Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling

Details

This review paper from Wengan He and colleagues maps the methodological evolution of protein structure prediction across four phases: early evolutionary coupling features, learned sequence representations (AlphaFold2, RoseTTAFold, ESMFold), heterogeneous molecular modeling (AlphaFold-Multimer, AlphaFold3), and generative design (RFdiffusion). Three cross-cutting transitions structure the narrative: from explicit features to learned representations; from protein-only to heterogeneous molecular systems; and from prediction to generative design. The review is a preprint submitted to Elsevier and has not yet undergone peer review. Its primary value is synthesis — pulling together disparate model families under a unified conceptual lens rather than presenting new experimental results or benchmarks. For researchers entering the space or practitioners evaluating tools like AlphaFold3 or RFdiffusion for drug discovery pipelines, it provides useful historical orientation. No new models, datasets, or benchmarks are introduced. As a survey, it organizes existing knowledge rather than advancing the technical frontier. The field has demonstrably shifted from prediction-oriented inference toward design-oriented generative modeling — RFdiffusion represents the current frontier for practitioners.