Turning Gene Circuits into Neural Networks with CRISPR

Synthetic biology has long pursued the goal of programming cells to perform custom computations, from simple logic gates to complex decision-making networks. As researchers refine these capabilities, integrating machine learning architectures into biological systems represents a frontier where cellular processes could execute complex information processing through molecular interactions.

Biological systems already process information through convoluted molecular networks that sense signals, make decisions, and coordinate responses. These natural networks integrate inputs to produce outcomes during processes like cell signaling, where multiple inputs converge to determine whether a cell divides, migrates, or changes state. Scientists have proposed molecular circuit designs that could execute neural network operations using the biochemical machinery of living cells. These molecular circuits would process biochemical signals as inputs and generate gene expression responses as outputs.

A group of scientists from Carnegie Mellon University, Vanderbilt University School of Medicine, and the University of Pittsburgh presents a theoretical and computational analysis of CRISPR-based molecular circuits, using mathematical modeling and numerical simulations to evaluate their behavior under biologically relevant constraints. The collaboration combined expertise in mathematical modeling, molecular physiology, and systems biology, providing the interdisciplinary foundation needed for this analysis. Their work shows that the CRISPR gene editing system can be repurposed to build neural network architectures whose theoretical behavior is consistent with known cellular constraints.

Understanding the CRISPR Platform

The CRISPR platform originated as a bacterial immune defense mechanism that bacteria use to remember and combat viral infections. At its core, the system relies on a protein called Cas9 that can be guided by RNA molecules to locate and cut specific DNA sequences with precision. Scientists discovered they could reprogram this system by changing the guide RNA sequences, allowing them to target virtually any gene of interest.

For this research, the authors used a modified version called dCas9, where the “d” stands for “dead” or catalytically inactive. This variant loses its cutting ability but retains the capacity to bind DNA at precise locations. When researchers fuse dCas9 with regulatory domains, the protein can either activate or suppress gene expression without making permanent changes to the genome, making it an exceptionally versatile tool for controlling cellular behavior on demand. Because the system’s targeting rules can be mathematically formalized, CRISPR serves as a suitable substrate for information processing.

From Gene Regulation to Neural Computation

Since dCas9 alters gene expression based on molecular cues, it effectively performs the role of logic gates and activation functions used in computing. The research team developed a theoretical architecture for molecular perceptrons, the basic building blocks of neural networks, using dCas9 proteins, guide RNAs, and antisense RNAs. In this context, “neural network” refers to a mathematical structure, not biological neurons or synapses, where outputs correspond to gene expression levels rather than electrical signals.

To undestand this connection, consider how digital and molecular systems map onto each other. The following provides a rough translation between computer terms and their molecular counterparts:

Digital Component → Molecular Equivalent

  • Logic gate input → Guide RNA concentration
  • Activation threshold → RNA binding equilibrium
  • Output signal → Gene expression level
  • Weight parameter → Effective binding affinity

Guide RNAs direct dCas9 to specific genes, activating the system when present at sufficient concentrations. Antisense RNAs act as inhibitors by binding to guide RNAs and preventing them from interacting with dCas9. This molecular competition creates a threshold mechanism similar to how neurons in the brain decide whether to fire. When antisense RNA levels are high, the system stays quiet. When guide RNA levels exceed the inhibitory threshold, the circuit activates.

Mathematical Behavior and Activation Patterns

Through mathematical analysis at steady state (meaning once the system has settled), the scientists showed that their system behaves like a rectified linear unit, or ReLU, a mathematical rule commonly used in artificial intelligence that increases output only after input passes a specific threshold. This resembles transcription factor cooperativity, where binding of multiple factors produces sharp transitions in gene expression once a critical concentration is reached. The authors note that this behavior emerges from the steady-state equations of the reaction network, while transient dynamics and stochastic noise remain subjects for future analysis.

To understand how this works in principle, consider a simple example. When guide RNA concentrations remain low and antisense RNA levels stay high, the circuit produces no output because most guide RNAs are sequestered. As guide RNA production increases and crosses the threshold where free molecules can bind dCas9, output activity begins rising proportionally with input strength. Eventually, the system reaches saturation when all available dCas9 molecules are bound, preventing further increases regardless of additional guide RNA. This input-output pattern matches the behavior of activation functions used in digital neural networks.

Managing Shared Molecular Resources

Having established the behavior of single circuits, the analysis extends to multicircuit systems where resource competition becomes crucial. When multiple circuits operate simultaneously, they all compete for the same restricted pool of dCas9 proteins. Because high dCas9 expression can harm cells, the protein must be kept at low concentrations. Elevated dCas9 levels impose metabolic burden by consuming cellular energy for protein synthesis and sequestering ribosomes that would otherwise translate essential housekeeping proteins. Additionally, excess dCas9 may interfere with endogenous regulatory machinery by binding unintended genomic sites. In simple terms, too much of this protein stresses the cell, so circuits must share a limited supply. These cellular costs set the upper bound used in the simulations for feasible dCas9 concentration. This creates a bottleneck where one circuit’s consumption of dCas9 reduces availability for others.

When multiple RNA species all take from the same restricted dCas9 supply, each circuit’s responsiveness can diminish as the shared resource becomes depleted. The researchers handled this constraint by developing mathematical models that predict how resource competition affects network performance. Their framework accounts for factors such as binding strengths between different guide RNAs and dCas9, allowing engineers to compensate for resource constraints during the design phase.

Building on this foundation, a central insight of the study is that resource competition does not necessarily corrupt computation, but instead reshapes it in predictable ways. Many molecular computing schemes fail under shared resource constraints, with outputs becoming unpredictable or circuits interfering destructively. This work reveals why certain computational properties survive resource competition while others do not. Specifically, the invariance of classification boundaries under dCas9 competition is not intuitive and represents a key result. This property emerges directly from the form of the steady-state equations rather than from ad hoc parameter tuning, making it a robust feature of the architecture. This distinguishes the framework from many prior molecular computing proposals, where shared resources lead to qualitative failure rather than predictable rescaling.

Classification and Regression Under Molecular Constraints

The same resource-aware design principles apply to both discrete classification problems and continuous prediction problems, though each behaves differently when molecular resources become restricted. Classification deals with category-type outputs such as identifying cell states or detecting specific conditions. As a classification example, cells must determine whether they have encountered a pathogen, entered a stress state, or received conflicting developmental signals. Regression predicts continuous values such as signal intensity or molecular concentration. As a regression example, cells must modulate responses proportionally to input strength, such as adjusting metabolic rates based on nutrient availability.

Numerical simulations revealed that classification problems remain robust under resource constraints. Decision boundaries, the lines that separate different categories in data, stay constant even when multiple circuits compete for the same dCas9 molecules. Linear classifiers could separate data into regions using straight boundaries, creating simple categorical divisions. Multilayer architectures produced curved decision boundaries capable of handling more complex patterns, such as distinguishing samples that fall within specific ranges while excluding those outside these bounds. When the group of researchers varied binding strengths between different guide RNAs and dCas9, the decision boundaries shifted position but maintained their fundamental shape and separation logic.

For regression problems, the team modeled multilayer networks that approximate band-pass functions (meaning the system responds only within a certain input range). These functions activate within specific input ranges, similar to how developmental systems respond selectively to particular signal strengths. While output magnitude varied with different binding conditions, the critical thresholds that determine when activation begins remained stable across simulations.

Comparison With Earlier Molecular Computing Systems

Previous molecular computing approaches used protein sequestration to create logic operations and toggle switches, phosphorylation cycles to build oscillators and timers, and enzymatic reactions to perform signal filtering and amplification. Each system excelled at a particular operation. Programmability and scalability, however, remained limited.

DNA strand displacement computing, another prominent approach, encodes logic through carefully designed nucleic acid sequences that undergo programmed hybridization reactions. While powerful for abstract computation, these systems typically operate in cell-free environments and require extensive sequence design for each new function. The CRISPR approach extends these capabilities by offering modular design where circuit behavior can be altered by simply changing guide RNA sequences rather than engineering entirely new proteins or redesigning entire strand displacement cascades. This flexibility allows researchers to rapidly prototype different network architectures and computational functions using the same core molecular components, with the added advantage that CRISPR components are naturally compatible with cellular contexts.

Applications and Theoretical Implications

The ability to approximate regulatory functions has potential practical applications. Many biological processes map inputs like hormones, metabolites, or environmental stressors to specific gene expression outputs. By understanding these input-output relationships, scientists could build synthetic circuits that replicate natural regulatory programs for therapeutic purposes. Such designs extend earlier efforts to map transcriptional responses into quantifiable digital functions.

Consider engineered immune cells that could distinguish between inflammatory and tumor signals by processing multiple biochemical markers simultaneously. Such cells might activate therapeutic responses only when they detect the specific combination of signals indicating disease, avoiding false positives that could trigger harmful reactions. Other applications include biosensors that respond adaptively to fluctuating environmental conditions or cells that release drugs only when specific molecular patterns are present.

These circuits perform inference with predefined input-output mappings encoded by molecular production rates and binding affinities. The weights and thresholds are set by these biochemical parameters rather than updated dynamically through a training process. This distinction matters because the circuits execute fixed computational functions rather than adapting their behavior over time.

At the system level, this work carries conceptual implications for understanding computation in living matter. If molecular circuits can implement the same mathematical operations as silicon computers, the boundary between biological and digital information processing becomes less distinct. Where silicon-based systems excel at speed and precision, molecular systems offer advantages in adaptability and energy economy. Biological circuits operate at ambient temperature using chemical energy, can self-repair, and integrate seamlessly with living tissues. These properties position molecular neural networks as complementary technologies.

However, experimental realization will require dealing with several practical concerns. Molecular reactions in cells experience significant stochastic noise, particularly when species concentrations are low, which could blur decision boundaries or introduce variability in regression outputs. Additionally, the timing of gene expression and RNA degradation occurs on timescales of minutes to hours, potentially limiting computational speed. Controlled expression systems that manage dCas9 levels precisely will be essential to avoid the cytotoxicity and metabolic burden described earlier. These dynamic factors highlight the gap between theoretical feasibility and laboratory implementation, meaning the current work serves as a blueprint rather than a finished technology that will require experimental validation before practical use.

Building Toward Practical Systems

The group’s resource-aware design framework provides blueprints for engineering more elaborate multilayer architectures. These principles allow researchers to predict how circuits would behave when constrained by cellular resources, enabling the construction of more reliable biological computers designed to function predictably when implemented.

Potential experimental testbeds could include model organisms such as E. coli or yeast, where measurable outputs like reporter fluorescence rates could validate predicted input-output relationships. The modular nature of CRISPR components means that circuits can be customized for different operations by swapping guide RNA sequences, similar to how software can be reprogrammed without changing hardware. This programmability accelerates the development cycle from concept to working prototype.

The study establishes that molecular neural network architectures are compatible with known biochemical constraints. By managing resource use and anticipating competitive effects, scientists can design biological computers capable of executing complex algorithms when constructed in cellular contexts. This collaboration between modelers and molecular biologists was crucial: mathematical analysis guided the theoretical framework, and biological insights shaped the parameter choices and constraints.

This theoretical groundwork maps how biological systems might manage computational logic within the constraints of living cells. Computation emerges naturally from properly organized molecular interactions across different physical substrates. As researchers continue refining these theoretical models and move toward experimental implementation, cells capable of executing predefined computational operations move from theoretical possibility to engineering reality.


Citation:
Montufar Leon, C., Wang, Y., Britto Bisso, F., Mehta, A., & Cuba Samaniego, C. (2025). CRISPR-based neuromorphic computing for solving regression and classification. bioRxiv. https://doi.org/10.64898/2025.12.03.692209

Related Articles

PHP Code Snippets Powered By : XYZScripts.com
★

Track This Company

Create a free account to add companies to your watchlist and get notified when patents, funding, or key developments change.