Local-Global Multimodal Contrastive Learning for Molecular Property Prediction

TL;DR

The paper tackles molecular property prediction by integrating local chemical patterns and global molecular topology with chemistry-aware textual semantics. It introduces LGM-CL, a local-global multimodal contrastive learning framework that learns transferable representations by self-supervised alignment between dual graph encoders (Graph Transformer for global structure and Attentive FP for local structure) and a text modality formed from SMILES and LLM-generated chemistry-aware descriptions, with an additional downstream fusion of molecular fingerprints via Dual Cross-attention. Key contributions include a chemistry-aware prompt-based SMILES augmentation, a dual-encoder graph contrastive objective $\mathcal{L}_{CL}$ to align local and global views, a DeBERTa-based cross-modal text–SMILES alignment using $z_s$ and $z_a$ with NT-Xent, and a hierarchical multimodal fusion for downstream prediction. Extensive experiments on MoleculeNet demonstrate robust, state-of-the-art performance across classification and regression tasks, with ablations and interpretability analyses validating the value of the local-global and multimodal design for practical molecular property prediction.

Abstract

Accurate molecular property prediction requires integrating complementary information from molecular structure and chemical semantics. In this work, we propose LGM-CL, a local-global multimodal contrastive learning framework that jointly models molecular graphs and textual representations derived from SMILES and chemistry-aware augmented texts. Local functional group information and global molecular topology are captured using AttentiveFP and Graph Transformer encoders, respectively, and aligned through self-supervised contrastive learning. In addition, chemically enriched textual descriptions are contrasted with original SMILES to incorporate physicochemical semantics in a task-agnostic manner. During fine-tuning, molecular fingerprints are further integrated via Dual Cross-attention multimodal fusion. Extensive experiments on MoleculeNet benchmarks demonstrate that LGM-CL achieves consistent and competitive performance across both classification and regression tasks, validating the effectiveness of unified local-global and multimodal representation learning.

Figures (9)

• Figure 1: Overview of Model. (A) Pipeline of the model in the pre-training stage. (B) Pipeline of the model when applied to downstream tasks.
• Figure 2: Overview of global and local graph encoders. (A) The Graph Transformer encoder, which model global structural dependencies of molecular graphs through multi-head self-attention with adjacency-aware bias. (B) The Attentive FP encoder, which captures local chemical environments via attention-guided message passing followed by a GRU-based update and pooling operation.
• Figure 3: Schematic illustration of the dual cross-attention component and text encoder in our framework. (A) The details of the dual cross-attention component. (B) DeBERTa-based encoder for SMILE and LLM-augmented textual descriptions.
• Figure 4: Workflow for enriching SMILES strings with chemically meaningful natural language descriptions generated by Mistral-7B-Instruct-v0.3.
• Figure 5: Ablation study of dual cross-attention fusion strategies across six molecular property prediction tasks, reporting ROC-AUC for classification and RMSE for regression.