Quantitative Recovery of Biomolecular Structural Ensembles from Cryo-EM

Cornell University, Ithaca NY

Investigators

Abstract

Project Summary Cryogenic-sample electron microscopy (cryo-EM) presents a unique opportunity to understand protein conformational change. In cryo-EM experiments, a solution of biomolecules is snap-frozen, trapping them in conformations close to the ones they adopt in solution. Modulo the effects of freezing on the conformational ensemble, we should be able to quantitatively recover the probabilities of protein conformational states from cryo-EM. In prior work, we developed a framework for addressing this challenge. For a hypothesized conformational ensemble, we simulate a cryo-EM experiment and see how well the resulting images match the experimental data. The better the match, the more likely the conformational ensemble is to be correct. Preliminary results suggest that this approach can recover accurate conformational probabilities from cryo- EM data, even in high noise regimes when it is impossible to accurately classify individual images. We plan to build on these promising preliminary results by developing new algorithms that will aid us in recovering conformational probabilities accurately and at scale. To scale our formalism to larger datasets and more conformations, we first aim to accelerate the comparison of simulated and experimental cryo-EM images using a combination of hierarchical clustering and heuristic search. By leveraging these twin strategies, we can efficiently match experimental to simulated images, minimizing the computational power required by our formalism. We also seek to improve the way we represent our conformational ensemble. Diffusion models, a family of generative neural networks, have seen considerable practical use for protein design. By training our diffusion model to match experimental data, we gain a flexible, fast, and accurate way to encode the ensemble of protein conformations. Moreover, our work will open the door to foundation models for protein conformational ensembles: generic models that can be used to predict conformational changes for all proteins. In parallel with our algorithmic work, we will attempt to push the limits of heterogeneity analysis for cryo-EM by attempting to recover the conformational ensemble of intrinsically disordered regions (IDRs). Historically, this has been seen as an impossible task due to the prohibitive difficulty of determining what IDR conformation a given cryo-EM image represents. But our formalism doesn’t require solving this problem: it is enough to recover the statistical properties of the image set. Consequently, we believe our method can succeed where traditional approaches have failed. Ultimately, the proposed work will transform cryo-EM into a tool that can efficiently recover protein conformational ensembles with quantitative probability estimates. This will shed new light into how proteins move to accomplish their biological tasks, helping scientists understand protein function, facilitate the design of new proteins, and discover new druggable mechanisms.