Connecting ensembles to function

Accurately predicting protein function remains a central challenge in biological machine learning. Function is multifaceted. A single protein can catalyze a reaction, bind a ligand, and transmit an allosteric signal. Each of these depends on context: expression level, pH, macromolecular crowding, and other perturbations. We are building efforts to connect machine learning models, including large language models, protein language models, and ensemble predictors to identify where ensemble-level information improves functional predictions.

We are building the shared tools, metrics, and methods to test how dynamics drive function, serving both mechanism and design.

Molecular interactions sustain life. Our structural models often reduce them to static interfaces. That view overemphasizes enthalpy and neglects the entropy arising from macromolecular motion and solvent, leaving gaps in our mechanistic and thermodynamic accounts of binding. Entropy helps drive binding by reshaping the number of accessible protein and solvent states. Prism is building the biophysical understanding needed to apply that insight to small molecule therapeutic design.

Closing this gap requires developing metrics to measure and predict entropy from structural data and connecting this to prospective design campaigns.