Protein Language Model Supervised Scalable Approach for Diverse and Designable Protein Motif-Scaffolding with GPDL

Abstract

Proteins perform essential roles in numerous biological processes, largely driven by the three-dimensional structure of several key motif residues. Recently, a variety of energy-based and machine learning backbone generation methods have been developed to solve the motif-scaffolding task. However, it is still challenging to generate diverse and accurate scaffold structures around motifs for models either fine-tuned pre-trained multiple sequence alignment-based (MSA-based) structure prediction models or trained from scratch. Here, we introduced Generative Protein Design by Language model (GPDL) for effectively replacing traditional MSA-based pretraining. Using our scalable design strategy, GPDL successfully solved 22 out of 24 benchmark problems and outperformed other methods by generating 33.5% more unique designable clusters than RFdiffusion. This demonstrates that our approach can generate accurate and physically plausible structures across diverse protein design scenarios. GPDL also showed strong robustness in orphan proteins that have low sequence similarity with the training set. Our approach underscores the promise of protein language models in protein design and has the potential to accelerate the discovery of novel functional proteins for a wide range of biological and therapeutic applications.