2026/06/08 by Karson Chrispens, Marcus Collins, James S. Fraser +3 · 1 voice
Materials Science · Biochemistry, Genetics and Molecular Biology · #Enzyme Structure and Function #Protein Structure and Dynamics #Advanced Electron Microscopy Techniques and Applications
paper · pdf · doi:10.82153/jkxj-tw08
Structural biology increasingly relies on machine learning-based structure predictors, enabling accurate atomic-level structure prediction at scale. However, predicting conformational ensembles rather than single structures remains a fundamental challenge. These predictors are trained on static PDB structures that are themselves imperfect representations of the underlying experiments. X-ray crystallography and cryo-EM measure ensemble-averaged signals over many molecular conformations, yet this heterogeneity is compressed into a single set of coordinates. Recent inference-time guidance methods offer a promising path forward by steering structure predictors toward agreement with experimental data, including ensembles. However, existing approaches are bespoke to specific models, making systematic comparison across predictors, guidance strategies, and loss functions difficult. Here, we present sampleworks_, a modular framework for generating and evaluating biomolecular conformational ensembles from structure predictors guided by experimental data. As an initial benchmark, we asked whether three current structure predictors, Boltz-2, Protenix, and RosettaFold3, can recover experimentally supported alternative conformations when guided by simulated, noise-free electron density maps. We assembled a dataset of 791 segments from 40 high-resolution PDB entries, which have multiple physically plausible and experimentally supported conformations, but which are absent from structure predictor training data. Without guidance, predicted baseline ensembles were strongly biased toward whichever conformation dominates the training set. Density guidance substantially improved the ability to capture multiple states, but still retained biases that reflect the training set conformational distribution. Together, these results show that current structure predictors can be steered toward experimental data but do not yet reliably generalize to physically plausible ensembles. sampleworks_ provides a platform for diagnosing these limitations and developing ensemble-aware structure predictors that better reflect the heterogeneous data measured by structural biology data.