Generative Neuromorphic Programming of Mammalian Cells

Abstract

Cells make decisions through densely interconnected regulatory networks that compute graded, analog responses. Synthetic biology has made cells programmable, but the field's dominant abstractions poorly match the analog nature of cells. Designing analog circuits that work in cells is challenging: predicting how components combine remains difficult, and the vast combinatorial space of parts, weights and topologies exceeds what experiments can explore. Here we establish generative programming of mammalian cells through a neuromorphic framework that compiles desired behavior into DNA. Composable regulatory devices built from RNA-targeting endoribonucleases (ERNs) supply signed weights and nonlinear activation, enabling multi-layer circuits with diverse multi-input analog responses. To predict and design such circuits, we develop biomorphic neural networks (BMNs): a compositional architecture that mirrors biological interactions, in which each basic process (transcription, translation, cleavage) is captured by a reusable neural block learned from the behavior of whole circuits that contain it. Recomposed into architectures absent from training, these models predict responses with errors often comparable to variability between experimental repeats. Inverting this process, our software package, the biocompiler, jointly optimizes topology, parts, and weights to realize specified target behaviors. Challenged with three target behaviors, the biocompiler proposed three previously unseen architectures; we built all three, and each reproduced its target behavior in mammalian cells in a single design pass without manual tuning. Generative neuromorphic programming thus offers a systematic route to engineering the analog computation native to living cells.