SPIN-CGNN: Improved fixed backbone protein design with contact map-based graph construction and contact graph neural...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Deep Learning

Background:

• Deep learning methods, including Graph Neural Networks (GNNs), have advanced protein sequence design from fixed backbones.
• Conventional GNN approaches using K-nearest-neighbors (KNN) graphs underutilize critical edge information.

Purpose of the Study:

• To introduce SPIN-CGNN, a novel GNN model utilizing protein contact maps for improved edge information in sequence design.
• To evaluate SPIN-CGNN's performance against state-of-the-art methods in sequence recovery and design quality.

Main Methods:

• Developed SPIN-CGNN incorporating protein contact maps for nearest neighbor identification.
• Implemented auxiliary edge updates and selective kernels within the GNN architecture.
• Benchmarked SPIN-CGNN against existing methods using AlphaFold2 refolding ability, sequence recovery, perplexity, and amino acid composition metrics.

Main Results:

• SPIN-CGNN demonstrated comparable refolding ability to state-of-the-art methods.
• Achieved significant improvements in sequence recovery, perplexity, native amino acid composition, hydrophobic position conservation, and low complexity region design.
• Performance was validated across unseen, hallucinated, and diffusion-generated protein structures.

Conclusions:

• SPIN-CGNN offers enhanced performance in protein sequence design, particularly in sequence recovery and native-like characteristics.
• Deep learning-designed sequences, especially for generated structures, still require optimization in low complexity regions compared to native sequences.