subject 43883-090**: Most efficient sleeper (90.21%) despite
moderate duration (8.14 hours)
- **Subject 43883-087**: Best REM sleep profile (23.26%) and
consistent sleep patterns
- **Subject 43883-083**: Highest deep sleep percentage (20.88%),
suggesting good physical recovery during sleep

## Biological Implications

1. **Sleep Homeostasis**: The data reveals individual differences

in sleep need and efficiency, supporting the concept of sleep
homeostasis varying between individuals.

2. **Sleep Architecture Balance**: The distribution of sleep

stages (light, deep, REM) appears relatively normal across most
subjects, with expected ratios for healthy adults.

3. **Circadian Rhythms**: The relatively low social jetlag

observed suggests well-aligned circadian rhythms for most
participants, though individual differences exist.

4. **Recovery Processes**: Deep sleep percentages (associated with

physical recovery) vary significantly between subjects,
potentially reflecting differences in physical activity or
metabolic demands.

739

41
 5. **Cognitive Processing**: REM sleep (associated with cognitive
processing and memory consolidation) shows notable variation,
which may relate to differences in cognitive demands or
emotional processing needs.

## Recommendations for Future Research

1. Incorporate additional biometric data (heart rate, physical

activity, stress levels) to correlate with sleep patterns

2. Collect subjective sleep quality ratings to compare with

objective measures

3. Extend data collection to longer periods to better capture

seasonal variations and long-term patterns

4. Analyze the impact of environmental factors (temperature, light

, noise) on sleep metrics

5. Investigate the relationship between sleep patterns and daytime

performance/cognitive function

This analysis demonstrates the power of wearable technology in

capturing detailed sleep metrics, providing insights into both
population-level sleep patterns and individual variations that
have important implications for health and well-being.
740

741 E Biomni generated report for the scRNA and scATAC case study
User Prompt for scRNA and scATAC Joint Data Analysis

Task G e n e r e g u l a t o r y n e t w o r k (GRN) analysis with pySCENIC

+ snATAC
Goal Map transcription factor (TF) circuits that drive
skeletal development across anatomical regions and
developmental stages.

DATA
snRNA seq : {PATH}/filtered_core_rna.h5ad
snATAC seq : {PATH}/filtered_core_atac.h5ad
Workdir : {PATH}/buffer_o3_run6/
pySCENIC database : {PATH}/pyscenic/
742

42
 All intermediates & logs must be written here.

1 LOAD & ALIGN

1.1 Read both .h5ad files into memory (AnnData).
1.2 Verify identical cell barcodes and c e l l t y p e labels.
1.3 Ensure gene/peak annotation uses the same genome build.

2 TF SEED LIST
2.1 Start with canonical skeletal TFs (SOX9, RUNX2, etc.).
2.2 Augment the list by:
Related database in data_lake
Differential expression (DE) across cell types.
Differential accessibility (DA) of promoter peaks. (Add
any TF with DE padj < 0.05 AND DA padj < 0.05.)

3 P E R C E L L T Y P E GRN INFERENCE
3.0 Filter cell types: keep only those with 500 nuclei (or
50 pseudobulks).
3.1 For each qualifying cell type **ct**:
a) adata_ct = adata[adata.obs["cell_type"] == ct]
b) ‘print(f"[{ct}] starting GRNBoost2 ")‘
c) Run GRNBoost2 save
‘{workdir}/grnboost2/raw/{ct}_adjacency.csv‘

4 MOTIF + ATAC PRUNING ( p e r c e l l t y p e )

4.1 Pseudobulk the matching snATAC cells of **ct** (min depth 20
k).
4.2 Keep an edge only if TF motif overlaps 1 accessible peak
in that pseudobulk.
4.3 Outputs: ‘ct_pruned_edges.csv‘, print out informative
messages of intermediate results.

5 REGULON BUILD
5.1 Convert pruned edges to regulons (.loom + .csv).
5.2 Discard regulons with <20 target genes.

6 REGULON ACTIVITY (AUCell)

6.1 Compute AUCell scores per nucleus.
6.2 Save AUCell.h5ad (scores + metadata).

7 COMPARATIVE ANALYSIS
7.0 Concatenate all AUCell matrices.
743

43
 7.1 Within each **ct**, test stage / region effects.
7.2 Across cell types, highlight TFs unique to a lineage.

8 VISUALISATION
8.1 UMAP colored by selected regulons.
8.2 Heatmap: regulon c e l l t y p e ( r o w z s c o r e d AUCell).
8.3 Line plots: stage trajectory of key TF activities.

9 REPORT (Output: report.md or report.pdf)

9.1 Executive overview ( 2 0 0 words)
Purpose, datasets, key methods, headline discoveries.

9.2 Key findings (tables / figures)

T o p 1 0 regulons by AUCell score *per lineage*
S t a g e s p e c i f i c TFs ( l o g 2 f o l d 1.5, FDR<0.05)
Region restricted regulons (limb vs. craniofacial, etc
.)
Novel TFs not in seed list, ranked by t a r g e t c o u n t .

9.3 Visual summary

UMAP coloured by 5 hallmark regulons.
Heatmap: Regulon c e l l t y p e ( r o w Z AUCell).
Bar plot: edge counts before/after ATAC pruning.

9.4 Method & QC appendix

Parameter table (motif DB, thresholds, software versions)
.
Runtime / memory stats.
E d g e r e t e n t i o n percentages for each pruning step.
744

Agent Output for scRNA and scATAC Joint Data Analysis

# Gene Regulatory Network Analysis in Skeletal Development

I’ve completed a comprehensive analysis of gene-regulatory-

networks (GRNs) in skeletal development using existing pySCENIC
and snATAC-seq data. The analysis maps transcription-factor (
TF) circuits across anatomical regions and developmental stages
.

## Analysis Overview

1. **Inventory**: Located all necessary files including adjacency

matrices, pruned edges, regulon tables, and AUCell data for 10
745

44
 distinct cell types.

2. **Data Loading**: Successfully loaded regulon data and AUCell

matrices, mapping cells to their corresponding cell types,
developmental stages, and anatomical regions.

3. **Quality Control**:
- Edge retention rates were consistently ˜10% across all cell
types
- All cell types had 566-589 regulons with at least 20 target
genes
- AUCell score distributions showed consistent patterns with
LimbMes exhibiting highest activity

4. **Biological Analysis**:
- Identified 458 differentially active regulons between early
and late developmental stages
- Found 4,177 differentially active regulons between anatomical
regions
- Mapped lineage-specific master regulators like SATB2 (cranial
mesenchyme), SOX5 (chondroprogenitors), and TWIST1 (
preosteoblasts)
- Discovered novel regulators including AUTS2, ZFHX3, and PBX1
with high activity across skeletal lineages
- Performed GO/KEGG enrichment for top regulons revealing
distinct biological processes for each cell type

5. **Visualizations**: Created a comprehensive set of figures

including:
- Heatmap of regulon activity across cell types
- UMAP embeddings colored by cell type and key regulon activity
- Bar plots of raw vs. pruned edge counts and retention rates

6. **Comprehensive Report**: Generated a detailed ‘report_full.md‘

containing:
- Executive summary of key findings
- QC analysis with tables and figures
- Key discoveries with visualizations
- Novel regulator spotlight
- Detailed methods section

## Key Discoveries

1. **Spatiotemporal Regulation**: Extensive differences in regulon

activity between developmental stages and anatomical regions,
746

45
 highlighting complex spatiotemporal control of skeletal
development.

2. **Lineage-Specific Programs**: Distinct regulatory programs for

each skeletal lineage, with both unique and shared regulons
between related cell types.

3. **Novel Regulators**: Identification of several transcription

factors not previously associated with skeletal development
that show high regulatory activity.

4. **Functional Enrichment**: Cell-type-specific enrichment

patterns, with osteoblast regulons enriched for bone
development, chondrocyte regulons for cartilage development,
and mesenchymal regulons for pattern specification.

The complete analysis provides new insights into the

transcriptional control of skeletal development and identifies
potential targets for future research into skeletal disorders
and regenerative medicine.
747

748 F Biomni generated report for the cloning protocol case study
User prompt for cloning protocol design

I have a plasmid plentiCRISPR v2 blast:

CGTTGGAGTCCACGTT...[full sequence removed for the clarity of the
manuscript]...CGCCCTTTGA. I hope to clone a CRISPR sgRNA
targeting human B2M into this plasmid. Could you give me the
final map of the plasmid and step-by-step guidance on how
should I perform the cloning?
749

Agent Output for cloning protocol design

# Cloning B2M sgRNA into lentiCRISPR v2 blast Plasmid

## Final Plasmid Map

The final plasmid (12,816 bp) will contain:
1. **Lentiviral Elements**:
- 5’ and 3’ LTRs for viral integration
- Psi ( ) packaging signal
- RRE (Rev Response Element)
- WPRE element for enhanced expression
750

46
 2. **CRISPR Components**:
- U6 promoter driving sgRNA expression
- B2M sgRNA: 5’-CGCGAGCACAGCTAAGGCCA-3’
- sgRNA scaffold for Cas9 binding
- EF1a promoter driving Cas9 expression
- Cas9 coding sequence

3. **Selection Markers**:
- Blasticidin resistance gene for mammalian selection
- Ampicillin resistance gene for bacterial selection

## Cloning Strategy
The B2M sgRNA will be inserted between the U6 promoter and sgRNA
scaffold using BsmBI restriction sites, replacing the original
stuffer sequence.

## Step-by-Step Cloning Protocol

### 1. Design and Order Oligos

- **Forward oligo**: 5’-CACCGCGCGAGCACAGCTAAGGCCA-3’
- **Reverse oligo**: 5’-AAACTGGCCTTAGCTGTGCTCGCGC-3’

### 2. Anneal and Phosphorylate Oligos

1. Resuspend oligos to 100 M in nuclease-free water
2. Prepare annealing reaction:
- 1 L Forward oligo (100 M)
- 1 L Reverse oligo (100 M)
- 1 L 10X T4 Ligation Buffer
- 6.5 L nuclease-free water
- 0.5 L T4 PNK
3. Incubate in a thermocycler:
- 37 C for 30 minutes (phosphorylation)
- 95 C for 5 minutes (denaturation)
- Ramp down to 25 C at 5 C /min (annealing)
4. Dilute annealed oligos 1:200 in nuclease-free water (final: 50
nM)

### 3. Golden Gate Assembly

1. Prepare reaction:
- 100 ng lentiCRISPR v2 blast plasmid
- 1 L diluted annealed oligos (50 nM)
- 2 L 10X T4 DNA Ligase Buffer
- 1 L BsmBI-v2 (NEB)
- 1 L T4 DNA Ligase
751

47
 - Nuclease-free water to 20 L total
2. Incubate in a thermocycler:
- 37 C for 5 minutes
- 16 C for 5 minutes
- Repeat steps 1-2 for 10 cycles
- 55 C for 5 minutes (final digestion)
- 80 C for 5 minutes (enzyme inactivation)

### 4. Bacterial Transformation

1. Add 2 L of the Golden Gate reaction to 50 L of competent
Stbl3 E. coli cells
2. Incubate on ice for 30 minutes
3. Heat shock at 42 C for 45 seconds
4. Return to ice for 2 minutes
5. Add 950 L of SOC medium
6. Incubate at 37 C for 1 hour with shaking
7. Plate 100 L on LB agar plates with ampicillin (100 g /mL)
8. Incubate plates at 37 C overnight

### 5. Colony Screening and Verification

1. Pick 3-5 colonies and grow in 5 mL LB + ampicillin medium
overnight
2. Extract plasmid DNA using a miniprep kit
3. Verify correct insertion by Sanger sequencing using the U6
promoter forward primer:
5’-GACTATCATATGCTTACCGT-3’
4. Confirm the presence of the B2M sgRNA sequence in the plasmid

### 6. Lentivirus Production and Cell Transduction

1. Transfect the verified plasmid into HEK293T cells along with
packaging plasmids
2. Collect virus-containing supernatant after 48-72 hours
3. Filter and concentrate the lentivirus if needed
4. Transduce target cells with the lentivirus
5. Select transduced cells with blasticidin (5-10 g /mL)
6. Verify B2M knockout by flow cytometry, Western blot, or genomic
sequencing

## Applications
- Generation of B2M knockout cell lines for immunology research
- Creating cells resistant to CD8+ T cell recognition
- Development of universal donor cells for cell therapy
- Control for CRISPR screens
752

48
 753 Supplementary Figures



%GGYVEG]









S



&MS 




MS
X
1
IX

IX

IX

QR
RI
00
RR

RR

RR

SR
7S

7S

S
M7

M7
1

1

QR

QR
00

00

&MS

&MS
Figure 1: Performance on HLE with different LLM as the engine. We found that Sonnet 3.7 has better performance than 3.5 for
both LLM and Biomni. However, interestingly, O1 has better performance than Sonnet 3.7 for the LLM but not for Biomni.

49
 

%GGYVEG]





X

V

M

V
QR
G

ZI

ZI
I%

MI

XVMI
&MS
M6

IXV

I
QR

S6

S6
X2

M2
&MS

QR
G
I%

&MS
M6
QR
&MS

Figure 2: Ablation performance on retriever. We found that retriever can significantly reduce context size and has significant gain
for Biomni.

Figure 3: Subject-level performance on HLE biomedicine track. Due to the small sample size, the conclusion may not be meaningful.

50
 a Scientific Aims
1. How do individual sleep architecture patterns (light, deep, REM)
vary over time and between subjects?
2. What are the main contributors to sleep quality differences
(e.g., efficiency, latency, wake events)?
3. Can consistent sleep timing or deviations (e.g., social jetlag)
serve as potential behavioral biomarkers for cognitive or
Wearable- 10 participants,
derived sleep covering 227
physiological states?
data nights of sleep.
1. Data 2. Sleep Metrics 3. Sleep Stage
b I have wearable-derived sleep data from Understanding and Extraction Decomposition
multiple participants collected nightly over
several weeks. Please explore each Organizing
participant's sleep pattern, assess quality
and consistency, and identify biologically
meaningful inter-individual differences.
4. Inter-Subject 5. Pattern 6. Sleep Quality
Comparison Discovery Scoring

c d Findings
1.On overage, participants slept 7.82 hours with 86.9%
efficiency, spending 59% in light sleep, 15% in deep
sleep, and 19% in REM.
2. Sleep quality scores, derived from latency, stage
balance, and wake events, revealed that consistency in
sleep timing correlated strongly with better sleep
quality.
3.No significant difference was found between
weekday and weekend sleep duration, though mild
social jetlag (~0.5 hours) was common.

Figure 4: Study design, analysis workflow, and key findings of wearable-derived sleep data exploration. (a) Overview of the dataset,
including wearable-derived sleep monitoring across 10 participants over 227 nights. (b) AI-assisted analysis framework, outlining
six key steps: data understanding and organization, sleep metrics extraction, sleep stage decomposition, inter-subject comparison,
pattern discovery, and sleep quality scoring. (c) Example outputs from data analysis, including sleep efficiency distributions, sleep
timing deviations, and clustering of sleep architecture patterns across individuals. (d) Summary of findings: participants averaged
7.82 hours of sleep with 86.9% efficiency, showing that consistent sleep timing was strongly associated with better sleep quality.
Mild social jetlag ( 0.5 hours) was observed without significant differences between weekday and weekend sleep durations.

51
 a Scientific Aims
1.How does the human body regulate thermogenic responses
following food intake at the molecular level?
2.What are the key lipids, metabolites, and proteins associated
with thermogenic responses to meals?
3.Are there seasonal or temporal patterns in lipidomics,
metabolomics, and proteomics data that influence thermogenic
CGM-Omics Activity Data regulation?

1. Data 2. Baseline 3. Omics

b Here I have some CGM omics Understanding Establishment
Integration
from participants, can you help &Statistical
and Organizing Analysis
me fully explore the omics data
and generate some insights?

4. Pathway & 5. Visualization & 6. Insight

Enrichment Reporting Generation
Analysis

c d Findings
1. The integrated CGM-omics analysis
revealed coordinated molecular patterns
across lipidomics, metabolomics, and
proteomics, with distinct clustering and
strong cross-omics relationships.
2. Several high-variance features were
identified as potential biomarkers, and
seasonal variations suggested
environmental or lifestyle impacts.

Figure 5: Study design, Analysis workflow, and key findings of CGM-omics data exploration. (a) Overview of the data types
used, including CGM-omics (lipidomics, metabolomics, proteomics) and activity data. (b) Biomni analysis workflow outlining
six key steps: data understanding and organization, baseline establishment and statistical analysis, omics integration, pathway and
enrichment analysis, visualization and reporting, and insight generation. (c) Example outputs from exploratory data analysis and
cross-omics integration, including t-SNE and correlation heatmap across omics layers. (d) Summary of major findings: coordinated
molecular patterns across omics datasets, identification of potential biomarkers, and discovery of seasonal and temporal influences
on thermogenic responses.

52
 Biomni
Biomni-ReAct

Biomni
Biomni-ReAct

Biomni
Biomni-ReAct

Figure 6: Statistics for the number of tools, datasets, and software used across all the real-world benchmark tasks.

53
 Biomni
Biomni-ReAct

Figure 7: Statistics for the length of the trajectory across all the real-world benchmark tasks.

Figure 8: Action category statistics across steps.

54
Figure 9: Detailed analysis on the tools, datasets, and software used for Gene Perturbation task.

55
Figure 10: Detailed analysis on the tools, datasets, and software used for GWAS causal gene detection task.

56
Figure 11: Detailed analysis on the tools, datasets, and software used for microbiome data analysis task.

57
Figure 12: Detailed analysis on the tools, datasets, and software used for patient causal gene detection task.

58
Figure 13: Detailed analysis on the tools, datasets, and software used for rare disease diagnosis task.

59
Figure 14: Detailed analysis on the tools, datasets, and software used for drug repurposing task.

60
Figure 15: Detailed analysis on the tools, datasets, and software used for scRNA annotation task.

61
Figure 16: Detailed analysis on the tools, datasets, and software used for variant prioritization task.

62
Figure 17: Biomni user interface.

63
754 Supplementary Tables
Table 1: Tool descriptions for biomni.literature

Tool Name Tool Description

fetch supplementary info from doi Fetches supplementary information for a paper given its
DOI and returns a research log.
query arxiv Query arXiv for papers based on the provided search query
and return formatted search results.
query scholar Query Google Scholar for papers based on the provided
search query and return the first search result.
query pubmed Query PubMed for papers based on the provided search
query and return formatted results.
search google Search the web using Google search and return results with
title, URL, and description.
extract url content Extract the text content of a webpage using requests and
BeautifulSoup, removing unwanted elements and format-
ting the result.
extract pdf content Extract text content from a PDF file given its URL.

64
 Table 2: Tool descriptions for biomni.biochemistry

Tool Name Tool Description

analyze circular dichroism spectra Analyzes circular dichroism (CD) spectroscopy data to
determine secondary structure and thermal stability of
biomolecules.
analyze rna secondary structure features Calculate numeric values for various structural features of
an RNA secondary structure in dot-bracket notation.
analyze protease kinetics Analyze protease kinetics data from fluorogenic peptide
cleavage assays, fits the data to Michaelis-Menten kinetics,
and determines key kinetic parameters (kcat, KM, and cat-
alytic efficiency).
analyze enzyme kinetics assay Performs in vitro enzyme kinetics assay and analyzes the
dose-dependent effects of modulators.
analyze itc binding thermodynamics Analyzes isothermal titration calorimetry (ITC) data to de-
termine binding affinity and thermodynamic parameters.
analyze protein conservation Perform multiple sequence alignment and phylogenetic
analysis to identify conserved protein regions.

Table 3: Tool descriptions for biomni.bioengineering

Tool Name Tool Description

analyze cell migration metrics Analyze cell migration metrics from time-lapse microscopy
images and generate quantitative measurements of cell
movement.
perform crispr cas9 genome editing Simulates CRISPR-Cas9 genome editing process including
guide RNA design, delivery, and analysis.
analyze calcium imaging data Analyze calcium imaging data to quantify neuronal activity
metrics including cell counts, event rates, decay times, and
signal-to-noise ratios.
analyze in vitro drug release kinetics Analyzes in vitro drug release kinetics from biomaterial for-
mulations and determines the best fitting kinetic model.
analyze myofiber morphology Quantifies morphological properties of myofibers in mi-
croscopy images of tissue sections.
decode behavior from neural trajectories Model neural activity trajectories and decode behavioral
variables from neural spiking data.
simulate whole cell ode model Simulate a whole-cell model represented as a system of or-
dinary differential equations (ODEs).

65
 Table 4: Tool descriptions for biomni.biophysics

Tool Name Tool Description

predict protein disorder regions Predicts intrinsically disordered regions (IDRs) in a protein
sequence using IUPred2A algorithm.
analyze cell morphology and cytoskeleton Quantifies cell morphology and cytoskeletal organization
from fluorescence microscopy images.
analyze tissue deformation flow Quantify tissue deformation and flow dynamics from mi-
croscopy image sequence.

Table 5: Tool descriptions for biomni.cancer biology

Tool Name Tool Description

analyze ddr network in cancer Analyze DNA Damage Response (DDR) network alter-
ations and dependencies in cancer samples by reconstruct-
ing the DDR network from genomic data, identifying dis-
ruptions, and analyzing dependencies between DDR path-
way components.
analyze cell senescence and apoptosis Analyze flow cytometry data to quantify senescent and
apoptotic cell populations from FCS files.
detect and annotate somatic mutations Detects and annotates somatic mutations in tumor samples
compared to matched normal samples using GATK Mutect2
for variant calling, GATK FilterMutectCalls for filtering,
and SnpEff for functional annotation.
detect and characterize structural variations Detects and characterizes structural variations (SVs) in
genomic sequencing data using LUMPY for SV detec-
tion followed by annotation with COSMIC and/or ClinVar
databases.
perform gene expression nmf analysis Performs Non-negative Matrix Factorization (NMF) on
gene expression data to extract metagenes and their asso-
ciated sample weights for tumor subtype identification.

Table 6: Tool descriptions for biomni.cell biology

Tool Name Tool Description

quantify cell cycle phases from microscopy Quantify the percentage of cells in each cell cycle phase
(G1, S, G2/M) using Calcofluor white stained microscopy
images by segmenting cells and analyzing their morpholog-
ical features.
quantify and cluster cell motility Quantify cell motility features from time-lapse microscopy
images and cluster cells based on motility patterns.
perform facs cell sorting Performs Fluorescence-Activated Cell Sorting (FACS) to
enrich cell populations based on fluorescence characteris-
tics.
analyze flow cytometry immunophenotyping Analyze flow cytometry data to identify and quantify spe-
cific cell populations based on surface markers.
analyze mitochondrial morphology and potential Quantifies metrics of mitochondrial morphology and mem-
brane potential from fluorescence microscopy images.

66
 Table 7: Tool descriptions for biomni.molecular biology

Tool Name Tool Description

annotate open reading frames Find all Open Reading Frames (ORFs) in a DNA sequence
using Biopython. Searches both forward and reverse com-
plement strands if specified.
annotate plasmid Annotate a DNA sequence using pLannotate’s command-
line interface to identify features such as genes, promoters,
and origins of replication.
get gene coding sequence Retrieves the coding sequence(s) of a specified gene from
NCBI Entrez.
get plasmid sequence Retrieves plasmid sequences from either Addgene or NCBI
based on the provided identifier.
align sequences Align short sequences (primers) to a longer sequence, al-
lowing for one mismatch. Checks both forward and reverse
complement strands.
pcr simple Simulate PCR amplification with given primers and se-
quence, returning products and binding details.
pcr complex multi primers Simulate PCR amplification with multiple primers, consid-
ering all possible primer combinations and their potential
products.
digest sequence Simulates restriction enzyme digestion of a DNA sequence
and returns the resulting fragments with their properties.
golden gate Simulate a GoldenGate cloning reaction with Type IIS re-
striction enzymes to predict assembly products.
oligo assembly Assemble two DNA sequences into an oligo with over-
hangs. Automatically detects overhang type and length.
gibson assembly Simulate a Gibson Assembly reaction to join DNA frag-
ments with overlapping regions.
find restriction sites Identifies restriction enzyme sites in a given DNA sequence
for specified enzymes.
find restriction enzymes Finds common restriction enzyme sites in a DNA sequence.
design primers with overhangs Design two primers to amplify a target sequence with op-
tional overhangs.
find sequence mutations Compare query sequence against reference sequence to
identify mutations.
get molecular cloning instructions Returns a dictionary containing molecular cloning instruc-
tions and important notes.
calculate element distances Calculate pairwise distances between elements on a DNA
sequence/plasmid, providing forward, reverse (for circular
sequences), and shortest path distances.

67
 Table 8: Continue tool descriptions for biomni.molecular biology

Tool Name Tool Description

design knockout sgrna Design sgRNAs for CRISPR knockout by searching pre-
computed sgRNA libraries for a specific gene.
design golden gate oligos Design complementary oligonucleotides with Type IIS re-
striction enzyme overhangs for Golden Gate assembly
based on restriction site analysis of the backbone.
get oligo annealing protocol Return a standard protocol for annealing complementary
oligonucleotides without phosphorylation.
get golden gate assembly protocol Return a customized protocol for Golden Gate assembly
based on the number of inserts and specific DNA sequences.
get bacterial transformation protocol Return a standard protocol for bacterial transformation with
detailed steps.
design primer Design a single primer within a given DNA sequence win-
dow based on GC content and melting temperature con-
straints.
design verification primers Design Sanger sequencing primers to verify a specific re-
gion in a plasmid, using existing primers when possible and
designing new ones as needed.

Table 9: Tool descriptions for biomni.genetics

Tool Name Tool Description

liftover coordinates Perform liftover of genomic coordinates between hg19 and
hg38 genome builds with detailed step-by-step explana-
tions.
bayesian finemapping with deep vi Performs Bayesian fine-mapping from GWAS summary
statistics using deep variational inference to compute pos-
terior inclusion probabilities and credible sets for putative
causal variants.
analyze cas9 mutation outcomes Analyzes and categorizes mutations induced by Cas9 at tar-
get sites, generating detailed statistics on mutation types.
analyze crispr genome editing Analyzes CRISPR-Cas9 genome editing results by compar-
ing original and edited sequences to identify mutations and
characterize edited loci.
simulate demographic history Simulate DNA sequences with specified demographic and
coalescent histories using msprime.
identify transcription factor binding sites Identifies binding sites for a specific transcription factor in a
genomic sequence using position weight matrices from the
JASPAR database.
fit genomic prediction model Fit a linear mixed model for genomic prediction using geno-
type and phenotype data.
perform pcr and gel electrophoresis Performs PCR amplification of a target transgene and visu-
alizes results using agarose gel electrophoresis.
analyze protein phylogeny Perform phylogenetic analysis on a set of protein sequences,
including multiple sequence alignment, tree construction,
and visualization.

68
 Table 10: Tool descriptions for biomni.genomics

Tool Name Tool Description

annotate celltype scRNA Annotate cell types in single-cell RNA-seq data based on
gene markers and transferred labels using LLM.
create scvi embeddings scRNA Creates scVI and scANVI embeddings for single-cell RNA-
seq data, training models and saving the results to a new
AnnData object.
create harmony embeddings scRNA Performs batch effect correction on single-cell RNA-seq
data using the Harmony algorithm and saves the integrated
embeddings.
get uce embeddings scRNA Generate UCE (Universal Cell Embeddings) for single-cell
RNA sequencing data to enable cell type identification and
mapping to reference datasets.
map to ima interpret scRNA Map cell embeddings from the input dataset to the Inte-
grated Megascale Atlas reference dataset using UCE em-
beddings for cell type annotation.
get rna seq archs4 Given a gene name, fetch and return RNA-seq expres-
sion data (transcripts-per-million) across tissues from the
ARCHS4 database.
get gene set enrichment analysis supported database list Returns a list of supported databases available for gene set
enrichment analysis
gene set enrichment analysis Perform enrichment analysis for a list of genes to identify
pathways, transcription factors, or other biological relation-
ships.
analyze chromatin interactions Analyze chromatin interactions from Hi-C data to iden-
tify enhancer-promoter interactions and topologically asso-
ciated domains (TADs).
analyze comparative genomics and haplotypes Perform comparative genomics and haplotype analysis on
multiple genome samples. Aligns genomes to a reference,
identifies variants, analyzes shared and unique genomic re-
gions, and determines haplotype structure.
perform chipseq peak calling with macs2 Perform ChIP-seq peak calling using MACS2 to identify
genomic regions with significant binding.
find enriched motifs with homer Find DNA sequence motifs enriched in genomic regions us-
ing the HOMER motif discovery software.
analyze genomic region overlap Analyze overlaps between two or more sets of genomic re-
gions and generate a research log summarizing the analysis.

69
 Table 11: Tool descriptions for biomni.immunology

Tool Name Tool Description

analyze atac seq differential accessibility Perform ATAC-seq peak calling and differential accessibil-
ity analysis using MACS2.
analyze bacterial growth curve Analyzes bacterial growth curve data to determine growth
parameters like doubling time, growth rate, and lag phase.
isolate purify immune cells Isolates and purifies immune cells from tissue samples and
returns a research log of the process.
estimate cell cycle phase durations Estimate cell cycle phase durations using dual-nucleoside
pulse labeling data and mathematical modeling.
track immune cells under flow Track immune cells under flow conditions and classify their
behaviors.
analyze cfse cell proliferation Analyze CFSE-labeled cell samples to quantify cell division
and proliferation from flow cytometry data.
analyze cytokine production in cd4 tcells Analyze cytokine production (IFN-γ, IL-17) in CD4+ T
cells after antigen stimulation using flow cytometry data.
analyze ebv antibody titers Analyze ELISA data to quantify EBV antibody titers in
plasma/serum samples.
analyze cns lesion histology Analyzes histological images of CNS lesions to quantify
immune cell infiltration, demyelination, and tissue damage.
analyze immunohistochemistry image Analyzes immunohistochemistry images to quantify protein
expression and spatial distribution.

70
 Table 12: Tool descriptions for biomni.microbiology

Tool Name Tool Description

optimize anaerobic digestion process Optimize anaerobic digestion process conditions to maxi-
mize VFA production or methane yield.
analyze arsenic speciation hplc icpms Analyzes arsenic speciation in liquid samples using HPLC-
ICP-MS technique and returns a detailed research log.
count bacterial colonies Count bacterial colonies from an image of agar plate using
computer vision techniques.
annotate bacterial genome Annotate a bacterial genome using Prokka to identify genes,
proteins, and functional features.
enumerate bacterial cfu by serial dilution Quantify bacterial concentration (CFU/mL) using serial di-
lutions and spot plating.
model bacterial growth dynamics Model bacterial population dynamics over time using ordi-
nary differential equations.
quantify biofilm biomass crystal violet Quantifies biofilm biomass using crystal violet staining as-
say data and generates a detailed research log of the analy-
sis.
segment and analyze microbial cells Perform automated cell segmentation and quantify morpho-
logical metrics from fluorescence microscopy images.
segment cells with deep learning Perform cell segmentation on fluorescence microscopy im-
ages using deep learning models from the Cellpose/Omni-
pose library.
simulate generalized lotka volterra dynamics Simulate microbial community dynamics using the Gener-
alized Lotka-Volterra (gLV) model.
predict rna secondary structure Predict the secondary structure of an RNA molecule using
ViennaRNA and generate visualization files.
simulate microbial population dynamics Performs stochastic simulation of microbial population dy-
namics using the Gillespie algorithm.

71
 Table 13: Tool descriptions for biomni.pathology

Tool Name Tool Description

analyze aortic diameter and geometry Analyze aortic diameter and geometry from cardiovascu-
lar imaging data to measure aortic root diameter, ascending
aorta diameter, and calculate geometric parameters such as
tortuosity and dilation indices.
analyze atp luminescence assay Analyze luminescence-based ATP assay data to determine
intracellular ATP concentration and generate a detailed re-
search log of the analysis.
analyze thrombus histology Analyze histological images of thrombus samples stained
with H&E to identify and quantify different thrombus com-
ponents (fresh, cellular lysis, endothelialization, fibroblastic
reaction).
analyze intracellular calcium with rhod2 Analyzes intracellular calcium concentration using Rhod-2
fluorescent indicator from microscopy images.
quantify corneal nerve fibers Quantify the volume/density of immunofluorescence-
labeled corneal nerve fibers from microscopy images.
segment and quantify cells in multiplexed images Segment cells and quantify protein expression levels from
multichannel tissue images.
analyze bone microct morphometry Analyze bone microarchitecture parameters from 3D micro-
CT images, calculating metrics such as bone mineral den-
sity (BMD), bone volume (BV), trabecular number (Tb.N),
trabecular thickness (Tb.Th), and trabecular separation
(Tb.S).

72
 Table 14: Tool descriptions for biomni.pharmacology

Tool Name Tool Description

run diffdock with smiles Run DiffDock molecular docking simulation using a protein
structure and a SMILES string for the ligand. Uses Docker
to execute the DiffDock algorithm.
docking autodock vina Performs molecular docking using AutoDock Vina to pre-
dict binding affinities between small molecules and a recep-
tor protein.
run autosite Runs AutoSite on a protein structure to identify potential
binding sites and returns a research log with the results.
retrieve topk repurposing drugs from disease txgnn Computes TxGNN model predictions for drug repurposing
for a given disease and returns the top predicted drugs with
their scores.
predict admet properties Predicts ADMET (Absorption, Distribution, Metabolism,
Excretion, Toxicity) properties for a list of compounds us-
ing pretrained models.
predict binding affinity protein 1d sequence Predicts the binding affinity between small molecules and a
protein sequence using pre-trained deep learning models.
analyze accelerated stability of pharmaceutical formulations Analyzes the stability of pharmaceutical formulations under
accelerated storage conditions and generates a research log
of the results.
run 3d chondrogenic aggregate assay Generates a detailed protocol for performing a 3D chondro-
genic aggregate culture assay to evaluate compounds’ ef-
fects on chondrogenesis.
grade adverse events using vcog ctcae Grade and monitor adverse events in animal studies using
the VCOG-CTCAE standard.
analyze radiolabeled antibody biodistribution Analyze biodistribution and pharmacokinetic profile of ra-
diolabeled antibodies, including tissue distribution, half-
lives, and tumor-to-normal tissue ratios.
estimate alpha particle radiotherapy dosimetry Estimate radiation absorbed doses to tumor and normal or-
gans for alpha-particle radiotherapeutics using the Medical
Internal Radiation Dose (MIRD) schema.
perform mwas cyp2c19 metabolizer status Perform a Methylome-wide Association Study (MWAS) to
identify CpG sites significantly associated with CYP2C19
metabolizer status.
calculate physicochemical properties Calculate key physicochemical properties of a drug candi-
date molecule including molecular weight, cLogP, TPSA,
H-bond donors/acceptors, and other drug-like characteris-
tics.
analyze xenograft tumor growth inhibition Analyze tumor growth inhibition in xenograft models
across different treatment groups.
analyze western blot Performs densitometric analysis of Western blot images to
quantify relative protein expression.

73
 Table 15: Tool descriptions for biomni.physiology

Tool Name Tool Description

reconstruct 3d face from mri Generate a 3D model of facial anatomy from MRI scans of
the head and neck.
analyze abr waveform p1 metrics Extracts P1 amplitude and latency from Auditory Brainstem
Response (ABR) waveform data. P1 (Wave I) is typically
the first positive peak in the ABR waveform and is a critical
marker for auditory function assessment.
analyze ciliary beat frequency Analyze ciliary beat frequency from high-speed video mi-
croscopy data using FFT analysis.
analyze protein colocalization Analyze colocalization between two fluorescently labeled
proteins in microscopy images.
perform cosinor analysis Performs cosinor analysis on physiological time series data
to characterize circadian rhythms.
calculate brain adc map Calculate Apparent Diffusion Coefficient (ADC) map from
diffusion-weighted MRI data using the monoexponential
diffusion model.
analyze endolysosomal calcium dynamics Analyze calcium dynamics in endo-lysosomal compart-
ments using ELGA/ELGA1 probe data.
analyze fatty acid composition by gc Analyzes fatty acid composition in tissue samples using gas
chromatography data.
analyze hemodynamic data Analyzes raw blood pressure data to calculate key hemody-
namic parameters including systolic blood pressure, dias-
tolic blood pressure, mean arterial pressure, and heart rate.
simulate thyroid hormone pharmacokinetics Simulates the transport and binding of thyroid hormones
across different tissue compartments using an ODE-based
pharmacokinetic model.
quantify amyloid beta plaques Analyzes an image to detect, quantify, and characterize
amyloid-beta plaques commonly found in Alzheimer’s dis-
ease tissue samples.

74
 Table 16: Tool descriptions for biomni.synthetic biology

Tool Name Tool Description

engineer bacterial genome for therapeutic delivery Engineer a bacterial genome by integrating therapeutic ge-
netic parts for therapeutic delivery.
analyze bacterial growth rate Analyze bacterial growth data and extract growth parame-
ters from OD600 measurements.
analyze barcode sequencing data Analyze sequencing data to extract, quantify and determine
lineage relationships of barcodes.
analyze bifurcation diagram Performs bifurcation analysis on a dynamical system and
generates a bifurcation diagram.
create biochemical network sbml model Generate a mathematical model of a biochemical network
in SBML format.
optimize codons for heterologous expression Analyzes and optimizes a DNA/RNA sequence for im-
proved expression in a heterologous host organism.
simulate gene circuit with growth feedback Simulate gene regulatory circuit dynamics with growth
feedback, tracking gene expression levels and cell growth
over time.
identify fas functional domains Identifies functional domains within a Fatty Acid Synthase
(FAS) sequence and predicts their roles.

Table 17: Tool descriptions for biomni.systems biology

Tool Name Tool Description

perform flux balance analysis Perform Flux Balance Analysis (FBA) on a genome-scale
metabolic network model to predict metabolic flux distribu-
tions.
model protein dimerization network Model protein dimerization networks to find equilibrium
concentrations of dimers based on monomer concentrations
and binding affinities.
simulate metabolic network perturbation Construct and simulate kinetic models of metabolic net-
works and analyze their responses to perturbations.
simulate protein signaling network Simulate protein signaling network dynamics using ODE-
based logic modeling with normalized Hill functions.
compare protein structures Compares two protein structures to identify structural dif-
ferences and conformational changes.
simulate renin angiotensin system dynamics Simulate the time-dependent concentrations of renin-
angiotensin system (RAS) components.

Table 18: Tool descriptions for biomni.support tools

Tool Name Tool Description

run python repl Executes a Python command in the notebook environment
and returns the output as a string.
read function source code Read the source code of a function from any module path.

75
 Table 19: Tool descriptions for biomni.database

Tool Name Tool Description

query uniprot Query the UniProt REST API using either natural language
or a direct endpoint to retrieve protein information.
query alphafold Query the AlphaFold Database API for protein structure
predictions and information.
query interpro Query the InterPro REST API using natural language or
a direct endpoint to retrieve information about protein do-
mains or families.
query pdb Query the RCSB PDB database using natural language or a
direct structured query to find protein structures.
query pdb identifiers Retrieve detailed data and/or download files for PDB iden-
tifiers.
query kegg Take a natural language prompt and convert it to a structured
KEGG API query, then execute the query.
query stringdb Query the STRING protein interaction database using natu-
ral language or direct endpoint.
query iucn Query the IUCN Red List API using natural language or a
direct endpoint to retrieve species conservation status infor-
mation.
query paleobiology Query the Paleobiology Database (PBDB) API using natu-
ral language or a direct endpoint.
query jaspar Query the JASPAR REST API using natural language or a
direct endpoint to retrieve transcription factor binding pro-
files.
query worms Query the World Register of Marine Species (WoRMS)
REST API using natural language or a direct endpoint.
query cbioportal Query the cBioPortal REST API using natural language or
a direct endpoint to access cancer genomics data.
query clinvar Take a natural language prompt and convert it to a structured
ClinVar query to search for genetic variants.
query geo Query the NCBI Gene Expression Omnibus (GEO) using
natural language or a direct search term.
query dbsnp Query the NCBI dbSNP database using natural language or
a direct search term.
query ucsc Query the UCSC Genome Browser API using natural lan-
guage or a direct endpoint.
76
 Table 20: Continue tool descriptions for biomni.database

Tool Name Tool Description

query ensembl Query the Ensembl REST API using natural language or a
direct endpoint to retrieve genomic data.
query opentarget genetics Query the OpenTargets Genetics API using natural lan-
guage or a direct GraphQL query to retrieve information
about genetic targets and variants.
query opentarget Query the OpenTargets Platform API using natural lan-
guage or a direct GraphQL query to access drug targets,
diseases, and mechanisms data.
query gwas catalog Query the GWAS Catalog API using natural language or a
direct endpoint to retrieve genetic association studies data.
query gnomad Query gnomAD for variants in a gene using natural lan-
guage or direct gene symbol.
blast sequence Identifies a DNA or protein sequence using NCBI BLAST
and returns information about the best alignment.
query reactome Query the Reactome database using natural language or
a direct endpoint to retrieve information about biological
pathways.
query regulomedb Query the RegulomeDB database using natural language or
direct endpoint specification to get information about regu-
latory elements.
query pride Query the PRIDE (PRoteomics IDEntifications) database
using natural language or a direct endpoint.
query gtopdb Query the Guide to PHARMACOLOGY database (GtoPdb)
using natural language or a direct endpoint.
region to ccre screen Retrieves information about candidate cis-regulatory ele-
ments (cCREs) that intersect with a specified genomic re-
gion.
get genes near ccre Identifies the nearest genes to a specified candidate cis-
Regulatory Element (cCRE) by querying the SCREEN
database.
query remap Query the ReMap database for regulatory elements and
transcription factor binding sites using natural language or
direct API endpoints.
query mpd Query the Mouse Phenome Database (MPD) for mouse
strain phenotype data using natural language or direct end-
point access.
query emdb Query the Electron Microscopy Data Bank (EMDB) for 3D
macromolecular structures using natural language or direct
endpoint77
access.
 Table 21: Data lake descriptions

File Name Description

affinity capture-ms.csv Protein-protein interactions detected via affinity capture and
mass spectrometry.
affinity capture-rna.csv Protein-RNA interactions detected by affinity capture.
BindingDB All 202409.tsv Measured binding affinities between proteins and small
molecules for drug discovery.
broad repurposing hub molecule with smiles.csv Molecules from Broad Institute’s Drug Repurposing Hub
with SMILES annotations.
broad repurposing hub phase moa target info.csv Drug phases, mechanisms of action, and target information
from Broad Institute.
co-fractionation.csv Protein-protein interactions from co-fractionation experi-
ments.
Cosmic Breakpoints v101 GRCh38.csv Genomic breakpoints associated with cancers from COS-
MIC database.
Cosmic CancerGeneCensusHallmarksOfCancer v101 GRCh38.csv Hallmarks of cancer genes from COSMIC.
Cosmic CancerGeneCensus v101 GRCh38.csv Census of cancer-related genes from COSMIC.
Cosmic ClassificationPaper v101 GRCh38.csv Cancer classifications and annotations from COSMIC.
Cosmic Classification v101 GRCh38.csv Classification of cancer types from COSMIC.
Cosmic CompleteCNA v101 GRCh38.tsv.gz Complete copy number alterations data from COSMIC.
Cosmic CompleteDifferentialMethylation v101 GRCh38.tsv.gz Differential methylation patterns from COSMIC.
Cosmic CompleteGeneExpression v101 GRCh38.tsv.gz Gene expression data across cancers from COSMIC.
Cosmic Fusion v101 GRCh38.csv Gene fusion events from COSMIC.
Cosmic Genes v101 GRCh38.csv List of genes associated with cancer from COSMIC.
Cosmic GenomeScreensMutant v101 GRCh38.tsv.gz Genome screening mutations from COSMIC.
Cosmic MutantCensus v101 GRCh38.csv Catalog of cancer-related mutations from COSMIC.
Cosmic ResistanceMutations v101 GRCh38.csv Resistance mutations related to therapeutic interventions
from COSMIC.
czi census datasets v4.csv Datasets from the Chan Zuckerberg Initiative’s Cell Census.
DisGeNET.csv Gene-disease associations from multiple sources.
dosage growth defect.csv Gene dosage changes affecting growth.
enamine cloud library smiles.pkl Compounds from Enamine REAL library with SMILES an-
notations.
genebass missense LC filtered.pkl Filtered missense variants from GeneBass.
genebass pLoF filtered.pkl Predicted loss-of-function variants from GeneBass.
genebass synonymous filtered.pkl Filtered synonymous variants from GeneBass.
gene info.csv Comprehensive gene information.
genetic interaction.csv Genetic interactions between genes.
go-plus.json Gene ontology data for functional gene annotations.
gtex tissue gene tpm.csv Gene expression (TPM) across human tissues from GTEx.
gwas catalog.pkl Genome-wide association studies (GWAS) results.
marker celltype.csv Cell type marker genes for identification.
McPAS-TCR.csv T-cell receptor sequences and specificity data from McPAS
database.
miRDB v6.0 results.csv Predicted microRNA targets from miRDB.
miRTarBase microRNA target interaction.csv Experimentally validated microRNA-target interactions
from miRTarBase.
miRTarBase microRNA target interaction pubmed abtract.txt PubMed abstracts for microRNA-target interactions in miR-
TarBase.
miRTarBase MicroRNA Target Sites.csv Binding sites of microRNAs on target genes from miRTar-
Base.

78
 Table 22: Continue data lake descriptions

File Name Description

mousemine m1 positional geneset.csv Positional gene sets from MouseMine.
mousemine m2 curated geneset.csv Curated gene sets from MouseMine.
mousemine m3 regulatory target geneset.csv Regulatory target gene sets from MouseMine.
mousemine m5 ontology geneset.csv Ontology-based gene sets from MouseMine.
mousemine m8 celltype signature geneset.csv Cell type signature gene sets from MouseMine.
mousemine mh hallmark geneset.csv Hallmark gene sets from MouseMine.
msigdb human c1 positional geneset.csv Human positional gene sets from MSigDB.
msigdb human c2 curated geneset.csv Curated human gene sets from MSigDB.
msigdb human c3 regulatory target geneset.csv Regulatory target gene sets from MSigDB.
msigdb human c3 subset transcription factor targets from GTRD.csv Transcription factor targets from GTRD/MSigDB.
msigdb human c4 computational geneset.csv Computationally derived gene sets from MSigDB.
msigdb human c5 ontology geneset.csv Ontology-based gene sets from MSigDB.
msigdb human c6 oncogenic signature geneset.csv Oncogenic signatures from MSigDB.
msigdb human c7 immunologic signature geneset.csv Immunologic signatures from MSigDB.
msigdb human c8 celltype signature geneset.csv Cell type signatures from MSigDB.
msigdb human h hallmark geneset.csv Hallmark gene sets from MSigDB.
omim.csv Genetic disorders and associated genes from OMIM.
proteinatlas.tsv Protein expression data from Human Protein Atlas.
proximity label-ms.csv Protein interactions via proximity labeling and mass spec-
trometry.
reconstituted complex.csv Protein complexes reconstituted in vitro.
synthetic growth defect.csv Synthetic growth defects from genetic interactions.
synthetic lethality.csv Synthetic lethal interactions.
synthetic rescue.csv Genetic interactions rescuing phenotypes.
two-hybrid.csv Protein-protein interactions detected by yeast two-hybrid
assays.
variant table.csv Annotated genetic variants table.
Virus-Host PPI P-HIPSTER 2020.csv Virus-host protein-protein interactions from P-HIPSTER.

79
 Table 23: Software descriptions

Software Name Description

biopython [Python Package] A set of tools for biological computa-
tion including parsers for bioinformatics files, access to on-
line services, and interfaces to common bioinformatics pro-
grams.
biom-format [Python Package] The Biological Observation Matrix
(BIOM) format is designed for representing biological sam-
ple by observation contingency tables with associated meta-
data.
scanpy [Python Package] A scalable toolkit for analyzing single-
cell gene expression data, specifically designed for large
datasets using AnnData.
scikit-bio [Python Package] Data structures, algorithms, and edu-
cational resources for bioinformatics, including sequence
analysis, phylogenetics, and ordination methods.
anndata [Python Package] A Python package for handling annotated
data matrices in memory and on disk, primarily used for
single-cell genomics data.
mudata [Python Package] A Python package for multimodal data
storage and manipulation, extending AnnData to handle
multiple modalities.
pyliftover [Python Package] A Python implementation of UCSC
liftOver tool for converting genomic coordinates between
genome assemblies.
biopandas [Python Package] A package that provides pandas
DataFrames for working with molecular structures and bio-
logical data.
biotite [Python Package] A comprehensive library for computa-
tional molecular biology, providing tools for sequence anal-
ysis, structure analysis, and more.
gget [Python Package] A toolkit for accessing genomic
databases and retrieving sequences, annotations, and other
genomic data.
lifelines [Python Package] A complete survival analysis library for
fitting models, plotting, and statistical tests.
scvi-tools [Python Package] A package for probabilistic modeling of
single-cell omics data, including deep generative models.

80
 Table 24: Continue software descriptions

Software Name Description

gseapy [Python Package] A Python wrapper for Gene Set Enrich-
ment Analysis (GSEA) and visualization.
scrublet [Python Package] A tool for detecting doublets in single-
cell RNA-seq data.
cellxgene-census [Python Package] A tool for accessing and analyzing the
CellxGene Census, a collection of single-cell datasets.
hyperopt [Python Package] A Python library for optimizing hyperpa-
rameters of machine learning algorithms.
scvelo [Python Package] A tool for RNA velocity analysis in single
cells using dynamical models.
pysam [Python Package] A Python module for reading, manipulat-
ing and writing genomic data sets in SAM/BAM/VCF/BCF
formats.
pyfaidx [Python Package] A Python package for efficient random
access to FASTA files.
pyranges [Python Package] A Python package for interval manipula-
tion with a pandas-like interface.
pybedtools [Python Package] A Python wrapper for Aaron Quinlan’s
BEDTools programs.
rdkit [Python Package] A collection of cheminformatics and ma-
chine learning tools for working with chemical structures
and drug discovery.
deeppurpose [Python Package] A deep learning library for drug-target
interaction prediction and virtual screening.
pyscreener [Python Package] A Python package for virtual screening
of chemical compounds.
openbabel [Python Package] A chemical toolbox designed to speak
the many languages of chemical data, supporting file for-
mat conversion and molecular modeling.
descriptastorus [Python Package] A library for computing molecular de-
scriptors for machine learning applications in drug discov-
ery.

81
 Table 25: Continue software descriptions

Software Name Description

pymol [Python Package] A molecular visualization system for ren-
dering and animating 3D molecular structures.
openmm [Python Package] A toolkit for molecular simulation using
high-performance GPU computing.
pytdc [Python Package] A Python package for Therapeutics Data
Commons, providing access to machine learning datasets
for drug discovery.
pandas [Python Package] A fast, powerful, and flexible data analy-
sis and manipulation library for Python.
numpy [Python Package] The fundamental package for scientific
computing with Python, providing support for arrays, ma-
trices, and mathematical functions.
scipy [Python Package] A Python library for scientific and tech-
nical computing, including modules for optimization, linear
algebra, integration, and statistics.
scikit-learn [Python Package] A machine learning library featuring var-
ious classification, regression, and clustering algorithms.
matplotlib [Python Package] A comprehensive library for creating
static, animated, and interactive visualizations in Python.
seaborn [Python Package] A statistical data visualization library
based on matplotlib with a high-level interface for drawing
attractive statistical graphics.
statsmodels [Python Package] A Python module for statistical modeling
and econometrics, including descriptive statistics and esti-
mation of statistical models.
pymc3 [Python Package] A Python package for Bayesian statistical
modeling and probabilistic machine learning.
pystan [Python Package] A Python interface to Stan, a platform for
statistical modeling and high-performance statistical com-
putation.
umap-learn [Python Package] Uniform Manifold Approximation and
82
Projection, a dimension reduction technique.
 Table 26: Continue software descriptions

Software Name Description

faiss-cpu [Python Package] A library for efficient similarity search
and clustering of dense vectors.
harmony-pytorch [Python Package] A PyTorch implementation of the Har-
mony algorithm for integrating single-cell data.
tiledb [Python Package] A powerful engine for storing and ana-
lyzing large-scale genomic data.
tiledbsoma [Python Package] A library for working with the SOMA
(Stack of Matrices) format using TileDB.
h5py [Python Package] A Python interface to the HDF5 binary
data format, allowing storage of large amounts of numerical
data.
tqdm [Python Package] A fast, extensible progress bar for loops
and CLI applications.
joblib [Python Package] A set of tools to provide lightweight
pipelining in Python, including transparent disk-caching
and parallel computing.
opencv-python [Python Package] OpenCV library for computer vision
tasks, useful for image analysis in biological contexts.
PyPDF2 [Python Package] A library for working with PDF files, use-
ful for extracting text from scientific papers.
googlesearch-python [Python Package] A library for performing Google searches
programmatically.
scikit-image [Python Package] A collection of algorithms for image pro-
cessing in Python.
pymed [Python Package] A Python library for accessing PubMed
articles.
arxiv [Python Package] A Python wrapper for the arXiv API, al-
lowing access to scientific papers.

83
 Table 27: Continue software descriptions

Software Name Description

scholarly [Python Package] A module to retrieve author and publica-
tion information from Google Scholar.
cryosparc-tools [Python Package] Tools for working with cryoSPARC, a
platform for cryo-EM data processing.
mageck [Python Package] Analysis of CRISPR screen data.
igraph [Python Package] Network analysis and visualization.
pyscenic [Python Package] Analysis of single-cell RNA-seq data and
gene regulatory networks.
cooler [Python Package] Storage and analysis of Hi-C data.
trackpy [Python Package] Particle tracking in images and video.
flowcytometrytools [Python Package] Analysis and visualization of flow cytom-
etry data.
cellpose [Python Package] Cell segmentation in microscopy images.
viennarna [Python Package] RNA secondary structure prediction.
PyMassSpec [Python Package] Mass spectrometry data analysis.
python-libsbml [Python Package] Working with SBML files for computa-
tional biology.
cobra [Python Package] Constraint-based modeling of metabolic
networks.
reportlab [Python Package] Creation of PDF documents.
flowkit [Python Package] Toolkit for processing flow cytometry
data.
hmmlearn [Python Package] Hidden Markov model analysis.
msprime [Python Package] Simulation of genetic variation.
tskit [Python Package] Handling tree sequences and population
genetics data.
cyvcf2 [Python Package] Fast parsing of VCF files.
pykalman [Python Package] Kalman filter and smoother implementa-
tion.

84
 Table 28: Continue software descriptions

Software Name Description

fanc [Python Package] Analysis of chromatin conformation data.
ggplot2 [R Package] A system for declaratively creating graphics,
based on The Grammar of Graphics. Use with subpro-
cess.run([’Rscript’, ’-e’, ’library(ggplot2); ...’]).
dplyr [R Package] A grammar of data manipulation, providing a
consistent set of verbs that help you solve the most common
data manipulation challenges. Use with subprocess.
tidyr [R Package] A package that helps you create tidy data,
where each column is a variable, each row is an observa-
tion, and each cell is a single value. Use with subprocess.
readr [R Package] A fast and friendly way to read rectangu-
lar data like CSV, TSV, and FWF. Use with subpro-
cess.run([’Rscript’, ’-e’, ’library(readr); ...’]).
stringr [R Package] A cohesive set of functions designed to make
working with strings as easy as possible. Use with subpro-
cess calls.
Matrix [R Package] A package that provides classes and methods
for dense and sparse matrices. Required for Seurat. Use
with subprocess calls.
Rcpp [R Package] Seamless R and C++ Integration, allowing R
functions to call compiled C++ code. Use with subprocess
calls.
devtools [R Package] Tools to make developing R packages easier,
including functions to install packages from GitHub. Use
with subprocess calls.
remotes [R Package] Install R packages from GitHub, GitLab, Bit-
bucket, or other remote repositories. Use with subprocess
calls.
DESeq2 [R Package] Differential gene expression analysis based
on the negative binomial distribution. Use with subpro-
cess.run([’Rscript’, ’-e’, ’library(DESeq2); ...’]).
85
 Table 29: Continue software descriptions

Software Name Description

clusterProfiler [R Package] A package for statistical analysis and visual-
ization of functional profiles for genes and gene clusters.
Use with subprocess calls.
DADA2 [R Package] A package for modeling and correcting
Illumina-sequenced amplicon errors. Use with subprocess
calls.
xcms [R Package] A package for processing and visualization of
LC-MS and GC-MS data. Use with subprocess calls.
FlowCore [R Package] Basic infrastructure for flow cytometry data.
Use with subprocess calls.
edgeR [R Package] Empirical Analysis of Digital Gene Expression
Data in R, for differential expression analysis. Use with
subprocess calls.
limma [R Package] Linear Models for Microarray Data, for differ-
ential expression analysis. Use with subprocess calls.
harmony [R Package] A method for integrating and analyzing single-
cell data across datasets. Use with subprocess calls.
WGCNA [R Package] Weighted Correlation Network Analysis for
studying biological networks. Use with subprocess calls.
samtools [CLI Tool] A suite of programs for interacting with
high-throughput sequencing data. Use with subpro-
cess.run([’samtools’, ...]).
bowtie2 [CLI Tool] An ultrafast and memory-efficient tool for align-
ing sequencing reads to long reference sequences. Use with
subprocess.run([’bowtie2’, ...]).
bwa [CLI Tool] Burrows-Wheeler Aligner for mapping low-
divergent sequences against a large reference genome. Use
with subprocess.run([’bwa’, ...]).

86
 Table 30: Continue software descriptions

Software Name Description

bedtools [CLI Tool] A powerful toolset for genome arithmetic,
allowing operations like intersect, merge, count, and
complement on genomic features. Use with subpro-
cess.run([’bedtools’, ...]).
macs2 [CLI Tool] Model-based Analysis of ChIP-Seq data, a tool
for identifying transcript factor binding sites.
fastqc [CLI Tool] A quality control tool for high throughput se-
quence data. Use with subprocess.run([’fastqc’, ...]).
trimmomatic [CLI Tool] A flexible read trimming tool for Illumina NGS
data. Use with subprocess.run([’trimmomatic’, ...]).
mafft [CLI Tool] A multiple sequence alignment program
for unix-like operating systems. Use with subpro-
cess.run([’mafft’, ...]).
Homer [CLI Tool] Motif discovery and next-gen sequencing anal-
ysis.
FastTree [CLI Tool] Phylogenetic trees from sequence alignments.
muscle [CLI Tool] Multiple sequence alignment tool.
plink2 [CLI Tool] A comprehensive toolkit for genome associa-
tion studies that can perform a range of large-scale analyses
in a computationally efficient manner. Use with subpro-
cess.run([’plink2’, ...]).
gcta64 [CLI Tool] Genome-wide Complex Trait Analysis (GCTA)
tool for estimating the proportion of phenotypic variance
explained by genome-wide SNPs and analyzing genetic re-
lationships. Use with subprocess.run([’gcta64’, ...]).
iqtree2 [CLI Tool] An efficient phylogenetic software for maximum
likelihood analysis with built-in model selection and ultra-
fast bootstrap. Use with subprocess.run([’iqtree2’, ...]).

87
 Table 31: Completeness scoring rubric for experimental documentation.

Score Level Description

1 Severely Incomplete Missing many major components necessary for ex-
periments. Cannot be followed for implementation.
Lacks critical methodological details and equipment
specifications.
2 Significantly Incomplete Missing few major components. Requires substantial
additional effort from user. Key experimental param-
eters or control measures absent.
3 Moderately Complete Contains basic components with some detail gaps.
Requires moderate additional effort from user. Most
critical parameters defined but needs optimization.
4 Mostly Complete Contains nearly all necessary components. Requires
minimal additional effort from user. Well-described
procedures with adequate troubleshooting guidance.
5 Completely Thorough Contains all necessary components in appropriate de-
tail. Can be followed precisely with no additional
effort. Comprehensive methodology, specifications,
and contingency plans.

Table 32: Accuracy scoring rubric for experimental documentation.

Score Level Description

1 Severely Inaccurate Contains many major errors or misconceptions.
Methods would lead to invalid or unreliable results.
Fundamental scientific principles violated or misap-
plied.
2 Significantly Inaccurate Contains few major errors or several minor errors.
Some methodological approaches are flawed. May
lead to partially valid but questionable results.
3 Moderately Accurate Generally correct but contains some minor errors or
imprecisions. Core methodology is sound but may
have optimization issues. Results would be accept-
able but not optimal.
4 Mostly Accurate Contains minimal errors. Methods align with ac-
cepted scientific practices. Results would be reliable
with only minor corrections needed.
5 Completely Accurate Contains no errors. Methods perfectly align with best
scientific practices. Results would be highly reliable
and reproducible. Includes appropriate controls and
validation steps.

88