Protein language model supervised motif-scaffolding design with GPDL.

Area of Science:

• Computational biology
• Protein engineering
• Bioinformatics

Background:

• Proteins' biological functions are dictated by their 3D structures, particularly key motif residues.
• Generating accurate and diverse protein scaffolds around motifs remains a challenge for existing computational methods.
• Traditional multiple sequence alignment-based (MSA-based) pretraining methods have limitations in protein structure prediction and design.

Purpose of the Study:

• To introduce Generative Protein Design by Language model (GPDL) as an effective alternative to traditional MSA-based pretraining for protein design.
• To evaluate GPDL's performance in generating diverse and accurate protein scaffolds across various benchmark problems.
• To assess GPDL's robustness, especially for orphan proteins with low sequence similarity.

Main Methods:

• Developed Generative Protein Design by Language model (GPDL) to replace MSA-based pretraining.
• Employed a scalable design strategy for GPDL.
• Tested GPDL on 24 benchmark protein design problems.

Main Results:

• GPDL successfully addressed 22 out of 24 benchmark problems.
• GPDL generated 33.5% more unique designable clusters compared to RFdiffusion, indicating superior diversity.
• Demonstrated accurate and physically plausible structure generation across diverse protein design scenarios.
• Showcased strong robustness on orphan proteins with limited sequence similarity to the training data.

Conclusions:

• GPDL effectively replaces traditional MSA-based pretraining in protein design.
• The approach generates accurate, diverse, and physically plausible protein structures.
• Protein language models hold significant promise for accelerating the development of novel functional proteins for biological and therapeutic applications.