Sequence Scanning Improves Accuracy for Antibody-Antigen and Molecular Glue Ternary Complex Cofolding

Accurately modeling biomolecular interactions is crucial for understanding biological function and can accelerate drug discovery efforts. Deep-learning-based cofolding methods such as AlphaFold3 and Boltz-2 can successfully predict many different biomolecular complexes. However, these models show lower success rates in predicting antibody-antigen and molecular glue (MG) ternary complexes, two cases of high relevance to pharmaceutical research. Here, we present sequence scanning, a new workflow, evaluated on Boltz and Protenix, that improves prediction accuracy in these domains by applying systematic constraints to enhance overall sampling diversity. For both use cases, we assess the performance of the method on large, diverse benchmark sets. We find improved accuracy of the sequence scanning workflow in predicting antibody-antigen interface contacts and in pose prediction quality for MGs compared to previously published results. We also discuss remaining challenges and potential improvements to the method.