MolContraCLIP: Structurally similar molecule retrieval algorithm based on graph neural network and CLIP model.

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Molecular similarity assessment is crucial for drug discovery and materials science.
• Conventional methods struggle to integrate 2D topological and 3D geometric molecular information effectively.

Purpose of the Study:

• To develop a novel graph neural network (GNN) framework that unifies 2D and 3D molecular representations.
• To adapt the cross-modal alignment strategy from Contrastive Language-Image Pretraining (CLIP) for molecular data.

Main Methods:

• A dual-channel GNN architecture using Graph Isomorphism Network (GIN) for 2D and Graph Attention Network (GAT) for 3D.
• Employing a CLIP-inspired contrastive learning strategy with InfoNCE loss to align 2D and 3D embeddings in a shared latent space.
• Utilizing the QM9 dataset for extensive experimental validation.

Main Results:

• The proposed model significantly outperforms traditional fingerprint-based methods and pure GNN baselines in molecular similarity assessment.
• Ablation studies confirmed the effectiveness of cross-modal contrastive learning in integrating structural information.
• The framework demonstrated robust generalizability across various molecular types.

Conclusions:

• The novel GNN framework successfully unifies 2D and 3D molecular information, advancing multimodal representation learning in cheminformatics.
• This pioneering adaptation of CLIP principles to non-visual domains opens new possibilities for molecular-text retrieval and drug design.