Department of Chemistry and Biochemistry
Physical Chemistry Seminar Series

Professor Dhiman Ray, University of Oregon
February 2, 2026—2:00pm
Tykeson 140

“Learning Mechanisms from Explainable Artificial Intelligence”

Molecular dynamics (MD) simulations are widely used to study atomistic mechanisms of chemical, material, and biological processes. Although most practical systems exhibit beyond millisecond timescales (e.g., drug-target binding and protein folding), we can only simulate up nanoseconds to microseconds using all atom MD simulation. Enhanced sampling algorithms address this problem by accelerating conformational sampling using an external biasing potential. The accuracy and efficiency of these algorithms are sensitive to the choice of collective variable (CV), a low-dimensional space along which the bias is applied. Deep neural networks can be used to construct CVs using a generic and system-agnostic feature space to compute an accurate free energy surface for complex molecular processes. However, their lack of interpretability and high cost of evaluation during trajectory propagation make NN-CVs difficult to apply to biomolecular processes.
In the past 1.5 years at the University of Oregon, my research group has been addressing this limitation from two different directions. First, we developed a surrogate model approach to express the output of a neural network as a linear combination of a subset of the input descriptors. In addition to providing mechanistic insights due to their explainable nature, the surrogate model CVs exhibits improved computational efficiency and negligible reduction in accuracy compared to the NN-CVs in reconstructing the underlying free energy surface. Second, we introduced a deep-multitask learning algorithm that can identify mechanistic pathways without the need for spending computational effort in sampling and converging the whole free energy landscape. This approach also generates mechanistic fingerprints from which subsequent ML models can be trained to predict dynamics properties of molecules. These explainable AI techniques provide mechanistic understanding of RNA and protein dynamics at an atomistic resolution.
1. Elangovan, R., Chatterjee, S., Ray, D. (2025). Data-driven enhanced sampling of mechanistic pathways. Proceedings of the National Academy of Sciences 122(49), e2517169122.
2. Chatterjee, S., & Ray, D. (2025). Acceleration with interpretability: A surrogate model-based collective variable for enhanced sampling. Journal of Chemical Theory and Computation, 21(4), 1561-1571.
3. Chatterjee, S., & Ray, D. (2026). Characterizing RNA Tetramer Conformational Landscape Using Explainable Machine Learning. The Journal of Physical Chemistry Letters.
4. Chatterjee, S., & Ray, D. (2026). Explainable Machine Learning Guided Enhanced Sampling of Protein Conformational Transition in HSP90. ChemRxiv.