Unobserved Sequence Space Has Many Functional Proteins

Recent studies reveal that the vast majority of functional proteins exist in regions of sequence space not observed in nature. Understanding this distribution is crucial for comprehending protein evolution and enhancing protein design. Despite the recent surge in AI/ML protein design tools, these technologies remain unproven. This investigation reveals that many functional proteins reside in unobserved sequence space, which cannot be accurately predicted using computational methods.

Researchers measured the fitness of diverse proteins across three protein families, constructing the largest dataset of natural protein orthologs with diverse functions. For each family, they identified numerous new-to-nature functional sequences with low similarity to existing orthologs. While sequence-based scoring metrics, such as Potts models and protein language models, provide reasonably accurate but inconsistent function predictions, empirical protein fitness landscapes prove to be complex.

Predictions of function often fail to capture the local shape and global trends of these landscapes. Most notably, researchers discovered extensive functional sequence space between existing proteins in each family, supporting the hypothesis that natural protein sequences represent only a tiny fraction of all possible functional sequences.