Structural Memorization in AlphaFold: Adversarial Mutations Reveal Template Reliance, Confidence Failures, and Implications for Protein Design

AlphaFold has transformed structural biology and spawned an ecosystem of derivative tools for protein design, binding prediction, and drug discovery. Whether AlphaFold has learned generalizable biophysical principles versus template-based pattern matching remains unclear—a distinction critical for applications beyond its training context. Here, we perform a systematic adversarial evaluation of AlphaFold 3 using point and deletion mutations across 200 proteins, including experimentally validated fold-switching proteins. Predicted structures remain invariant to mutations of up to 40% of residues—including deliberately destabilizing substitutions—and to deletions of 10%, even for fold-switching proteins known to adopt alternative conformations under precisely such perturbations and for which AlphaFold is expected to perform best. Confidence metrics prove unreliable, selecting the most accurate structure at most 35% of the time and correlating with training-set template quality rather than biophysical prediction accuracy—suggesting AlphaFold's uncertainty estimates reflect template availability more than genuine structural reasoning. ESMFold exhibits greater mutational sensitivity. Together, these findings suggest AlphaFold relies heavily on memorized templates rather than biophysical reasoning, with profound implications for the reliability of AlphaFold-based protein design, drug discovery, and modeling workflows.