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

Abstract

Molecular similarity assessment is pivotal in drug discovery and materials science, yet conventional methods often fail to integrate complementary 2D topological and 3D geometric information effectively. Inspired by Radford et al. (2021) the cross-modal alignment capability of Contrastive Language-Image Pretraining (CLIP), this study proposes a novel graph neural network (GNN) framework that unifies 2D and 3D molecular representations through a CLIP-inspired contrastive learning strategy. Our dual-channel architecture employs a Graph Isomorphism Network (GIN) for 2D topology encoding and a Graph Attention Network (GAT) for 3D spatial feature extraction. These modality-specific embeddings are aligned in a shared latent space via the InfoNCE loss, emulating CLIP’s paradigm to maximize mutual information between 2D and 3D molecular structures. Extensive experiments on the QM9 dataset demonstrate that our model significantly outperforms traditional fingerprint-based methods and pure GNN baselines in molecular similarity assessment. Ablation studies further validate the critical role of cross-modal contrastive learning in bridging structural information. The framework exhibits robust generalizability across diverse molecular types, offering a pioneering adaptation of CLIP’s principles to non-visual domains. This work advances multimodal representation learning in cheminformatics and opens avenues for future applications in molecular-text retrieval and drug design.