ChiGNN: A Protein Side Chain Conformation Generation Model Based on Torsional Diffusion

ChiGNN is a lightweight graph neural network model targeting the protein side chain conformation recovery problem. It adopts a torsional diffusion method with von Mises distribution, providing new insights for computational drug design and structural biology. Its core innovations include mathematical modeling to handle the periodicity of dihedral angles, a lightweight architecture that lowers the research barrier, and the ability to quantify uncertainty.

AlphaFold2 has made breakthroughs in backbone prediction, but the recovery rate of side chain χ₁ angles is only 70-75%, while side chains are the core of function (active sites, drug binding, hydrogen bond networks). Traditional methods like SCWRL4/Rosetta treat this as a deterministic optimization problem, ignoring probability distributions; ChiGNN uses a generative probabilistic approach and is the first to achieve calibrated uncertainty quantification in a lightweight architecture.

Torsional Diffusion and Circular Distribution: Using von Mises distribution (a natural distribution for circular spaces) as the noise source to handle the periodicity of dihedral angles and avoid boundary errors from Gaussian diffusion.

Model Architecture: 4 layers of GCNConv + batch normalization + residual connections, with only 80,404 parameters, which can run smoothly on Colab T4 GPU.

Training Dataset: 597 high-resolution (<2.0Å) structures selected from PDB-REDO. Preprocessing includes graph construction (Cα nodes, 8Å edges), node features (residue type, Cα coordinates, φ/ψ), and label extraction (χ₁-χ₄ calculated via BioPython); AdamW optimizer + cosine annealing, 50 epochs, with the best performance at epoch 42 (validation loss: 0.0885).

Test set metrics: χ₁ circular MAE is 56.41° (baseline ~60-65°, random ~90°); χ₁ recovery rate (±40°) is 53.9% (baseline ~47%, random ~33%). Although it does not reach SCWRL4's 83%, the lightweight model (80K parameters) has already exceeded the statistical baseline.

Uncertainty Quantification: The Spearman correlation coefficient between prediction confidence and error is 0.299 (p<0.001), which can identify high-uncertainty predictions.

Visualization: Rose plots reproduce the trimodal distribution of χ₁ (g⁻≈-60°, t≈180°, g⁺≈+60°), reflecting physical plausibility.

Limitations: Small data scale (597 vs tens of thousands in industrial settings); GCNConv lacks rotational equivariance; 30% recovery rate gap compared to SCWRL4.

Future Directions: Introduce equivariant graph neural networks (EGNN/GVP); expand training data; explore more complex diffusion scheduling strategies.

Accessibility: Open-source Colab notebooks, documentation, and pre-trained models. With a Google account, reproduction can be done in 30 minutes, promoting democratization of the field.

Drug Design Applications: Quickly evaluate side chain conformation uncertainty, identify key residues that need experimental validation, and optimize lead compound design.