We want to decipher the biophysical drivers of molecular interactions. Protein–protein interactions are one example: with the Dickinson lab, we use PANCS selections of designed binders against many targets to ask whether a sequence-only model can learn the rules for which proteins bind which, and to map the fitness landscapes of protein–protein interactions.
How do cells make decisions from high-dimensional inputs, when receptors read combinations of ligands and a population of cells responds to a perturbation across thousands of genes? One line of our work asks this question using non-equilibrium network models. Response constraints on the steady states of these networks enable local, imperfect feedback control; the same constraints make each output monotone in each input, which limits the decision boundaries a network can draw. We are now asking which structural features of signaling networks let them draw complex decision boundaries. With the Goyal lab, we showed that the gene–gene covariance of unperturbed cells predicts transcriptome-wide responses to genetic and chemical perturbations (CIPHER).


In C. elegans, neuropeptides act far beyond the synaptic connectome and are needed to sustain behavioral states. We ask what neuropeptides do for computation in the brain. With Monika Scholz, we model their action as a slow, activity-dependent, multiplicative modulation of each neuron. With the synaptic couplings unchanged, this modulation extends memory retrieval far beyond the classical Hopfield capacity and produces emergent multistability.


We study how cytoskeletal networks, the structural scaffolding of cells, learn and adapt through mechanical feedback and active processes. Our work shows how activity enables transitions between contractile and extensile states, and how mechanosensitivity enables a form of physical learning.
People: Lissa Rennert and Billie Meadowcroft
We ask the same questions of synthetic systems. How can physical systems store, retrieve and process information through their material properties? Using ideas from Hopfield networks and associative memory, we study how non-equilibrium dynamics can enhance memory in elastic and disordered systems. We also ask whether chemical reaction networks can compute, and whether quantum coherence is a resource for computation in molecular networks. This builds on our work on self-assembly and soft matter far from equilibrium.
Also working on this: Billie Meadowcroft and Lissa Rennert

We explore connections between modern AI architectures and non-equilibrium physics. Score-based generative diffusion can be enhanced by active noise sources, chemical reaction networks can perform in-context learning, and ideas from statistical mechanics illuminate the design of machine learning algorithms.
