What is the latest AI research on protein folding?

Protein folding research is experiencing one of its most productive periods in scientific history, with AI models now routinely predicting structures that would have taken years of crystallography to resolve. The field has moved well beyond AlphaFold2's landmark achievement, entering a new phase focused on dynamics, interactions, and design — predicting not just what proteins look like, but how they move, bind, and can be engineered from scratch.

The most significant ongoing development is the maturation of AlphaFold3, released by Google DeepMind, which extended structure prediction to protein-DNA, protein-RNA, and protein-ligand complexes — a critical leap for drug discovery. Researchers at institutions like the Wellcome Sanger Institute have been applying these capabilities to map previously uncharacterized proteins across the human proteome, with particular focus on intrinsically disordered regions that older models struggled to handle. Meanwhile, Meta AI's ESMFold continues to power large-scale evolutionary analyses, having already predicted structures for over 600 million proteins from metagenomic databases, giving researchers unprecedented access to nature's structural diversity.

A compelling emerging trend is the shift toward generative protein design. Labs including those at the University of Washington's Institute for Protein Design are combining diffusion-based models like RFdiffusion with structure predictors to close the loop between design and validation. This approach is yielding novel enzymes and therapeutic binders that don't exist in nature, with some candidates already entering early experimental validation pipelines. The integration of molecular dynamics simulations with AI predictions is also gaining traction, addressing a long-standing criticism that static snapshots miss the conformational flexibility essential for biological function.

Watch closely for developments in RNA structure prediction, which is rapidly catching up to the protein domain following the success of models like RhoFold+ and Evo. The intersection of protein and RNA folding — particularly for ribosomes, spliceosomes, and CRISPR machinery — represents the next major frontier. Additionally, expect benchmark competitions and open datasets to sharpen community focus on antibody-antigen interfaces, where current models still show meaningful accuracy gaps. The next 12 months will likely see the first AI-designed proteins advance into human clinical trials, marking a true inflection point for the field.

— Pulse