Anupam Ojha is a postdoctoral researcher at the Flatiron Institute. He develops advanced statistical frameworks to resolve the dynamic, hidden structural conformations of vital proteins. In this Postdoc Portrait interview, he shares how he came to the field of computational biophysics and what he hopes his work will achieve.
Advancing Structural Biology Through Statistical Mechanics
Q | What drew you to computational biophysics?
I grew up far from any research laboratory, in a remote part of north-western India, and came to science largely through textbooks and a stubborn interest in physical problems. What pulled me into computational biophysics, and has kept me there, is the unreasonable success of statistical mechanics in describing living systems. The same mathematical machinery that explains why a gas occupies a volume also predicts, in principle, why an enzyme catalyzes a reaction or why a receptor binds one ligand and not another. During my graduate work on drug-target kinetics, I became convinced that the bottleneck in modern biophysics is no longer measurement or computation alone, but the lack of a rigorous statistical layer that connects the two. That gap, between what experiments observe and what simulations predict, is the problem I have spent the rest of my career trying to formalize.
Q | What scientific problem are you trying to solve?
Proteins are not fixed structures. They exist as populations of conformations whose relative weights determine function. Single-particle cryo-electron microscopy (cryo-EM), a technique that reconstructs 3D structures from hundreds of thousands of 2D images of vitrified samples, is now powerful enough to detect this heterogeneity, but it still returns averaged maps that miss the rare and functionally critical states. Molecular dynamics simulations, in contrast, generate detailed trajectories of motion but routinely sample states inconsistent with experiment due to inaccuracies in the force field and limited time scales. My work develops statistical frameworks, built on Bayesian inference, maximum-entropy reweighting, and manifold learning, that integrate these two sources of information into a single, internally consistent conformational ensemble. I implement this approach in systems where the dynamics are the function itself.
Q | What’s one thing you learned from your research that you didn’t expect?
The most unexpected lesson came from carefully examining the particles discarded during cryo-EM reconstruction. Standard practice is to identify a dominant 3D class, refine it to high resolution, and remove particles that do not align well, typically about a quarter of the dataset. I had assumed this discard pile was mostly damaged or misclassified images. While working on ManifoldEM, which embeds the entire particle stack into a continuous low-dimensional manifold rather than sorting it into discrete classes, I observed that a substantial fraction of the rejected particles are genuine conformational states that fall outside the dominant class. Some correspond to intermediates of functional motions that cannot be observed directly by any other method. The lesson is that resolution and completeness are not the same thing. Optimizing for the sharpest map can systematically erase the very dynamics one is trying to study.
The Next Frontier in Machine Learning and Protein Modeling
Q | If your research succeeds, what are the long-term implications?
A general framework that integrates simulations with experimental observables would change the economics of several problems we currently treat as difficult. Predicting whether a drug candidate will have a useful residence time, currently a months-long campaign of simulation and biochemistry could become a routine calculation with quantified uncertainty. Designing biologics whose conformational populations are tuned for stability or switching could be possible with limited sequence and experimental input. For systems where the mechanism is still contested, such as viral entry or chaperone cycling, experimentally-consistent ensembles would allow us to test hypotheses about intermediates that no single technique can resolve on its own.
Q | What question are you most excited to answer next?
I want to know whether protein structure prediction models can be retrained end-to-end to output conformational ensembles with experimentally correct populations rather than single structures. Public databases now contain sufficient cryo-EM, nuclear magnetic resonance, and hydrogen-deuterium exchange mass spectrometry data for thousands of proteins across many states, so that the supervisory signal is present. What is missing is the right loss function, together with a statistical framework that connects predicted ensembles to heterogeneous experimental constraints. If this works, sequence-to-ensemble prediction becomes a usable tool for the rest of biology, and not only for specialists who can afford weeks of simulation.
Responses have been edited for length and clarity.
Are you a researcher who would like to be featured in the “Postdoc Portraits” series? Send in your application here.
