Topic analysis

The Unreasonable Redundancy of Nature's Protein Folds

Researchers at Ligo bio have found that while natural protein sequences are vast, their 3D folds are highly redundant, meaning scaling datasets like MGnify does not add as much new structural diversity as expected for training generative AI models used in enzyme design. They developed advanced data engineering techniques—including spectral bisection for protein fragmentation and TM-score-based cluster auditing—to identify around 25,000 distinct structural neighborhoods in natural proteins, and propose weighted sampling strategies to balance fold diversity and natural abundance when training these models.