Guided by the classic notion that a protein's amino acid sequence determines its three-dimensional structure, protein structure prediction has long been a central problem in structural biology. Over the past decade, deep learning methods, exemplified by AlphaFold2 (AF2), have revolutionized the field, an achievement recognized by the 2024 Nobel Prize in Chemistry. Yet this success does not mean the protein folding problem is fully solved. Key challenges, each an opportunity for future breakthroughs, include single-sequence prediction, modeling protein dynamics, predicting multimeric complexes, and incorporating experimental constraints.

Recently, a team from the Yang lab at Shandong University published a perspective article, “Opportunities and challenges in protein structure prediction” in the journal Quantitative Biology. The paper systematically reviews recent advances in four major frontiers of protein structure prediction and presents the authors’ perspectives on future research directions. The article points out that these frontier challenges share an underlying bottleneck: in such scenarios, the deep evolutionary information that underpins the success of AlphaFold-family methods is often sparse, absent or insufficient. A paradigm shift is needed for the field, moving from “prediction based on evolutionary data” toward “modeling constrained by physical principles and experimental data.”

For single-sequence prediction, AF2’s reliance on multiple sequence alignments (MSAs) limits efficiency and fails for “orphan” proteins such as fast-evolving viral proteins and hypervariable antibodies; protein language models (PLMs) offer an MSA-free alternative, as shown by ESMFold, though a truly novel paradigm beyond PLM-based methods remains elusive. For protein dynamics, the article surveys input perturbation through MSA clustering or masking, physics-based sampling, hybrid methods integrating energy features, and generative models such as diffusion and flow matching, while noting the risks of hallucination and the difficulty of selecting biologically meaningful conformations. For multimeric prediction, AlphaFold-Multimer (AFM) raised the interface success rate from about 31% in CASP14 to about 90% in CASP15, and AlphaFold3 brought further gains, yet unknown stoichiometry, the lack of reliable scoring functions, and antibody-antigen complexes remain difficult, with traditional docking excelling on such targets in CASP16. For experimental data, deep learning methods for cryo-EM modeling either fit predicted structures into density maps or build atomic models de novo, and crosslinking mass spectrometry supplies distance restraints (as in AlphaLink), pointing toward multimodal models that integrate one-dimensional sequences, two-dimensional restraints, and three-dimensional density maps.
DOI
10.1002/qub2.70035