The Unreasonable Redundancy of Protein Folds in Nature

With advances in generative models like AlphaFold3, the field of biomolecular modeling—including drug and antibody design—is growing rapidly. A key methodology for improving these models is scaling model, compute, and data.

In particular, the approach of converting vast amounts of protein sequence data into 3D structural data via structure prediction models, and using this as training data for next-generation models, is gaining attention. This is a strategy that increases the value of data by converting sequence scale into structural scale.

However, research has found a limitation: although natural protein sequences are extremely vast, the protein folds they form are highly redundant relative to the number of sequences. Simply folding more sequences does not sufficiently secure the structural diversity needed for model training.

As a result, this mismatch between sequence diversity and structural diversity is a key data engineering challenge that must be addressed when building advanced generative models, such as those for enzyme design.