Early detection of abiotic stress in plants through SNARE proteins using hybrid feature fusion model

Introduction

Over the past few years, agriculture has become increasingly vital in the global economy, playing a critical role in ensuring safe and efficient food production. This has led to a rising demand for smart farming techniques. The integration of computer vision and deep learning has enabled the use of advanced technologies in agricultural automation, significantly benefiting small-scale farming practices (Ünal, 2020). The Indian economy relies on agriculture by acquiring sustainability in agricultural production by including advanced technologies in farm management. Stress in plants is the main threat to productivity. Identifying stress at the initial stage can curb the loss because this is a very challenging task in agriculture (Chung, Breshears & Yoon, 2018). Abiotic stressors are environmental factors that negatively impact plant growth and development, leading to reduced yields below optimal levels. These stressors, which include factors like temperature extremes, water scarcity, and soil salinity, can significantly affect crops and commercial plants. In some cases, they can cause crop production to decrease by up to 70%, limiting plants to operate at just 30% of their genetic potential (Abbas & El-Manzalawy, 2020). Plant growth under unfavourable conditions is termed plant stress. Plant stress is categorized into biotic stress which is caused due to the invasion of bacteria, viruses, fungi, pathogens, and many more whereas abiotic stress occurs mainly due to the intervention of non-living organisms. stresses caused by heat, drought, and salinity (Cushman & Bohnert, 2000). In a few cases, these stressors acts as a defensive mechanism that protects the plant from deviations caused by external factors. The impact of stress shows its imprint in the decline of food production which further affects the supply-chain management of the agricultural industry (Jansen & Potters, 2017). Furthermore, fluctuations in climatic conditions can also lead to great damage (Rico-Chávez et al., 2022). Artificial Intelligence-based methodologies showed promising results in analyzing the mechanism of plant stress, thereby helping farmers monitor crop stress in the early stages (Fenu & Malloci, 2021). Abiotic stress greatly impacts the physical and biological status of the plant. The vital biochemical present in plants helps them to increase the quantity of the produce (Moghimi, Yang & Marchetto, 2018). Abiotic stress disturbs the gene structure of the plants which leads to the fatal destruction of the yield. Abiotic stress in plants is mainly contributed by different phases of plant growth, advancements in the genetic structure of the plants, and various biotic and abiotic stressors (Kazan, 2015). Proteins play an essential role in the biological functioning of plant tissue. Identifying protein-protein interaction in plants is vital to know the organ and tissue formation, cell structure, and plant defense mechanism (Yang et al., 2011). Therefore, analysis of the interaction between sequence-based proteins is important to identify the stress in plants (Zhang, Gao & Yuan, 2010; Khan & Kihara, 2016). Proteins are the most essential building blocks present in every life science. Understanding the underlying structure and functions of proteins is a challenging task. Protein sequences are composed of a series of alphabets and machine learning and deep learning methodologies are used to decode the sequence (Ofer, Brandes & Linial, 2021). Deep learning, a subset of Artificial Intelligence is extensively used to learn the data stored in multiple layers as sequences of hidden amino acid details from a protein sequence (LeCun, Bengio & Hinton, 2015). SNARE proteins (soluble N-ethylmaleimide-sensitive factor activating protein receptors) are vital structures that hold vital information about cell membrane formation (Kha, Ho & Le, 2022). By considering and analyzing the key importance of SNARE proteins in the transportation of essential information, several techniques have been used to decode the information. One of them is to investigate the SNARE from the unknown motif information with the help of bioinformatics (Kloepper et al., 2007). Furthermore, SNAREs are identified with the help of position position-specific scoring matrix and are fed into a 2D convolutional neural network in the form of images thereby extracting the vital information from the sequential data (Le & Nguyen, 2019). Numerous environmental stresses trigger the plant’s biological structure leading to changes in the life cycle such as changes in the functionality of the antioxidants involved, reconstruction of the endomembrane structure, accumulation of the visible solutes in the gene expression, and change in the transportation cycle of the protein substances (Wang et al., 2020). Nowadays, crop damage is mainly due to climatic changes and incorrect human actions resulting in the decline of food production (Mousavi-Derazmahalleh et al., 2019). Plants developed a resistant environment by inhibiting the nature of adapting to the stress as well as coping with uninvited changes in the environment made them benefit from the adversities created by the unfavorable nature leading them to build a strong defense against their constantly shifting environment (Kalinowska & Isono, 2018). Throughout their life cycles, plants are continuously exposed to several abiotic challenges, including biotic stresses and drought, excessive salinity, heat, cold, freezing, UV-B, and osmotic pressures (Sanderfoot, 2007). Proteins are a jack of all bio-molecular functions of a living cell and they are represented as a chain of patterns built with amino acid composition with unique lengths. Feature extraction is the main step that exposes the interaction and the involvement of amino acids in the function of the cell. Traditional methods extract features from the Basic Local Alignment Search Tool (BLAST) (Xu et al., 2021). To extract the descriptive features, evolutionary methods like Auto-Cross-Covariance (ACC) (Dong, Zhou & Guan, 2009) and Separated-Dimer (SD) (Saini et al., 2015). Our work is the first one to detect plant stress using SNARE proteins. In our work, we focused on extracting the features from the protein sequence with the help of convolutional neural networks, and thereby we emphasized the presence of SNARE proteins in estimating the abiotic stress in plants. The main contributions of our work are:

1. To project the involvement of SNARE proteins in alleviating the abiotic stress in plants.

2. An improvised feature extraction method to analyze the occurrence of amino acids in the protein-protein interaction of plant cells.

3. A deep neural network architecture for classifying the SNARE and non-SNARE proteins in plant protein sequences.

4. Visualization of hidden features of the network is done using manifold discovery and analysis (MDA).

Related works

Dey et al. (2022) identified the diseases that different pathogens attacked in the rice crop, leading to the manifestation of diseases such as brown spot, hispa, and NPK (nitrogen, phosphorus, and potassium) shortage. In order to precisely detect the stress in plants, the dataset was gathered from the paddy field and analyzed with several current models, including restnet50, VGG 16, VGG 19, and Inception V3.The deep CNN model performed well in terms of classification.

The work done by de Melo et al. (2022) focused on identifying water stress in sugarcane crop using thermal images as manual visualization is time-consuming. They used an integrated model called Inception-ResnetV2 which gave good accuracy with low time than traditional machine learning models. Thermal images would identify the stressed and non-stressed images by comparing the intensity distribution on the images.

Yi et al. (2020) worked on identifying the nutrient deficiency in sugarbeet crop with the help of sugarbeet nutrient deficiency dataset. They compared various CNN algorithms to find out the best model that would detect nitrogen, phosphorus, potasium and calcium deficiency. Among all the models, Denset has shown promising results. In the work done by Shona et al. (2022) cold stress in watermelon has been analyzed as this is rich in water content and nutritional fruit consumed in many countries. Here the crop is dried under a controlled environment by inducing the perfect temperature for the growth of this plant. The stress in this plant was detected based on morphological features like reduction in leaf size, shape, and leaves. U-Net architecture is used for the classification of stressed and non-stressed plants and the model achieved 100% accuracy. This article (Zhou et al., 2021) projected the importance of high throughput phenotyping in measuring crop traits. This work mainly aims to estimate the injury caused by flood currents to the soybean plant. Soya bean plants are grown under controlled conditions where the plant shows flooding symptoms and the images are collected using Ariel equipment. The flooding injury score was calculated for 724 species and five features was analyzed namely, canopy temperature, normalized difference vegetation index, canopy area, width, and length, and the model outperformed in identifying the flooding stress. This study carried out by Zhang et al. (2020) worked on identifying crop leaf purplings which usually occur when a plant is subjected to stress. Purple rapeseed leaves are examined using a UAV vehicle at pixel level with limited resolution and are segmented using U-Net architecture. The focus is to identify nitrogen stress. Leaf purpling is mainly due to a lack of nitrogen (N), phosphorus (p), and potassium (k). Accurate prediction is done with U-Net with a patch size of 256 × 256. The work (Chandel et al., 2021) focused on identifying the water stress in plants which causes huge losses to agricultural produce. The author carried out this work on maize, soybean, and okra with the help of traditional DL algorithms like Inception V3, AlexNet, and googleNet to classify and identify the plant stress. The data used for this work was collected in real-time from the crops and a total of 12,000 images were gathered from the fields using specialized cameras the convolutional neural network architectures were applied with a set of hyper-parameters and classification was done successfully in identifying the stressed and non-stressed species. Yu, Fang & Zhao (2021)’s work depicted the negative effects of heavy metal stress in tobacco plants. Unsupervised machine learning algorithms have shown remarkable results in identifying heavy metal stress. Hyperspectral images of the tobacco plant canopy were collected. Least-squares discriminant analysis (PLS-DA) and least-squares support vector machine (LS-SVM) algorithms are employed to identify heavy metal stress in plants and analyze different chemical compositions associated with it. Analysis of the nitrogen deficiency in sorghum plants has been done by Azimi, Kaur & Gandhi (2021). The author trained a CNN model to automatically detect the stress in plants. The shoot images were collected from Donald University and the proposed model achieved 96% accuracy in detecting the nitrogen stress. In the work done by Das et al. (2020), authors determined the method to identify the salt stress in the most produced and consumed food i.e., rice and it suffers from abiotic stress when adequate/low concentration of NaCl gets deposited in the soil. K-nearest neighbors (KNN) are used to classify the stressed and non-stressed and principal component analysis was used to extract features from the spectral images and partial least square regression and supervised machine learning models have performed efficiently in classifying the salt-stressed species.

Dao, He & Proctor (2021) compared the involvement of deep learning and machine learning in identifying drought stress using spectral analysis. Drought stress impacts are complicated and it affects leaves at multiple stages of severity. The derivative spectra achieved 97.5% accuracy with the help of DL algorithms. The goal of the work proposed in Zahid et al. (2022) is to discover water stress in the plant species of basil, coriander, parsley, coffee, bay leaf, and pea. With the aid of supervised machine learning techniques like random forest, SVM, and KNN, stress can be classified and predicted. By examining the morphological aspects of the data, the KNN was followed by the KNN94.64% accuracy and the SVM89.67% accuracy in terms of accuracy. Additionally, RF performed better than other classifiers in identifying water-stressed leaves and has a 99.42% accuracy rate. The work of Niu et al. (2021) identified the water stress in maize using multi-spectral images. The experiment was conducted in open maize fields in China and UAV-based RGB imagery was used to analyze the effectiveness of sensors. Random forest, ANN, and multivariant linear regression are used for the classification and identification of crop water stress and growth stages of the plant. Fractional vegetation cover (FVC) is the key measure in identifying the water stress and these models accurately calculated the FVC. Mondal et al. (2019) worked to identify the water stress in wheat. In this work, these images are captured by the Indian Agriculture Research Institute. The stressed species are classified using a random forest (RF) algorithm. Supervised machine learning techniques have shown outstanding results in analyzing the spectral data from the high-resolution images and accurate features are selected that could easily provide reliable data for analyzing the water stress in plants.

Ly et al. (2018) suggested a model to identify the environmental stress in wheat. This study mainly concentrates on identifying whether a plant is adaptable to stress or not. An extension of factorial regression called the genomic random regression model is used to predict the stress-resistant variant. Khatoon et al. (2021) has been done on analyzing the nutrient deficiency in tomato plants. The images were gathered from the field as of sufficient nutrient supply is very crucial in plant breeding. A deep neural network called DenseNet 121 has shown promising results in capturing the stressed species accurately.

The work done by Asefpour Vakilian (2020) explained the role of miRNA in regulating plant stress responses. Several machine learning algorithms like decision trees, naïve Bayes, and support vector machines are used to predict the plant stress based on miRNA concentration.

From the above works, we conclude that identifying abiotic stress in plants is found to be prominent in analyzing the productivity of the yield and several traditional methods tried to detect the cause of stress with some limitations. We herein propose a hybrid model to overcome the hurdles of the previous models in spotting and processing the important hidden features from the protein patterns. Our deep learning framework focus on extracting the reliable features from the given input patterns by incorporating traditional CNN along with RBF and to classify the sequences based on the presence of SNARE proteins we used Bi-directional LSTM for classification. Moreover, we performed comparisons with the existing methods to protect the supremacy of our proposed approach. The above literature survey is summarized in Table 1.

Materials and Methods

This section illustrates the sequential flow of protein-protein interaction in extracting the features from the Amino acid sequences.

Encoding the amino acid representation

To extract the features from the protein sequences, we calculated the PSSM for the FASTA sequences. PSSM is composed of the representation of the frequency of amino acids in the protein sequence. This is calculated by rendering the ordering of similar amino acids occurring in different sequences. For this reason, PSSM is widely used in several bioinformatics applications. The data retrieved from the Uniprot database is in the form of FASTA sequences and it is decoded into PSSM with the help of PSI-BLAST. The amino acids in a protein sequence as represented in the Fig. 1.

To generate this the PSI-BLAST searches for the non-redundant sequence representations by performing two iterations and the feature vector is generated represented in the form of M x 20 matrix and A = {P_a,b; a = 1……m and b = 1………20} and the PSSM is defined as follows

A={P1,1P1,2P1,3⋯⋯⋯⋯P1,20P2,1P2,2P2,3⋯⋯⋯⋯P2,20P3,1P3,2P3,3⋯⋯⋯⋯P3,20P4,1P4,2P4,3⋯⋯⋯⋯P4,20⋮⋮⋮⋮⋮Pm,1Pm,2Pm,3⋯⋯⋯⋯Pm,20}where P_a,b represents the probability of a mutation to both amino acids during the evolution of the features from the sequential data. The amino acid protein sequences are represented in the form of a vector that holds 20 different values which symbolizes the kind of amino acid involved in them. The main function of this step is to generate a vector that can be easily modeled by CNN. Now each of these 400D vectors is normalized between 0–1 by dividing each vector by the length of the sequences. By incorporating the features of NLP, we encoded amino acid representation. The main objective is to effectively use the applications of NLP in the field of biological sequences. To extract the information from the PSSM the whole sequence should be taken as a word. For this purpose, we used fastText a model developed by Facebook that considers a word as a continuous bag of n-grams extension of word2vec (Bojanowski et al., 2017). The main idea is to treat protein sequence as a sequence and Amino acids as words and subsequently, we have generated the feature vectors of reduced dimensionality. Figure 2 represents the amino acid values.

Proposed work

Bi-directional long short-memory

To model the protein sequences to learn the long-range dependencies of the sequential data and to identify the presence of SNARE proteins in a sequence we applied bi-directional long short-memory (Bi-LSTM) which updates the hidden states for sequential data from two directions. The architecture works as shown in Fig. 3 the Bi-LSTM layer consists of two layers of LSTM where one layer receives the input in the backward direction (h^p_t) and the other layer receives the input in the forward direction (h^f_t) where t = (1, 2, 3, ….. n). The output is given by combining the inputs of forward and backward layers. The Bi-LSTM layers compute Hid=(h1,⋯⋯,ht) and Out=(o1,⋯⋯,ot) output obtained from hidden layer and output layer as in Eqs. (5) and (6),

(5) Ht=ACT(Wshht+Whhht−1+Biah)

(6) Ot=Whoht+Biaowhere Wsh is the weight matrix between the input layer and the intermediate layer Biah is the bias vector computed for the intermediated layer and ACT is the non-linear vector activation function. To calculate the presence of SNARE proteins in the given seq we use the function F(seq) as shown in Eq. (7)

(7) F(seq)=Bi−LSTM(CNN+RBF(Embedding(Encoding(seq)).

Feature fusion model

We proposed a feature fusion model that combines the functionalities of CNN, RBF, and Bi-LSTM to detect abiotic stress in plants. The architecture of the model is shown in Fig. 4. This hybrid approach focuses on projecting the hierarchical feature extraction capabilities of CNN non-non-linear modeling of RBF networks and the sequential analysis of Bi-LSTM. This hybrid approach influences the strengths of different neural network architectures to potentially improve the overall performance and capture complex patterns in the data. The main contribution of the feature fusion model to our sequential data is due to the following attributes that contribute to identifying abiotic stress in plants effectively:

1. Hierarchical feature extraction:

CNNs are excellent at learning hierarchical features from raw data, such as images or sequences. The significant property is that they can capture low-level features and gradually learn more abstract features. When the CNNs are combined with subsequent layers like RBF and Bi-LSTM, the model is capable of capturing both local and global patterns in the data.

2. Nonlinear and radial basis function (RBF) modeling:

RBF networks are good at approximating complex nonlinear relationships. They can learn to map input features to nonlinear functions, which might be particularly useful for capturing intricate patterns that CNNs alone may not fully grasp.

3. Enhanced representation learning:

Through the deployment of the fusion model, there is a greater possibility of enrichment in feature representations of the data. CNNs can capture spatial or sequential patterns, RBF networks can capture complex nonlinear relationships, and Bi-LSTMs can handle sequential dependencies. These models together would be able to learn more informative and discriminative representations.

4. Better generalization:

Through this combination, the model can generalize better to unseen data by utilizing diverse approaches to capture various patterns.

Combining all the aforementioned benefits, our Feature fusion model forms a robust and resilient framework that magnifies the functionalities of these components in identifying the abiotic stress in plants based on the biological sequential pattern analysis of amino acids. This approach surpasses the limitations of the traditional methods by ensuring an elevated outcome.

Algorithm explanation

Input sequences denoted as P are collected and then a PSSM is generated with dimensions 20 × 20 and then A represents the PSSM. Aaug denotes the Augmented matrix and then the original and augmented PSSM is denoted as Pcom. Fcnn(Pcom) denotes the embedding vectors generated by CNN. Where f_ij represents the j^th feature value obtained by applying the i^th filter to the combined sequence. μ represents a set of RBF centers and RBF(μ,Fcnn(Pcom)) represents the combined feature embedding vectors obtained from cnn and rbf Where rbf_ij represents the RBF activation value for the i^th CNN feature vector and the j^th RBF center. Let F_com(P) represent the combined feature vectors from the CNN and RBF layers. h_t^(f) be the hidden state at time step t for the forward LSTM and h_t^(b) be the hidden state at time step t for the backward LSTM, o_t is the output vector at time step t. W_o is the weight matrix connecting the hidden states to the output layer. The concatenated hidden state vector at timestamp t (both forward and backward hidden states) is represented as H_t. b_o is the bias vector associated with the output layer.

Algorithm:

Algorithm for hybrid feature fusion model.

Input: Protein sequence P [seq1, seq2, seq3, ……….. seqn]

Output: Stress classification [Stressed, non-stressed]

Step-1: Process the Protein sequence and generate PSSM as P[MX20]

Pa,b:a=1⋯mandb=1⋯20 A=[Pm1,Pm2,⋯⋯⋯Pm20]

Step 2: Augment the generated PSSM as

Aaug=A[Pm1,Pm2,⋯⋯Pm20]+P|[Pm1|,Pm2|⋯⋯Pm20|] Pcom=A[Pm1,Pm2,⋯⋯⋯Pm20,Pm1|,Pm2|⋯⋯Pm20|]

Step 3: Generate Feature Embedding Vectors

Fcnn(Pcom)=[(f11,f12,⋯⋯f1m),(f21,f22,⋯⋯f2m),⋯⋯(fn1,fn2,⋯⋯fnm)]RBF(μ,Fcnn(Pcom))=[(rbf11,rbf12,⋯rbf1k),(rbf21,rbf22,⋯rbf2k),⋯(rbfn1,rbfn2,⋯rbfnk)]

Step-4: Classification of stressed and non-stressed

for each timestamp from 1 to T:

ht(f)=LSTMforward(Fcom(P)t,h(t−1)(f)) #forward LSTM

for each timestamp from T to 1:

ht(b)=LSTMbackward(Fcom(P)t,h(t−1)(b)) #backward LSTM

Combine Hidden states:

for each timestamp t from 1 to T:

Ht=[ht(f);ht(b)]

Output:

Ot=softmax(wo∗Ht+bo)

DOI: 10.7717/peerj-cs.2149/table-6

Results

The reliability and quantitativeness of our model are evaluated by evaluating the data with other architectures. The evaluation results and comparisons of different architectures are enclosed in this part.

Distribution of occurrence of the amino acids

Our work analyzed the composition of amino acids in SNARE and non-SNARE sequences by computing the frequency of occurrence of amino acids in each sequence is shown in Fig. 5.

Comparative analysis in identifying SNARE protein with CNN and other networks

As SNARE proteins constitute a fraction of the occurrence of abiotic stress in plants, we examined the input data by applying several algorithms like CNN, LSTM, CNN with Bi-LSTM, and CNN with RNN to evaluate the performance of these models with our CNN model. Table 4 shows the comparative analysis. Our model showed remarkable performance when compared with other machine learning algorithms. The hyperparameters of the model are shown in Table 5.

Table 4:

Comparative analysis of various model.

Classifier	Accuracy	Sensitivity	Specificity	MCC
CNN	65.1	80.4	55.7	0.4
CNN+RNN	68.6	75.5	66.1	0.2
CNN+Bi-LSTM	69.4	81.3	67.8	0.3
LSTM	70.1	82.6	58.5	0.1
Proposed model	74.6	88.8	73.1	0.4

DOI: 10.7717/peerj-cs.2149/table-4

Table 5:

Model hyperparameters.

Hyper parameters	Values
Epochs	50
Batch size	32
Learning rate	0.001
Optimizer	Adam

DOI: 10.7717/peerj-cs.2149/table-5

The composition of the protein sequences is visualized by calculating the similarity distance between them and the same is shown in Fig. 6. The compositions of the input patterns involved in identifying the abiotic stress in plants are calculated by observing the proportion of the SNARE proteins involved and the Fig. 7 shows the distribution of the input data.

To evaluate the correctness of the model in classifying the stressed and non-stressed input patterns is represented in the form of a confusion matrix as shown in Fig. 8. A graphical representation to predict the performance of the model on various thresholds to distinguish and classify the positive input data and the negative input data is shown in Fig. 9. The consolidated analysis of the performance metrics used to evaluate the classification of the stressed and non-stressed input data is shown in Fig. 10.

Visualization and analysis of feature space

Our study employs various visualization techniques to enhance our understanding of the feature space for identifying abiotic stress in plants using SNARE protein sequences. We utilize t-SNE (t-distributed Stochastic Neighbor Embedding), manifold Discovery and analysis (MDA), assessments of noise reduction, and examinations of intermediate layers to comprehensively analyze our model’s efficacy and interpretive aspects across multiple dimensions. The t-SNE visualization of input data is shown in Fig. 11.

Conclusion

Predicting abiotic stress in plants is essential in today’s scenario to safeguard agricultural productivity. SNARE proteins play a key role in regulating the metabolism of the amino acids that lead to the occurrence of abiotic stress in plants. Identification of SNARE proteins is of major concern in biological computation. SNARE proteins play a key role in regulating the metabolic activity of the cellular structure; hence, it is essential to develop models to identify their occurrence in the protein sequences. In our study, we proposed a hybrid model by replacing the shortcomings of traditional machine learning algorithms used to identify hidden patterns in amino acid sequences. In our approach, we extracted the features from the sequences with the help of a CNN, and then with the help of a radial basis neural network, we generated the high-dimensional feature vectors. To detect the stressed and non-stressed sequences we applied bi-directional LSTM to classify the sequences. We applied many experiments to validate the performance of our approach with the existing models. The experimental results obtained using our approach surpassed the existing methods. To our knowledge, this is the first approach used to identify the presence of SNARE proteins that detect the occurrence of abiotic stress in plants. Furthermore, our approach may facilitate the discovery of underlying functionalities of different proteins in the future by revealing the hidden structure of the sequences with the help of feature extraction methods.

Supplemental Information