CRISPR-based neuromorphic computing for solving regression and classification

The CRISPR-dCas9 system has emerged as a versatile platform for programmable gene regulation, offering unique advantages in modularity and orthogonality for constructing synthetic genetic circuits. Here, we present a novel architecture for biomolecular neural networks based on dCas9, guide RNAs, and antisense RNA sequestration. Through mathematical modeling and steady-state analysis, we demonstrate that this system functions as a molecular perceptron with a threshold activation function analogous to a saturated rectified linear unit (ReLU). However, a critical challenge in scaling these circuits is competition for the finite dCas9 pool, whose expression must remain low to avoid cytotoxicity. We address this constraint by developing a resource-aware design framework and characterizing how shared dCas9 availability affects network performance. Our results show that for classification tasks, decision boundaries remain invariant under resource competition, while for regression tasks, node thresholds are preserved despite sensitivity in output magnitude under heterogeneous binding conditions. We demonstrate the computational capabilities of this platform through both linear and nonlinear classification problems, as well the approximation of a band-pass function as a proof-of-concept regression task. This work expands the repertoire of molecular mechanisms capable of computation and establishes design principles for implementing CRISPR-based neuromorphic circuits that can execute complex computational tasks within the biochemical constraints of living cells.