A Linear Approach to Fault Analysis and Intervention in Boolean Systems

Let a logical domain be defined as

Let $u$ be the input variable and $y$ be the output variable. For the logical domain,$\{u,y\}\in{\mathcal{L}}$. $d$ is a logical variable for a drug such that $d=1$ means the drug is applied and $d=0$ means the drug is not applied, i.e., $d\in{\mathcal{L}}$. $f$ is a ternary variable for a fault with values {sa-1, sa-0, no-fault} which are represented as $f\in\{-1,0,1\}$.

For $\alpha$ number of primary inputs, the input state (Boolean vector) is defined as $U=(u_{1},\ldots,u_{\alpha})$, $U\in\{0,1\}^{\alpha}$. Similarly, for $\beta$ number of primary outputs, the output state is $Y\in\{0,1\}^{\beta}$; for $\lambda$ number of drugs, the drugs state is $D\in\{0,1\}^{\lambda}$; and for $\gamma$ number of faults, the fault state is $F\in\{-1,0,1\}^{\gamma}$. The notations $\alpha,\beta,\gamma,$ and $\lambda$ are used in this same context throughout the manuscript.

A faulty BCN can be modeled by considering the faults as separate inputs to the system provided a fault-free BCN and the probable gene mutations (fault points). The drugs are also treated as external input variables in the known BCN topology. Without loss of generality, it is assumed that a fault-free BCN shown in (1) can be converted to define a faulty BCN as:

where $\widehat{\Psi}:\{0,1\}^{\alpha}\times\{-1,0,1\}^{\gamma}\times\{0,1\}^{\lambda}\times\{0,1\}^{N}\rightarrow\{0,1\}^{N}$ and $\widehat{\Gamma}:\{0,1\}^{\alpha}\times\{0,1\}^{N}\rightarrow\{0,1\}^{\beta}$. Here $(\widehat{\Psi},\widehat{\Gamma})$ is a faulty version of a fault-free BCN $(\Psi^{\ast},\Gamma^{\ast})$. The faults ($F$) and drugs ($D$) are assumed to be constant over time, hence, the time stamp is not mentioned.

Similarly, a fault-free BM shown in (2) can be converted to define the faulty BM as:

where $\widehat{\Gamma}:\{0,1\}^{\alpha}\times\{-1,0,1\}^{\gamma}\times\{0,1\}^{\lambda}\rightarrow\{0,1\}^{\beta}$. $\widehat{\Gamma}$ is a modified BM after introduction of faults. Since the time indices of the input and output vectors of the BM are continuous in nature and devoid of causality, the indices can be omitted for all purposes. Hence (4) and all the subsequent analysis of BM no longer have a time stamp.

The drugs ($D$) are inhibitory and are inherently different from the inputs ($U$). The input $U$ is mostly not under the control of the therapist for in situ modeling of a gene regulatory system. The inputs can be growth factors, hormones, oxygen, or different kind of molecular stresses [8], whereas, drugs are user designed molecules targeted at known locations in Boolean systems. Hence the problem of control design in biological systems is finding drugs $D$. In both cases shown in (3) and (4), $\widehat{\Psi}$ and $\widehat{\Gamma}$ are assumed to be known because $\Psi^{\ast}$ and $\Gamma^{\ast}$ are known from their construction.

For the conversion of logical expression into the linear form, the variables in the logical domain are required to map in the vector form. Let a vector (delta) set $\Delta_{k}$ be defined as $\Delta_{k}=\{\delta^{i}_{k}|\,i=1,2,\ldots,k\}$, where $\delta^{i}_{k}$ is the $i$-th column of an identity matrix ${\rm{I}}_{k}$. A set $\Delta_{2}$ is used to denote the binary values, such that $\{1,0\}\sim\{\delta_{2}^{1},\delta_{2}^{2}\}$ respectively. Therefore, the variables defined in the logical domain are defined in the vector form given as, $\{{\mathbf{u,y,d}}\}\in\Delta_{2}$. Similarly, a vector set $\Delta_{3}$ is required to show a ternary variable in the vector form. Therefore, a vector form of fault variable is defined as ${\mathbf{f}}\in\Delta_{3}$, where $\delta_{3}^{1}$ shows a sa-1 fault, $\delta_{3}^{2}$ shows a sa-0 fault, and $\delta_{3}^{3}$ shows a no-fault condition. Although $f$ is not a binary variable, such multi-valued mapping has been done earlier in the literature [17].

The logical operators $\widehat{\Psi}$ and $\widehat{\Gamma}$ are converted into linear operators using the semi-tensor product (STP) approach formulated in [28, 29, 30, 31, 32].

For an STP of $p$ k-valued logical variables, a mapping $\ltimes_{i=1}^{p}:\Delta_{k}\rightarrow\Delta_{k^{p}}$. For $\alpha$ number of input variables, an input state $U$ is represented by a vector ${\mathbf{U}}={\mathbf{u}}_{1}\cdots{\mathbf{u}}_{\alpha}=\ltimes^{\alpha}_{i=1}\mathbf{u}_{i}$. As ${\mathbf{u}}_{i}\in\Delta_{2}$, ${\mathbf{U}}\in\Delta_{2^{\alpha}}$. Similarly, the output vector ${\mathbf{Y}}\in\Delta_{2^{\beta}}$, the fault vector ${\mathbf{F}}\in\Delta_{3^{\gamma}}$, and the drug vector ${\mathbf{D}}\in\Delta_{2^{\lambda}}$.

A structure matrix is derived from the network structure, and thus represents the characteristics of the network.

The dimensions of a structure matrix depend upon the dimensions of individual ${\mathbf{x}}_{i}$. In the matrix form, ${\rm{M}}_{\sigma}=[\delta_{2^{m}}^{i_{1}},$ $\ldots,\delta_{2^{m}}^{i_{2^{n}}}]=\delta_{2^{m}}[i_{1},\ldots,{i_{2^{n}}}]$.

As mentioned in Section 2.3, a fault is modeled as an external input. Therefore, each fault changes the structure matrix of the network. If a node $x$ is observed (by gene sequencing) to be mutated, its value is modified to $x^{\ast}$. If the observed mutation is mapped as fault $f$, then the change in the value of $x$ is given by the relation in matrix form as $\mathbf{x}^{\ast}={\rm{M}}_{f}\,\mathbf{x\,f}$, where $\{\mathbf{x}^{\ast},\mathbf{x}\}\in\Delta_{2},\mathbf{f}\in\Delta_{3}$. In this expression, $\rm{M}_{f}=\delta_{2}[1,2,1,1,2,2]$ is the structure matrix showing the effect of fault for different values of node $x$. For instance, if the original value of a node $x$ is 0 (${\mathbf{x}}=\delta_{2}^{2}$) and the fault observed is sa-1 (${\mathbf{f}}=\delta_{3}^{1}$), then ${\mathbf{x}}^{\ast}={\rm{M}}_{f}\delta_{2}^{2}\delta_{3}^{1}=\delta_{2}^{1}$. i.e., the value of $x$ changes to 1 as an effect of fault.

Similarly, the application of an inhibitory drug $d$ blocks the node $x$ by changing its value to 0. When the drug is not applied, the node $x$ retains its value. In logical form, it is represented as $x^{\ast}=x\cdot\bar{d}$ [11]. Using the STP techniques [31], this expression can be written in vector form as $\mathbf{x}^{\ast}={\rm{M}}_{D}\,\mathbf{x\,d}$, where $\{\mathbf{x}^{\ast},\mathbf{x,d}\}\in\Delta_{2}$ and ${\rm{M}}_{D}=\delta_{2}[2,1,2,2]$. ${\rm{M}}_{D}$ shows a structure matrix incorporating the effect of a drug applied at node $x$. The derivation of structure matrices ${\rm{M}}_{f}$ and ${\rm{M}}_{d}$ are provided in Section A of the Appendix.

The matrices ${\rm{M}}_{f}$ and ${\rm{M}}_{D}$ are constant. For every fault $f$ at a node $x$, ${\mathbf{x}}$ has to be replaced by ${\rm{M}}_{f}{\mathbf{xf}}$ to obtain a linear form. Similarly, for every inhibitory drug $d$ at a node $x$, ${\mathbf{x}}$ has to be replaced by ${\rm{M}}_{D}{\mathbf{xd}}$.

Let ${\mathcal{B}}$ be a set of matrices with all binary elements.

The network information of BM is available in the structure matrix ${\rm{H}}$ represented by (8). Therefore, the fault analysis and intervention results are derived from the structure matrix ${\rm{H}}$.

The following results discuss the necessary and sufficient conditions for detecting the existence of a particular fault using an arbitrary input.

The test set generated by this method is not optimal, but it is useful to know the important homeostatic inputs and the fault coverage.

Due to the limited number of homeostatic inputs, many of the fault vectors may remain undetectable in a real biological system. This affects the fault coverage. The following theorem provides a way to use multiple input vectors for unique fault detection.

In Proposition 4.3, if some fault vector results in ${\mathcal{U}_{\mathbf{F}}}=\emptyset$ then that fault vector is said to be undetectable. Such faults may not be harmful in the pathways under consideration, but may be harmful in other dependent pathways. Therefore it is important to detect such faults. In Corollary 4.2, if set ${\mathcal{F}_{{\mathcal{U}}^{\ast}}}$ represents equivalent fault vectors, then fault detection becomes ambiguous or undetectable. For these conditions, a new reporter design is necessary. Some faults may not be detectable at primary outputs, but may be detectable at the output of internal blocks. Therefore, observability of faults at every internal node is used to decide a new reporter.

Let a set of undetectable fault vectors for a set of input vectors $\mathcal{U}$ be denoted as $\overline{\mathcal{F}}_{\mathcal{U}}$. Set $\overline{\mathcal{F}}_{\mathcal{U}}$ contains all the faults for which ${\mathcal{U}_{\mathbf{F}}}=\emptyset$ and the faults which are equivalent. Algorithm 1 results into best possible reporter(s) for improved fault detection. ${\mathcal{F}^{ij}_{\mathcal{U}}}$ indicates a set of detectable faults considering node $x^{i}_{j}$ as ‘reporter’. Reporter is an additional input of the network which provides additional information for fault detection. Selection of each reporter augments the output vector by one. Algorithm 1 results in the set of best possible reporters, ${\mathcal{X}}$ (best reporter at ${\mathcal{X}}(1)$). It is clear from Algorithm 1 that the complexity of this algorithm is $O(2^{\sum^{m}_{i=1}{n_{i}}})$ or $O(2^{\gamma+\lambda})$ (using (7)). The modular design keeps the structure matrices for the individual blocks accessible for algebraic manipulation.

It is possible in principle to put some additional constraints depending on some prior biological knowledge of certain nodes in choosing the best reporter in each iteration. The accessibility of the new reporters is restricted to the output nodes of the primary sub-blocks only. The number of such sub-blocks is $\gamma+\lambda$, which is small compared to the total number of internal nodes in the system keeping the computational complexity of the algorithm manageable.

Improvement in controllability: For Proposition 4.12, if ${\mathcal{D}_{\widetilde{\mathbf{F}}}}=\emptyset$, then a new drug is required such that it can eliminate the effects of the faults. Let $\mathbf{d}_{\lambda+1}$ be a new drug. Then,

Naturally, the effect of faults will be maximum at the downstream protein(s) in the pathways. Although the secondary block does not have any fault or drug, a new drug can be targetted at the primary output of that block. These targets facilitate better possibilities of the drug intervention. Therefore, the best possible target site for the drug is searched from output towards input, level wise. Depending upon the target location of the drug, there are two possibilities. (i) Target point is internal node ($x^{i}_{j}$ from primary block) of a BM, and (ii) target point is one of the primary outputs (secondary block) of a BM. For these two cases, similar to (8), equation of the output can be written as

Details of the derivation process are provided in the Appendix (Section C). Algorithm 2 shows the process of improvement in controllability. However, all faults cannot be controlled with inhibitor type drugs. Some activator drugs are also required for improved control.

In (9), the information about network dynamics is shown by structure matrix ${\rm{L}}$. As the secondary block is fault-free, a structure matrix ${\rm{L}}$ alone is sufficient for the fault analysis. The cyclic attractors in the state transitions of the BCN can be calculated from the diagonal of its structure matrix [31]. The number and sizes of the attractor cycles in the network change with the fault present. For each input vector and fault vector, a structure matrix ${\rm{L}}_{\mathbf{UF}}$ and its corresponding cycles are estimated. Assume that there are $s$ different attractor cycles in the network. The cycles are numbered in the increasing size of their lengths. The length of an $i^{th}$ cycle is $l_{i}$. For a cycle of length $k$, the corresponding diagonal elements of the structure matrix $({\rm{L}}_{\mathbf{UF}})^{k}$ become one [31]. Therefore, the diagonal values are used to identify changes in the attractor cycles for different conditions of the inputs and faults.

Note that ${\rm{A}}\vartriangle{\rm{B}}$ is the symmetric difference between the matrices ${\rm{A}}$ and ${\rm{B}}$. $Tr({\rm{A}})$ is the trace of a matrix ${\rm{A}}$.