AlphaInterp: Probing AlphaFold 3's Internal Representations Reveals Evolutionary Determinants of Predicted Structure and Confidence
AlphaFold 3 predicts biomolecular structures with unprecedented accuracy, yet the computations transforming sequence and evolutionary data into structural coordinates remain poorly understood. Here, we present a systematic mechanistic interpretability analysis of AlphaFold 3, tracking its internal representations across the forward pass. Probing four critical network checkpoints reveals that the Pairformer compresses diffuse co-evolutionary inputs into a compact latent geometry where complex biophysical features become linearly decodable. Using causal activation patching, we demonstrate that predicted confidence is directly manipulable within this latent space, allowing geometric certainty to be transferred across entirely unrelated proteins. Furthermore, across adversarial-mutation, fold-switching, and generalization benchmarks, we show that AlphaFold 3s representational coherence strictly requires comparative evolutionary context. The latent space collapses when multiple sequence alignments are removed, regardless of sequence familiarity or training-set membership. This stability requires phylogenetic diversity rather than alignment depth, and a minimal set of highly divergent homologs is sufficient to anchor the latent space and activate the models structural priors. These findings indicate that AlphaFold 3s representational coherence is deeply tied to evolutionary scaffolding, suggesting it functions similarly to an advanced fold-recognition system and highlighting that protein structure prediction from sequence alone is not yet fully solved.