Two-Stage Ligand-Protein Complex Sequence Design with L-Caliby

Jinho Kim, Richard W. Shuai, Po-Ssu Huang

Keywords: Protein Design, Ligand-Binding Protein Design, Sequence Design, Ligand Conditioning, Potts Model, Generalization

TL;DR: L-Caliby, a Potts model with ligand conditioning and a two-stage pocket-then-scaffold design paradigm, outperforms LigandMPNN on both native and de novo ligand-binding protein benchmarks evaluated by AlphaFold3.

Abstract:

Designing proteins that bind small-molecule ligands requires sequence design methods that account for ligand interactions while maintaining robust backbone foldability. A recent work, Caliby, has demonstrated that Potts model-based sequence design achieves state-of-the-art AlphaFold2 self-consistency for protein structural design, without features to model protein-ligand interactions. Here, we introduce L-Caliby, a Potts model for ligand-conditioned protein sequence design that augments the Potts-based architecture with LigandMPNN's ligand encoder and protein-ligand interaction module. L-Caliby employs a two-stage pocket-then-scaffold (Pocket→\to
→Scaffold) design paradigm using two independently trained design models: a pocket model that optimizes protein-ligand interactions, and a scaffold model that ensures robust foldability of the remaining sequence. On 61 held-out native ligand-protein complexes evaluated by AlphaFold3 in single-sequence mode, L-Caliby (Pocket→\to
→Scaffold) outperforms LigandMPNN on ligand metrics by more than fourfold. On 1,220 de novo protein-ligand complexes generated by RFdiffusion3 with unseen ligands, L-Caliby similarly outperforms LigandMPNN by 28% on PoseBusters-valid ligand success rate. These results reveal that a two-stage Potts model with ligand conditioning can satisfy the different requirements needed for pocket and scaffold redesign tasks and simultaneously optimize both.

Submission Number: 83