Advancing Structure Prediction of Biomolecular Interaction via Contact-Guided Sampling with HelixFold-S1

Abstract: Biomolecular structure prediction is essential to molecular biology, yet accurately predicting the structures of complexes remains challenging, especially when co-evolutionary signals are absent. While recent methods have improved prediction accuracy through extensive sampling, aimless sampling often provides diminishing returns due to limited conformational diversity. Here, we introduce HelixFold-S1, a contact-guided sampling strategy that improves structural accuracy. Rather than relying on indiscriminate sampling, HelixFold-S1 predicts contact probabilities between molecular entities and uses these predictions to prioritize sampling of likely binding sites and modes. This targeted approach generates a diverse set of structural candidates, enhancing the likelihood of identifying accurate conformations. We demonstrate that HelixFold-S1 consistently outperforms baseline sampling strategies across a range of biomolecular interactions, including protein-antibody, protein-protein, protein-ligand, protein-RNA, and protein-DNA interfaces. Furthermore, the predicted contact probabilities serve as a reliable indicator of structural difficulty, guiding the allocation of sampling resources. These results highlight the potential of targeted sampling strategies to advance the structural modeling of complex biomolecular interactions during inference.