Can AI predict the 3d structure of any protein from its amino acid sequence ?

For decades, deducing a protein’s three-dimensional shape directly from its linear amino-acid sequence was considered one of biology’s hardest problems. Today AI systems can do so with remarkable precision—but how complete is that capability across the full universe of proteins?

Background

AlphaFold 2, developed by DeepMind and unveiled at CASP14, demonstrated near-experimental accuracy in blind structure-prediction trials and now underpins the majority of structural biology workflows (Nature enrichment, May 9, 2026).

Current AI methods—exemplified by AlphaFold—leverage deep learning architectures trained on large curated libraries of experimentally solved protein structures. These models learn statistical correlations between sequence and conformation, enabling end-to-end prediction of 3D coordinates from primary amino-acid strings. In benchmark assessments, AlphaFold’s median accuracy approaches that of low-resolution experimental techniques for many globular proteins (Senior et al., Nature 2020; Jumper et al., Nature 2021).

Despite rapid advances, open challenges persist. Accuracy remains lower for proteins with non-canonical folds, large intrinsic disorder, or sparse evolutionary signal. Community-wide assessments such as CASP continue to track progress and highlight edge cases where human insight or additional experimental data are still required. Ongoing research targets improved robustness, uncertainty quantification, and generalization to orphan sequences and membrane proteins (Nature enrichment, May 9, 2026).