Spring 2025 Update: AI Is Cracking “Inscrutable” Proteins — Here’s What People Still Do Best

This spring brought a wave of encouraging coverage: AI tools are getting sharper at detecting and describing previously undiscovered proteins and their features. That’s great news for anyone who’s stared at a low-confidence model and thought, “So what?” We’re moving from “folds exist” to “features matter.”

But there’s a gap between pattern and purpose. On one side are interaction-savvy predictors like AlphaFold 3, which model complexes with proteins, nucleic acids, ions, and small molecules. On the other are real-world labels — the pocket boundaries, catalytic residues, and surface chemistry that make a prediction useful to an experimentalist. Our community sits in that gap. We take proposed sites and stress-test them: is the geometry compatible with a ligand family? Is a cavity hydrophobic in the right places? Does the residue micro-environment agree with the reaction chemistry?

Practically, that means feeding human-curated positives/negatives back into training and ranking. It also means triaging long-tail proteins where homology is thin and confidence is lumpy. The result is simple: more testable hypotheses per week. AI is widening the searchlight; people are telling it where to linger.