Ligand-Conditioned Discrete Diffusion for Protein Sequence-Structure Co-Design

Researchers introduce ProtLiD², a discrete diffusion model that co-designs protein sequences and structures while conditioning on ligand information, achieving significant improvements in fold confidence and ligand-binding accuracy compared to existing methods. The model demonstrates practical advantages in both whole-protein and active-site pocket design tasks.

ProtLiD² represents a meaningful advancement in computational protein design by addressing a fundamental limitation in existing approaches: the inability to jointly optimize protein sequences, structures, and ligand-binding simultaneously within a discrete token framework. While continuous diffusion models have shown promise in coordinate spaces, discrete diffusion models offer computational advantages and better alignment with how biological sequences naturally exist as discrete units. This research bridges that gap through geometry-aware cross-attention mechanisms that incorporate both chemical and spatial information about ligands.

The technical improvements are substantial. The model achieves a TM-score improvement from 0.672 to 0.802 in whole-protein design and reduces active-site backbone RMSD to 1.97Å compared to 3.40-3.46Å in predecessor methods. Perhaps more importantly, ligand-aware pass rates increased from approximately 15% to nearly 60% under standard docking criteria. Training on over one million ligand-protein complexes provides the scale necessary for robust generalization.

For the synthetic biology and computational chemistry sectors, this represents incremental but meaningful progress toward practical protein engineering applications. The maximum confidence-margin guided ReMask decoding strategy introduces a self-correction mechanism that could improve reliability in real-world applications. However, the real-world impact depends on validation through wet-lab experimentation and integration into existing drug discovery pipelines. The open-source code release signals confidence in reproducibility and may accelerate adoption among researchers. This work primarily advances the AI/machine learning field rather than presenting immediate commercial or market implications, though it strengthens the technical foundation for AI-driven drug discovery platforms.

• →ProtLiD² jointly generates protein sequences and structures with direct ligand conditioning using discrete diffusion, addressing limitations in previous token-based models.
• →The model achieves 19% improvement in TM-score and reduces active-site RMSD by 43% compared to existing methods.
• →Training on 1M+ ligand-protein complexes and introducing self-correction decoding strategies enhance both accuracy and confidence in predictions.
• →Ligand-aware pass rates improved from 6-15% to 23-60% across different docking thresholds, indicating practical feasibility gains.
• →Open-source release enables broader research adoption in computational protein design and AI-driven drug discovery.