EQCNNet: A New Paradigm for Protein-Ligand Binding Affinity Prediction Integrating Equivariance and Quantum Inspiration

EQCNNet is an SE(3)-equivariant quantum-inspired convolutional neural network designed specifically for protein-ligand binding affinity prediction. By combining Clebsch-Gordan products, spherical harmonics, and quantum-inspired convolutional layers, it accurately learns the 3D geometric structure of complexes, addressing the high computational cost or insufficient generalization ability of traditional methods and providing a new path for drug discovery.

One of the core steps in drug development is predicting the binding affinity between small molecule ligands and target proteins. Traditional physics-based methods (e.g., molecular dynamics simulations) are accurate but computationally expensive; traditional machine learning ignores 3D geometric properties and has insufficient generalization ability. Equivariant neural networks have great potential in molecular modeling, but how to integrate 3D geometric information with node-level interactions remains an open question.

The core innovation of EQCNNet lies in the combination of SE(3) equivariance and quantum-inspired convolution mechanisms:

1. SE(3) Equivariant Representation Learning: Achieved via Clebsch-Gordan products and spherical harmonics, ensuring rotation-translation invariance and capturing the relative spatial relationships of atoms;
2. Distance Encoding and Graph Construction: Using Gaussian radial basis functions to encode atomic distances, dynamically constructing graphs (nodes are atoms, edges represent adjacent relationships within 10.0Å);
3. Quantum-Inspired Convolution Layer: QCNNLayer simulates quantum convolution operations to enhance the expressive ability for complex molecular interactions.

Dataset: Adopted the PDBbind v.2019 refined set (4000+ complexes), with division strategies of 30% sequence identity (strict out-of-distribution) and 60% (moderate homology); Ablation Experiments: Designed 5 variants to verify the effectiveness of components, and the Wilcoxon signed-rank test showed that the complete model had p<0.05 over all variants, verifying the synergistic effect.

Model Configuration: maxl=2, max-sh=2, num-cg-levels=4, number of channels=32, number of atom types=5, batch size=16, learning rate=1e-3, training epochs=100; Dependencies and Environment: Based on PyTorch Geometric, dependencies include PyTorch>=1.9.0, torch-geometric, cormorant, etc. NVIDIA A100/V100 GPUs (16GB memory) are recommended.

EQCNNet provides a new path for structure-based drug design, with advantages including:

1. Physical consistency (SE(3) equivariance aligns with physical intuition);
2. Data efficiency (equivariant priors reduce data dependency);
3. Interpretability (CG product hierarchical structure provides decision understanding). Its generalization ability was verified on the CSAR-HiQ benchmark.

EQCNNet is an important advancement of geometric deep learning in molecular science. By integrating equivariant representation, quantum-inspired computing, and molecular modeling, it opens up a new direction for efficient and accurate drug screening tools, which is expected to accelerate new drug discovery and contribute to precision medicine.