Training physical neural networks (PNNs) directly in their substrate, without external algorithms, remains a central challenge for neuromorphic computing and synthetic biology. Synthetic gene circuits can implement logic gates and even neural network architectures in bacteria, but they rely on pre-programmed configurations and do not learn autonomously. Here we engineer a PNN in Escherichia coli that learns autonomously through a DNA-encoded local learning rule. Our system—a DNA memory based on persistent copy-number tuning in duplicate-origin plasmids—stores synaptic weights as population-level plasmid ratios and converts task performance into persistent weight updates through antibiotic-mediated population modulation. Using only negative feedback, bacterial agents learn from experience through reinforcement learning to increase their proficiency in simplified decision trees in the game of tic-tac-toe and other 3×3 board games. This strategy generalises to any gene circuit, including modules with non-linear interactions, which we exemplify by realising a fundamental XOR gate with designed weights. Because the learning rule is local, activity-dependent, and autonomous, it scales to larger architectures and can extend to other rules that minimise prediction error in networks of cells carrying memregulons, with or without intercellular communication. These results demonstrate autonomous learning in living cells and outline a route to training biological computing architectures relevant to biotechnology and medicine.
Alfonso Jaramillo (Sun,) studied this question.