University of Chicago, Dinner Group

During a summer research appointment at the University of Chicago, I worked as an AI Research Assistant studying how well AlphaFold's confidence metrics predict real protein-ligand binding behavior.

To predict how a given ligand would bind to a target protein, I fused the ligand and protein sequences using a glycine linker, allowing AlphaFold to fold the pair as a single continuous chain and generate a predicted bound structure. I ran this process across more than 200 protein-ligand pairs, extracting each prediction's pLDDT, pTM, and PAE scores, then compared those confidence metrics against experimentally measured relative enrichment scores to look for correlations. Using Python (NumPy, Pandas), I processed and analyzed the resulting datasets to identify which confidence metrics actually tracked with real binding outcomes.

The work fed into a broader question the group cares about: how much can you trust an AI-generated structural prediction without experimental validation. I presented my findings, the trends and limitations in the dataset, to the research group. In the end, there was not enough of a coralation for it to be an accurate measure.