Search Results

Molecules with designed interactions can serve as substrates for information processing; demonstrated examples include algorithmic self-assembly of DNA tiles and DNA strand displacement reactions. These two well-established paradigms correspond to solid-phase-like behavior, where spatially structured molecular assemblies grow by crystalline attachment, and gas-phase-like behavior, where a dilute mixture of freely diffusing molecules is governed by mass-action kinetics of reactions. A third paradigm, liquid-liquid phase separation, is a fundamental phenomenon in physical and biological systems, giving rise to membraneless compartments that facilitate complex information processing. Recent studies have shown that multicomponent liquid mixtures described by lattice models are analogous to Boltzmann machines and can implement neural computation. Mean-field models have also suggested a connection to Hopfield networks and illustrated complex decision-making during condensation from a reservoir. As an alternative to previous approaches to train liquid interaction energies using backpropagation through a loss function, here we derive "local" learning rules based on contrastive Hebbian learning that may be more biologically and physically plausible. The trained systems demonstrate a range of computational tasks, including associative recall, linear and nonlinear input-output relations, pattern classification, and spatial phase separation. Our work suggests that the computational potential of liquid-phase molecular systems could be unlocked in real physical systems with local learning rules.