Statistical physics of biological organization

We want to know how a cell reads its chemical environment through the receptors on its surface. Receptors are the first interface, and whatever they fail to encode is lost to everything downstream. What sets that ceiling, and how far can a cell push past it by spending energy?

One question we keep returning to is how well an individual cell senses, rather than how well an average one does. Treating a population as a single communication channel puts cells at around one bit, which we take to be an artifact of the averaging. We work on frameworks that recover the distribution of single-cell channel capacities and tie a cell's fidelity to its internal state.

How receptors tell chemically similar ligands apart is a second thread. Multisite phosphorylation and receptor degradation can act together as a band-pass filter over ligand affinity, rather than the high-pass filter proofreading arguments suggest. We are pursuing what that buys a cell, particularly in EGFR signaling, where several old puzzles about low-affinity ligands look different from this angle.

Spatial gradients pose the same problem in a different geometry. The standard local-excitation/global-inhibition picture is under pressure from recent experiments, and we are exploring alternatives in which receptor kinetics alone generate an emergent integral feedback controller, along with the odd possibility that a cell sharpens a gradient by destroying the signal it is trying to detect.

Underneath all of it, we think of receptors as physically learnable machines: chemical reaction networks that encode low-dimensional information about the world. We use ideas from physical learning to reverse engineer the strategies they use.

We want to understand how microbial ecosystems hold on to their diversity. Hundreds of species share a gut or a rumen, competing for an overlapping set of resources, and classical niche theory says that should be hard.

What draws us in is that the coexistence turns out to be low-dimensional. Reading latent dimensions as ecological niche dimensions gives a way to measure niche dimensionality in a real community, and host-associated communities come out surprisingly low-dimensional, with dynamics that reduce to a few ecological normal modes. We want to know what gives rise to that structure and what it implies for ecological theory.

The practical question is whether this can be turned into design. A generative model of microbiomes together with their hosts lets us propose communities matched to a host phenotype, and we are building digital twins we can probe with counterfactual interventions. Three applications are underway: a rumen digital twin for methane mitigation in livestock, gut communities engineered to resist pathogen colonization in hospitalized patients, and community assembly in controlled in vitro systems.

We are drawn to the process of building models in physics. Maximum entropy gives the least-committed distribution consistent with a set of constraints, a clean answer to an awkward question, and it leaves a harder one open: which constraints belong in the model at all?

One extension we work on carries the argument from distributions to trajectories. Maximum caliber does this, and we use it to correct Markov models with dynamical data they were never trained on, to infer transition rates from steady-state populations, and to repair molecular simulations.

The other asks how to choose constraints from the data rather than by hand. Minimax entropy gives a principle for doing so, and it underpins the rest of the lab: latent variables we can read as niche dimensions, generative models of protein families that separate phylogeny from function, and data manifolds reconstructed by treating nearby samples as thermodynamic states.