Graph attention with structural features improves the generalizability of identifying functional sequences at a...

Area of Science:

• Computational Biology
• Protein Engineering
• Structural Biology

Background:

• Predicting protein-protein interface sequence compatibility is a critical challenge in biology.
• Existing sequence-based models struggle to generalize to distantly related protein sequences.

Purpose of the Study:

• To enhance the generalizability of protein interface prediction by integrating deep learning protein models.
• To develop and validate a novel deep learning architecture for predicting functional protein variants.

Main Methods:

• Designed and screened deep mutational libraries of SARS-CoV-2 Spike Receptor Binding Domain (RBD) for ACE2 receptor binding.
• Developed a graph attention network (GAN-PLM) combining protein structure graphs and protein language model (PLM) embeddings.
• Compared GAN-PLM performance against baseline supervised learning and sequence embedding models.

Main Results:

• A dataset of over 43,000 SARS-CoV-2 RBD variants was generated, exploring an expanded sequence space.
• Purely sequence-based models showed poor generalization to unseen variants.
• The developed GAN-PLM model significantly outperformed baseline models in predicting functional ACE2-binding variants across diverse sequences.

Conclusions:

• Integrating structure- and sequence-based features into deep learning models enhances prediction generalizability for protein interface function.
• The GAN-PLM approach offers a powerful tool for understanding and engineering protein interactions.
• This has broad implications for infectious disease research and therapeutic protein design.