Trainable computation in many-to-many molecular networks

Molecular networks have functions encoded by the binding interactions of their components which are evolved through mutations and selection. However, networks might also need to learn new functions based on examples of behavior presented within a cell's lifetime, where binding interactions are not easily tunable. We show that a large class of molecular networks–networks with many-to-many interactions–can, nevertheless, be trained to learn associations on within-lifetime timescales through a local, biologically plausible learning process inspired by energy-based machine learning. Training exposes the network to external stimuli and corresponding target responses, while simple auto-regulation adjusts expression levels of "hidden" species that serve as training parameters analogous to weights of neural networks. We explore how the same simple feedback rule in different environmental contexts can cause cells to undergo Pavlovian conditioning, learn quantitative input-outputs in a supervised manner, and learn generative behaviors relevant for bet-hedging. These findings establish a broadly applicable, model-free mechanism by which molecular systems can reprogram their own function within a lifetime.