Biophysical Society Thematic Meeting | Trieste 2024

Emerging Theoretical Approaches to Complement Single-Particle Cryo-EM

Poster Abstracts

4-POS Board 4 AMORTIZED IDENTIFICATION OF BIOMOLECULAR CONFORMATIONS IN CRYO-EM USING SIMULATION-BASED INFERENCE Lars Dingeldein 1,2 ; David Silva-Sánchez 4,6 ; Luke Evans 4 ; Nikolaus Grigorieff 8,9 ; Edoardo D'Imprima 7 ; Roberto Covino 1,3 ; Pilar Cossio 4,5 ; 1 Frankfurt Institute for Advanced Studies, Frankfurt am Main, Germany 2 Goethe University Frankfurt, Institute of Physics, Frankfurt am Main, Germany 3 Goethe University Frankfurt, Institute for Computer Science, Frankfurt am Main, Germany 4 Flatiron Institute, Center for Computational Mathematics, New York, NY, USA 5 Flatiron Institute, Center for Computational Biology, New York, NY, USA 6 Yale University, Department of Mathematics, New Haven, CT, USA 7 Humanitas Research Hospital, Rozzano, Italy 8 University of Massachusetts Chan Medical School, RNA Therapeutics Institute, Worcester, MA, USA 9 Howard Hughes Medical Institute, Chevy Chase, MD, USA Comparing experimental measurements to theoretical models is a fundamental component of scientific investigations. Often, this means comparing observations from simulators that capture the important physics to experimental observations. The aim is to gain a mechanistic understanding by identifying model parameters that can reproduce the experimental data. However, inferring these parameters is often challenging and time-consuming. To tackle these issues, new machine learning-based approaches known as Simulation-Based Inference have been developed, although they have not yet been applied to Cryo-Electron Microscopy (Cryo-EM). Here, we introduce a new method (cryoSBI) using Simulation-based inference to approximate the Bayesian posterior distribution. We utilize Neural Posterior Estimation, a technique that directly approximates the Bayesian posterior using a simulator (e.g., a Forward Model) and a neural density estimator. The key advantage is that the cryoSBI training only happens once with simulated data. Afterwards, inference for each particle takes just one forward pass through the neural network. This provides a significant advantage as the posterior is amortized: the particle pose and imaging parameters do not need to be estimated, resulting in a considerable computational speedup compared to explicit likelihood methods. Through extensive experiments with synthetic and real cryo-EM data, cryoSBI extracts molecular conformations from observations while providing a meaningful measure of confidence in the inferred parameter.The primary advantage of cryoSBI lies in its direct estimation of the posterior distribution, circumventing the computationally intensive pose estimation typically associated with cryo-EM. This leads to a substantial reduction in computational overhead compared to traditional explicit likelihood methods. CryoSBI is especially beneficial in situations that involve large conformational changes and large amounts of experimental data, as is often the case in cryo-EM.

56