The Unreasonable Redundancy of Nature's Protein Folds

Recent advances in deep learning for protein structure prediction, exemplified by DeepMind's AlphaFold3, have enabled the design of improved drugs and biologics by scaling models, compute, and training data. However, researchers at Ligo discovered a critical limitation: while natural protein sequences are vast, their corresponding 3D folds are far more redundant than sequence counts suggest, meaning that simply folding more natural sequences may not provide the structural diversity needed to advance enzyme design. This mismatch between sequence abundance and fold redundancy has important implications for how biomolecular design models should approach future scaling.