Jonathan D. Feldman

AlphaInterp analyzes AlphaFold 3's internal mechanisms to show how evolutionary data influences protein structure prediction, revealing dependencies on comparative context and phylogenetic diversity.

The paper integrates three computational methods to predict E. coli protein-protein interactions and functional networks, identifying over 20K high-confidence interactions.

The paper evaluates AlphaFold and ESMFold by introducing adversarial mutations and finds structural invariance and unreliable confidence metrics, suggesting reliance on templates rather than biophysical principles.

This paper introduces a framework that integrates biophysical priors into protein embeddings to improve DeltaDeltaG prediction, showing that general-purpose encoders can outperform specialized models.

AF3Complex improves structural predictions of protein complexes by building on AlphaFold 3 with new modifications, outperforming AlphaFold 3 and achieving top results in CASP16.