top of page
13 | Turning diffusion models into interpretable Bayesian inference machines.
Keeping score, we use nonequilibrium path thermodynamics to co-opt diffusion models for classification and regression.
Our approach calculates the class log-likelihood p(x|c) exactly for each input feature, making us able to say exactly how much each degree of freedom contributes to (un)successful statistical inference. The framework can be applied to high-dimensional data across the physical sciences.
B. Kuznets-Speck, [...], & Y. Goyal, Interpretable thermodynamic score-based classification of relaxation excursions, bioRxiv, (2025).
12 | Predicting perturbation responses from gene co-fluctuations with statistical physics.
We apply linear response from statistical physics to fit transcriptional-wide changes in gene expression upon CRISPR perturbation, and infer driver perturbations from genes fluctuate together in control (nonperturbed) conditions.
These baseline fluctuations are, to an extent, stereotyped in primary tissues, revealing universal gene modules that modulate response.
Response among hundreds and thousands of genes lie along just a few 'soft' covariance directions/modes characterized by basic homeostatic processes.
B. Kuznets-Speck, [...], & Y. Goyal, Fluctuation structure predicts genome-wide perturbation outcomes, bioRxiv, (2025).
11 | Generative prediction of causal gene sets responsible for complex traits
Complex trait phenotypes from asthma to cancer metastasis are determined by an orchestra of gene interactions. We took a data driven path, combining gene perturbations with complex disease RNAseq studies to learn which effective perturbations drive phenotypic change. To probe the sets of genes that work together to cause these diseases, we melded generative AI with statistical physics/network science and constrained optimization in a framework we coin a transcription-wide variational autoencoder (TWAVE).
B. Kuznets-Speck, B.K. Ogonor, T.P. Wytock, & A.E. Motter, Generative prediction of causal gene sets responsible for complex traits, Proc. Natl. Acad. Sci. U.S.A. 122 (24) e2415071122, https://doi.org/10.1073/pnas.2415071122 (2025).
10 | Experimental and theoretical tests of the foldon theory of protein folding
In the Bustamante lab, we use single molecule optical tweezing experiments and molecular dynamics simulations at multiple scales to test the foldon theory of protein folding. By pulling proteins apart and letting them refold, we learn that they leverage secondary structure to fold and unfold along only a handful of unique paths in a cascade like manner, leveraging one foldon as a template for the next to fold, and so on.
We use adaptive importance sampling to learn approximately counter-diabatic forces that make work measurements reversible even for short finite time protocols. With these forces, we estimate free energy differences of complex systems with significantly lower variance and fewer trajectories than by naive switching protocols.
A. Zhong* & B. Kuznets-Speck*, M.R. DeWeese, arXiv:2304.12287
08 | Inferring equilibrium rates with nonequilibrium protocols
We use statistics of the distribution of heat dissipation, conditioned on making a rare transition under applied force to estimate the rate of the transition when the system is unperturbed, at equilibrium.
07 | Variational path sampling (VPS) to control and learn the rates of rare events in noisy complex systems
VPS makes YOU the (puppet) master of your [nonequilibrium] potential
- Checkout our new theory and practical simulation framework to estimate the rate/mechanism of rare transitions in nonequilibrium many-body systems!
- We circumvent the problem of sampling rare events by making them typical-- learning and applying an optimal control force which mimics the force naturally felt along transition paths.
06 | Counter-diabatic control of biophysical processes
We developed a graph-theoretic formalism to control arbitrary biophysical systems out of equilibrium, and put concrete limits on what can be done to shape their function with partial control when some microscopic details of the underlying system remain unknown.
05 | Dissipation bounds transition rate amplification far from equilibrium ​
We uncover a universal tradeoff between speed and energetic resource consumption: the excess heat dissipated along a transition path sets a limit on the extent a transition rate can be accelerated
04 | Molding the (mal)function of von Willebrand factor​
​With coarse-grained MD simulations like the one in this clip, I'm studying how the protein that clots your blood and regenerates wounded tissue behaves and misbehaves in its nonequilibrium environment.
​
​
Here's a sneak peek from our manuscript exploring the trade-offs between speed, fidelity and energy consumption and selective pressure in Kinase-Phosphatase push-pull loops.
​​
​
03 | The price of a bit: energetic costs and the evolution of cellular signaling
​​
We recently published this awesome work, where we design non-equilibrium selection protocols to robustly steer the evolution of high-dimensional populations.
02 | Controlling the speed and trajectory of evolution in high-dimensional clonal populations
01 | Thermodynamic Properties of Molecular Communication ​
Here we take a duel view of Maxwell's Demon, showing that the Landauer bound falls out as the information capacity per unit energy cost to faithfully operate a communication channel comprised of low/high concentration reservoirs.
bottom of page