Where to Steer: Latent-Space Optimization vs. Coordinate Guidance for Conformational Ensemble Generation from Electron Density

As protein machine learning moves beyond milestone achievements in structure prediction, attention is turning to predicting the ensembles of conformations. These ensembles are what enable proteins to perform their function. One source of the training data needed to predict these ensembles is experimental data generated by X-ray crystallography and cryo-electron microscopy. However, modeling this data remains a challenge. Diffusion-based structure predictors such as Protenix offer one route forward. As conditional diffusion models, the predictors can be guided to predict ensembles consistent with experimental measurements. However, the best way to guide structure predictors with experimental data remains an open question. Here, we use the sampleworks platform to make a direct comparison of two of the most successful methods, coordinate-space guidance (CSG) and inference-time latent-space optimization (ITO), across the same test set of 40 proteins on simulated, noise-free electron density maps. We find that the two methods are indistinguishable in aggregate across the test set, with each method being slightly better suited to particular proteins. Preliminary evidence suggests ITO is better suited to large conformational changes, with CSG being better suited to small, local rearrangements.