Site-Dependence as a Design Objective:

Computer models like AlphaFold 3 (AF3) are frequently used to design new peptides (short proteins) by predicting how strongly they will stick to a target. However, a high confidence score from the model only proves it is sure about the final shape. It does not prove the peptide is attaching to the exact spot researchers want, or for the right scientific reason.We looked closely at a recent study that used AF3 to design 25-amino-acid peptides meant to bind to a specific zinc-containing spot on a bacterial enzyme (VIM-2). We found three major flaws in using just the standard confidence score:The measurement is misleading: The model rewarded designs for "burying" the zinc ions, but those ions were already almost completely covered before any peptide was even added.The metal doesn't seem to matter: For all the top-scoring peptides, the AI model was just as confident whether the crucial zinc ions were present or entirely missing.The search picks up "flukes": Relying on a single, random prediction for each sequence means the search easily falls into a trap called the "winner's curse," picking designs that just got a luckily high noise-driven score.To fix this, we created a new strategy called the counterfactual site-dependence objective. Instead of just looking at how well a peptide binds to the normal target, we compare that score against how it performs on "knockout" targets—such as versions without the zinc, mutated versions, or entirely different decoy targets. If a peptide is truly designed for the specific zinc spot, its score should drop significantly when that spot is altered.We prove that this score difference mathematically aligns with the true odds of specific binding, protects against false positives, and cuts down on random prediction noise. While our new method requires running the AI model through a few extra steps—costing about 1.9 times more computing power than the original method—it ensures the AI is actually finding peptides that bind where intended. Finally, we tested and proved our framework using a pre-registered benchmark of known protein-metal complexes.