How do Co-Folding Models Organize Structural Information?

Abstract

Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix-Compress-Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo-holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.