AI's Protein Predictions Rely on Hidden Shape Clues
New research reveals how AI predicts protein structures, impacting drug discovery.
A research report asks how ESM-2, a model designed to infer missing amino acids in protein sequences, represents and uses DSSP-labeled secondary structure—helix, strand, or coil—in ESM-2 150M. Linear probes could decode secondary structure, but ablating that linear plane reduced true residue likelihood by about 1% for helix and strand compared with random vector ablation.
Ablating the class-mean plane at layer 29 was more effective: a refit probe fell to 0.22 macro-F1, and true residue likelihood dropped 6% across secondary structures. Deliberately adding the helix class-mean vector to coil rows shifted residue predictions toward helix-forming amino acids, increasing predicted probability on coil rows by 13%. The report says the linear representation of secondary structure is used by ESM-2 but is not a complete picture of what the model represents about secondary structure.
- ESM-2, an AI that predicts protein shapes, uses hidden patterns of secondary structure to make predictions.
- Removing these patterns reduced the AI's accuracy by about 6%, showing they are important.
- Understanding AI's inner workings can lead to more reliable tools for drug discovery and medicine.
Why It Matters
This research helps make AI in biology more transparent, potentially leading to better drugs and treatments.