Pattern recognition in living cells through the lens of machine learning

Abstract

At a coarse level, pattern recognition within cells involves sensing of environmental signals by surface receptors, and activating downstream signalling pathways that ultimately drive a transcriptome response, enabling biological functions such as differentiation, migration, proliferation, apoptosis or cell-type specification. This kind of decision-making process resembles a classification task that, inspired by machine learning concepts, can be understood in terms of a decision boundary: the combination of inputs relative to the classification region defined by this boundary defines context-specific responses. In this report, we contextualize machine learning concepts within a biological framework to explore the structural and functional similarities (and differences) between artificial neural networks, signalling pathways and gene regulatory networks. We take a preliminary look at neural network architectures that may better suit biological classification tasks, explore how learning fits into this paradigm, and address the role of competitive binding in cellular computation. Altogether, we envision a new research direction at the intersection of systems and synthetic biology, advancing our understanding of the inherent computational capacities of signalling pathways and gene regulatory networks.

Keywords: biocomputation; neural networks; pattern recognition; signalling pathways; synthetic biology.

Figure 1.

Pattern recognition in living cells. (A) Single-layer perceptrons can implement linear decision boundaries, while (B) multilayer perceptrons feature nonlinear and more complex decision boundaries. (C) For illustrative purposes, consider a simplified scenario in which environmental signals $x_1$ and $x_2$ are recognized by receptors on the cell surface, activating downstream signalling pathways that ultimately drive a transcriptome response, enabling biological functions such as differentiation, migration, proliferation, apoptosis or cell-type specification. Across various regulatory layers—including ligand–receptor interactions, downstream signal transduction and gene regulatory networks—the biological counterpart to artificial neural networks (ANNs), named biomolecular neural networks (BNNs), offers a plausible framework for understanding how information is processed to produce a cellular response. Analysing the input space for a given combination of $x_1$ and $x_2$ reveals a decision boundary that determines whether an appropriate cellular response is activated (highlighted grey area) or not.

Figure 2.

Building multilayer perceptrons using chemical reactions. Biomolecular perceptrons can be implemented through a variety of chemical reactions that are ubiquitous to signalling pahtways and gene regulatory networks: phosphorylation/dephosphorylation cycles, where a kinase Y adds a phosphate to the $P^{\ast }$ protein, and converts it into $P$ at a rate $k_1$, whereas a phosphatase $Z$ removes the phosphate of $P$ and converts it into $P^{\ast }$ at a rate $k_2$; molecular sequestration, where a molecule $Z$ binds to a molecule $Y$ to form an inactive complex at a rate defined by $\gamma$; and enzymatic reactions, where a substrate $Y$ binds to an enzyme $Z$ to form a complex $C$, at a rate defined by $\gamma$, that catalyses $Y$’s degradation, after which $Z$ is able to dissociate from the complex into its free form. (B) The Sotfplus-like and sigmoid-like activation functions arise when mapping the perceptron's outputs to its respective inputs, and whose shape depends on the rate constants of each reaction. (C) The single-layer perceptron exhibits a tunable decision boundary. By cascading this unit into layers, we obtain more complex, nonlinear decision boundaries. (D) Architecture of a multilayer perceptron with a hidden layer made of Node 1 (linear decision boundary on top) and Node 2 (linear decision boundary at the bottom).

Figure 3.

Learning is conceptualized as reshaping the decision boundary to adapt to a new environment. (A) A representative example of a multilayer perceptron processing the concentrations of molecules $x_1$ and $x_2$ to promote cell growth. (B) Evolution is framed as an iterative parameter search that introduces random mutations to the BNN, leading to adaptation. (C) Decision boundaries associated with the multilayer perceptron before perturbation of the environment. From left to right, the outputs of Node 1, Node 2 and Node 3. (D) The weights associated with Node 1 ($w_{1,n}$) and Node 2 ($w_{2,n}$) changed to that the decision boundary of Node 3 adapted to those environmental perturbations.

Figure 4.

Competitive binding enables complex decision boundaries with a minimal set of reactions. (A) Schematic representation of a single-layer neural network that generates a linear classifier, where each neuron is based on the sequestration reaction between a sigma ($\sigma$) and its sigma factor (anti-$\sigma$) given at the ${\gamma }_1$ rate. The sigma factor is allowed to bind to the RNAp at a rate ${\gamma }_2$ to form a complex $c_{\mathrm{1,2}}$ that enables transcription. (B) When isolated, each neuron presents a linear decision boundary (first column, in grey) that turns nonlinear when considering competitive binding between the sigma factor and either the anti-sigma or the RNAp, with increasing complexity when increasing the competitive binding ratios (${\gamma }_2/{\gamma }_1$).