Author: Emma Wilmott

From Prediction to Scientific Inference: Grounding Generative Models in Biological Experiments

Biological measurements and predictive models are both scaling rapidly, yet the interface between them remains a bottleneck. In structural biology, deposited structures have enabled major advances in AI but capture only part of the underlying experimental evidence. Connecting models more directly to measurements could better exploit information about conformational populations, experimental context, and uncertainty.

This talk presents experiment-grounded generative inference: combining learned priors with measurements to infer sample-specific biological states. I will develop these ideas primarily in structural biology, where the states of interest are biomolecular conformations. I will focus on ROCKET, which uses AlphaFold as a learned prior for inference from X-ray crystallography and cryo-EM measurements, and EmbedOpt, which steers conditional protein diffusion models by optimizing their conditioning embeddings at inference time. Both approaches use pretrained models as structured search spaces, allowing experimental evidence to guide inference beyond a model’s default prediction. I will also briefly discuss GOTO, an optimal-transport-based objective that provides informative optimization signals for joint inference of molecular states and cryo-EM imaging parameters.

I will close with an outlook on MIMIC and possible routes for making measurements first-class inputs and training signals in multimodal biomolecular models. The broader goal is to use learned priors to interpret experimental evidence—and ultimately use that evidence to improve the priors themselves.

References:
Golkar, Li, et al. — MIMIC (arXiv, 2026)
Fadini, Li et al. — AlphaFold as a prior (Nature Methods, 2026)
Li et al. — EmbedOpt (arXiv, 2026)
Woollard, Li et al. — GOTO (bioRxiv, 2025)

Structural Ensembles as Molecular Phenotypes of Translation

Structural studies of ribosomes by single particle cryogenic electron microscopy (cryoEM) have traditionally relied on purified or reconstituted samples, with particles often trapped in desired states using genetic, pharmacological, or biochemical perturbations. While informative, such approaches can fail to capture the full diversity of structural states and associated factors present in cells. In this talk, I will describe our development of cryoDRGN and cryoPRISM, a rapid ex vivo workflow for high-resolution analysis of ribosomes directly from cell lysates, and show how this approach enables visualization and quantitative analysis of ribosomal ensembles spanning assembly, translation, and hibernation under near-native conditions. I will focus on the idea that these structural ensembles can be interpreted as molecular phenotypes, and that perturbations such as small molecules, protein factors, or environmental stresses reshape the structural landscape by shifting the distribution of states. Importantly, I will discuss how such molecular phenotypes can reveal the mechanisms by which these perturbations modulate translation. Examples will be drawn from our recent work in bacterial and archaeal systems, illustrating how we are using cryoPRISM to understand how translation is modulated across diverse biological contexts.

Is Cryo-EM as Good as It Gets? Where We Stand, What We’re Missing, and Why It Matters.

Single-particle cryo-electron microscopy (cryo-EM) has become a transformative technique for determining the three-dimensional structures of biomolecules, building on nearly a century of innovation since the invention of the electron microscope in the early 1930s. Over the past decade, advances in microscope hardware, computational power, and image-processing algorithms have enabled atomic-level structure determination of proteins and large macromolecular complexes. During this time, cryo-EM has resolved many challenging structural biology targets that previously eluded X-ray crystallography.

The rapid expansion of high-end microscopes, powerful computing resources, and user-friendly software has propelled cryo-EM to even rival X-ray crystallography in the annual number of structure depositions. Yet, despite this remarkable progress, significant technical bottlenecks and limitations remain. At a time when cryo-EM seems to have reached maturity as a structural biology method, it is crucial to take a closer look at its current constraints and identify where future improvements are most needed.

This presentation will examine these challenges and explore strategies to further enhance the resolution, quality, and throughput of cryo-EM

Deconstructing Algorithms for “Discrete” Heterogeneity in Cryo-EM

Classic approaches to analyzing conformational or compositional heterogeneity in cryo-EM model attempt to recover a discrete set of structures from the data.  But despite how common these algorithms are, exactly how these structures relate to the structures the molecule observes in solution is not fully understood. In this talk, we revisit the likelihood-based algorithm for finding discrete classes.  By considering the limit of a large number of structures, we show that this algorithm can be viewed as an approach to solving a stochastic inverse problem for the latent structural probability density.  This gives us a new understanding of how the algorithm behaves in the limit of large data.  Moreover, our analysis gives a systematic way to construct new algorithms for recovering probability densities associated withheterogeneous data.  As an initial contribution, we introduce an algorithm based on the Wasserstein gradient flow.  Our work points the way towards more efficient and noise-robust algorithms for recovering heterogeneity with quantitative accuracy.

Diego Sanchez Espinosa, Erik Thiede, Yunan Yang.

The GoldX Fiducial Eraser

Gold nanoparticles are standard fiducial markers in cryogenic electron tomography but often cause reconstruction artifacts due to their high electron density. We present a post-processing tool that automatically detects and suppresses these particles by replacing them with the local background. The method yields cleaner tomograms without loss of sample information or introduction of new artifacts, enabling more accurate visualization and interpretation of native cellular structures. 

References:
Van Blerkom P, Bezault A, Sauvanet C, Hanein D, Volkmann N. The GoldX Fiducial Eraser. International Journal of Molecular Sciences. 2024; 25(13):7442. https://doi.org/10.3390/ijms25137442

Protein structural ensembles from 3D and 2D cryo-EM data

Understanding the molecular mechanisms employed by biological systems to carry out their functions is often essential for rationally targeting associated diseases. In many cases, determining the three- dimensional (3D) structure of these systems provides valuable insights. However, it is frequently the interplay between structural and dynamical properties that determines the behavior of complex systems. While both experimental and computational methods are invaluable tools for studying protein structure and dynamics, limitations in each individual technique can hinder their capabilities [1]. Here, I will present two integrative computational-experimental approaches to combine cryo-electron microscopy (cryo-EM) data (and more) into molecular dynamics (MD) simulations to determine accurate protein structural ensembles [2,3]. I will showcase the capabilities of these methods using different applications to biological systems of outstanding interest. First, I will illustrate how accurate protein structural ensembles can be obtained from 3D cryo-electron microscopy maps with our recently developed EMMIVox approach [4]. Then, I will focus on characterizing the structural and dynamic properties of the CyaA toxin by combining coarse-grained MD simulations with Hydrogen/Deuterium eXchange Mass Spectrometry (HDX-MS), Small Angle X-ray Scattering (SAXS), and 2D single-particle cryo-EM data. Finally, I will show how AlphaFold2 and single-particle cryo-EM images can be synergistically used to characterize the alternative, functional states of a G-Protein Coupled Receptor.

References:
1] M. Bonomi, G. T. Heller, C. Camilloni, M. Vendruscolo. Curr. Opin. Struct. Biol. 42 (2017) 106
[2] M. Bonomi, C. Camilloni, A. Cavalli, M. Vendruscolo. Sci. Adv. 2 (2016) e1501177
[3] P. Cossio, G. Hummer. J. Struct. Biol. 184 (2013) 427
[4] S. E. Hoff, F. E. Thomasen, K. Lindorff-Larsen, M. Bonomi. bioRxiv (2023) 2023.10.18.562710

The Inaugural Flatiron Institute Cryo-EM Conformational Heterogeneity Challenge

Despite the rise of single particle cryo-EM as a premier method for resolving macromolecular structures at atomic resolution, methods to address molecular heterogeneity in vitrified samples have yet to reach maturity. With an increasing number of new methods to analyze the multitude of heterogeneous states captured in single particle images, a systemic approach to benchmarks and metrics in this field is needed. With this motivation, we issued a challenge to the community to analyze two curated cryo-EM image stacks of the thyroglobulin molecule with conformational heterogeneity: the first was a typical experimental dataset, and the second was synthetically generated, allowing control over the distribution of molecular structures in the particle images. This synthetic dataset also enabled direct comparison between participants’ submissions and the ground truth molecular structures and distributions. Participants were asked to submit 80 molecular volumes representing the heterogeneous ensemble in the dataset and estimate their respective populations in the particle stack provided. Participation in the challenge was strong, with submissions from nearly all developers of heterogeneity methods, resulting in 40 submissions across both datasets. Submissions qualitatively exceeded expectations, with the molecular motions identified by methods resembling both each other and the ground truth distribution. However, quantitatively assessing these similarities was a challenge in and of itself.  In the process of assessing the submissions to this challenge, we developed several validation metrics, most of which require reference to the underlying ground truth volumes, but one which does not. These approaches allowed us to assess the similarity and accuracy in map quality, molecular motions, and distribution estimates of different submissions. These metrics and the efforts of all participants will help chart a path forward for the improvements of heterogeneity methods for cryo-EM and future challenges to test these new methods as they continue to be developed by the community.

From Pixels to Bayesian Posteriors: Fast Molecular Conformation Estimation from Cryo-EM with simulation-based inference

Biomolecules are highly dynamic systems. They reorganize between a network of metastable states connected by rare structural intermediates, which is referred to as their conformational ensemble. The ensemble, including the rare intermediate structures, determines biomolecular function in the cell. However, mapping biomolecular conformational ensembles is still an outstanding challenge in both experimental and computational approaches. Cryo-electron microscopy (cryoEM) has emerged as a powerful paradigm for characterizing protein conformational ensembles. However, even though the frozen sample contains information on the entire ensemble, current approaches reconstruct only a few conformations by averaging over many microscopy images. In fact, accurately identifying molecular conformations depicted in a single cryoEM image is still a challenging task. Here, we integrate simulations and probabilistic deep learning to develop the cryoEM simulation-based inference (cryoSBI) framework for inferring molecular conformations and their uncertainties from individual cryoEM images. Given an observed image, cryoSBI enables us to directly estimate the Bayesian posterior using forward model simulations, an embedding network, and a neural posterior estimation framework. CryoSBI is amortized. Training happens only once, after which inference for each experimental image takes only milliseconds to evaluate. Pose and imaging parameters do not have to be estimated, resulting in a high computational speed compared to explicit likelihood methods. For synthetic and experimental data, we could systematically disentangle the molecular conformation from the noisy observation with a confidence interval for the inference and learn about the most relevant features of the observed particles. Our approach paves the way to characterizing entire conformational ensembles from experimental data.

Bio:
Roberto Covino is W3 Professor of Computational Life Science at the Institute of Computer Science at Goethe University Frankfurt, and a Senior Fellow at the Frankfurt Institute for Advanced Studies. His research uses theory, simulation, and statistical modelling of experiments to understand how biomolecular functions emerge from the interplay between structure, dynamics, and complexity. He focuses on understanding the mechanism of key events in proteins and cellular membranes. RC studied physics and theoretical physics at the University of Bologna and obtained his PhD in physics at the University of Trento. He was appointed Professor of AI in Protein Science at Bayreuth University in 2023 and received the call to Goethe University in 2024.

High-resolution in situ structures of mammalian respiratory supercomplexes

Mitochondria are essential for ATP production via oxidative phosphorylation, involving respiratory complexes within the inner membrane. Despite extensive in vitro studies, understanding their mechanisms in physiological environment is challenging due to loss of the native environment during purification. Here, we directly image porcine mitochondria by developing a high-resolution in-situ cryo-electron microscopy technique, which enabled us to determine near-atomic structures of various respiratory supercomplexes in the native membrane.