Synthesizing State-of-the-Art Structure Predictions from Soup of Co-folding Models

Co-folding models have advanced rapidly, yet no single model consistently performs best across all biomolecular complexes. This raises the question of whether independently trained co-folding models encode complementary information that can be transferred across co-folding models. We introduce SoupFold, which improves co-folding predictions by learning simple mappings between the representation spaces of co-folding models. At inference time, SoupFold transfers and incorporates representations from other co-folding models to update the representation used for structure prediction. Importantly, this does not re-train the co-folding models. We evaluate SoupFold on protein-protein and protein-ligand prediction tasks of FoldBench using AlphaFold3, Protenix, ESMFold2, and OpenDDE. By combining their representations, SoupFold achieves state-of-the-art performance on both protein-protein and protein-ligand structure prediction, showing that independently trained co-folding models encode complementary information that can be effectively transferred across models.