Combinatorial decision-making driven by surface condensation in multicomponent mixtures

Unlike artificial neural networks, biological systems rely on physical interactions in crowded environments to perform complex information processing tasks. An emerging and characteristic example is condensates—self-organized biomolecular assemblies where interactions among many components encode for emergent multiphase behavior. Here, we explore how the physics underlying condensates can be tuned for solving an important class of computations, i.e. classification. Bridging non-equilibrium statistical physics with differentiable optimization, we train molecular systems to perform classification through the formation of different phases. In our model of surface condensation, input species are surface-localized (analogous to transcription factors bound on DNA) and recruit catalytic species (polymerase or repressor) from the bulk (nucleoplasm) to drive output function (gene activation or silencing). We find that the addition of regulatory species, akin to hidden nodes in machine learning, greatly improves the expressivity of these molecular ensembles in solving classification problems. Finally, we show that trained molecular interactions can solve distinct and multiple tasks by simply changing the concentration of the nucleoplasm, analogous to transfer learning. These results provide a framework to discuss how phase transitions in heterogeneous mixtures of molecules can drive information processing.