Physical Neural Nets Model for Eukaryotic Receptor Signaling

Receptor Tyrosine Kinases (RTKs) often have multiple phosphorylation sites whose dynamics can be modeled as a Markov chain over a hypercube. However, how these receptors store information about their environment using this bit-like structure is poorly understood. We present a physical learning framework to explore how diverse behaviors can emerge out of simple learning goals and biophysically inspired limitations. We implement chemical complexity by parameterizing reaction rates to a smaller number of first and second order occupancy dependency parameters. With the construction of a chemicals network topology and parameterized transition rates we encode desired biological behaviors as functions of the steady state of chemical species concentrations. As a test case, we searched for networks that perform proofreading and found surprising strategies that outperform traditional proofreading models. We also discuss how cells may use the multi-site barcode to encode ligand identities and concentrations.