Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling

Protein structure prediction has undergone a profound transformation driven by deep learning, shifting from early multiple sequence alignment (MSA)-driven monomer folding to broader frameworks capable of modeling complex macromolecular systems and generating novel designs. Existing reviews have frequently examined this progress from the perspective of representative model families, application domains, or protein design goals. In contrast, this review focuses on the methodological evolution of the field itself, analyzing developments through three interconnected dimensions: representations and data, architectures and learning strategies, and confidence and evaluation.

The evolution of the field can be structured into four sequential methodological phases accompanied by three cross-cutting transitions. The first phase relies on explicit evolutionary modeling, where homologous sequences are used to compute covariance statistics and direct coupling analysis to predict residue-to-residue contacts. The second phase introduces learned sequence representations through pretrained protein language models (PLMs) and end-to-end architectures like AlphaFold2, RoseTTAFold, and ESMFold, which replace handcrafted evolutionary features with internal latent representations. The third phase expands the modeling scope from isolated monomer folding to the integrated treatment of heterogeneous macromolecular assemblies—such as protein-protein complexes, nucleic acids, and small-molecule ligands—in models like AlphaFold-Multimer, RoseTTAFoldNA, and AlphaFold3. Finally, the fourth phase transitions from prediction-oriented structure inference to design-oriented generative modeling using diffusion and flow-based frameworks.

Underlying these phases are fundamental changes in how data and architectures interact. Early pipelines were strictly modular and fragmented, separating feature extraction, contact prediction, and downstream structure reconstruction, which allowed errors to propagate unchecked. Modern architectures employ end-to-end differentiable learning strategies, jointly optimizing internal representations and atomic coordinate predictions. Furthermore, the data regime has expanded from curated structural supervision combined with deep homologous sequence alignments toward massive unlabeled sequence corpora and diverse covalent and non-covalent biomolecular interaction data.