A genetically encoded local learning rule enables physical learning in engineered bacteria
Abstract
Training physical neural networks directly in matter remains difficult because most platforms do not store and update weights within the same physical substrate. Here we show that engineered Escherichia coli can implement a genetically encoded local learning rule acting on persistent biological memory. We introduce memregulons: bacterial memory elements in which coupled plasmids store an analogue weight as the relative abundance of P1 within the P1 + P2 plasmid pool. Environmental inputs activate local promoters, and a shared sublethal kanamycin signal converts promoter activity into differential growth that lowers the stored P1 fraction. Across single-strain trajectories, combined plate-reader and flow-cytometry measurements are consistent with activity- and variance-dependent weight change. At the single-cell level, repeated negative learning reshapes the stored weight distribution as it approaches the lower boundary. In mixed populations and co-cultures, one shared negative learning signal selectively rewrites active memregulons, enabling externally routed supervised tic-tac-toe lessons. We then generalise the architecture across orthogonal chemical inputs and combinatorial promoters, and combine measured co-culture updates with external game analysis and simulate negative pruning in nonlinear-classifier architectures. These results establish physical learning in living matter through local negative updates of genetically encoded analogue weights.
Related
- Genetically engineered RNA inducers of proximity for protein regulation
- ELM-CLAIM: a claim-calibration framework for adaptive, memory-like, learning-like, and goal-directed claims in engineered living materials.
- Mapping the Nonhistone Interactome of the TRIM24 Bromodomain Using Genetically Encoded Complementary Photocrosslinkers
- Advancing Targeted Nanoparticle mRNA Delivery Using Machine Learning
- EnzyREx: deep learning enables reaction-based comprehensive enzyme screening
BibTeX
@article{prakash2026genetically,
title = {{A genetically encoded local learning rule enables physical learning in engineered bacteria}},
author = {Satya Prakash and Clenira Varela and Mark Walsh and Roberto Galizi and Mark Isalan and Alfonso Jaramillo},
journal = {Nature Communications},
year = {2026},
month = oct,
doi = {10.1038/s41467-026-78417-9},
url = {https://www.nature.com/articles/s41467-026-78417-9},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A genetically encoded local learning rule enables physical learning in engineered bacteria
AU - Satya Prakash
AU - Clenira Varela
AU - Mark Walsh
AU - Roberto Galizi
AU - Mark Isalan
AU - Alfonso Jaramillo
PY - 2026
DA - 2026/10/16
JO - Nature Communications
DO - 10.1038/s41467-026-78417-9
UR - https://www.nature.com/articles/s41467-026-78417-9
AB - Training physical neural networks directly in matter remains difficult because most platforms do not store and update weights within the same physical substrate. Here we show that engineered Escherichia coli can implement a genetically encoded local learning rule acting on persistent biological memory. We introduce memregulons: bacterial memory elements in which coupled plasmids store an analogue weight as the relative abundance of P1 within the P1 + P2 plasmid pool. Environmental inputs activate local promoters, and a shared sublethal kanamycin signal converts promoter activity into differential growth that lowers the stored P1 fraction. Across single-strain trajectories, combined plate-reader and flow-cytometry measurements are consistent with activity- and variance-dependent weight change. At the single-cell level, repeated negative learning reshapes the stored weight distribution as it approaches the lower boundary. In mixed populations and co-cultures, one shared negative learning signal selectively rewrites active memregulons, enabling externally routed supervised tic-tac-toe lessons. We then generalise the architecture across orthogonal chemical inputs and combinatorial promoters, and combine measured co-culture updates with external game analysis and simulate negative pruning in nonlinear-classifier architectures. These results establish physical learning in living matter through local negative updates of genetically encoded analogue weights.
SN - 2041-1723
ER -