ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning

TL;DR

ProtoMol tackles molecular property prediction by jointly modeling molecular graphs and textual descriptions. It introduces dual-branch hierarchical encoders, a layer-wise bidirectional cross-modal attention mechanism, and a unified semantic prototype space with learnable class-specific anchors to enable fine-grained cross-modal alignment. The training objective combines prototype alignment, task-specific supervision, and prototype-guided contrastive losses, and the approach demonstrates state-of-the-art performance on multiple classification and regression benchmarks. The framework also provides interpretable prototype activations and shows robustness across backbones and hyperparameters, with potential for extending to additional modalities and large-scale bioinformatics tasks.

Abstract

Multimodal molecular representation learning, which jointly models molecular graphs and their textual descriptions, enhances predictive accuracy and interpretability by enabling more robust and reliable predictions of drug toxicity, bioactivity, and physicochemical properties through the integration of structural and semantic information. However, existing multimodal methods suffer from two key limitations: (1) they typically perform cross-modal interaction only at the final encoder layer, thus overlooking hierarchical semantic dependencies; (2) they lack a unified prototype space for robust alignment between modalities. To address these limitations, we propose ProtoMol, a prototype-guided multimodal framework that enables fine-grained integration and consistent semantic alignment between molecular graphs and textual descriptions. ProtoMol incorporates dual-branch hierarchical encoders, utilizing Graph Neural Networks to process structured molecular graphs and Transformers to encode unstructured texts, resulting in comprehensive layer-wise representations. Then, ProtoMol introduces a layer-wise bidirectional cross-modal attention mechanism that progressively aligns semantic features across layers. Furthermore, a shared prototype space with learnable, class-specific anchors is constructed to guide both modalities toward coherent and discriminative representations. Extensive experiments on multiple benchmark datasets demonstrate that ProtoMol consistently outperforms state-of-the-art baselines across a variety of molecular property prediction tasks.

Figures (6)

• Figure 1: Overall architecture of the proposed ProtoMol. The workflow begins by independently encoding molecular graphs and their corresponding textual descriptions using hierarchical GNN and Transformer-based encoders, respectively. At each representation layer, bidirectional cross-modal attention modules facilitate information exchange between the graph and text modalities, progressively enriching the learned embeddings. These hierarchical representations are then projected into a unified semantic prototype space, which acts as a shared anchor point for both modalities. At multiple levels, prototype-based contrastive learning and alignment objectives are applied to refine the representations, enhancing intra-class compactness and inter-class separability across modalities.
• Figure 2: Visualization results of baseline methods and ProtoMol on the HIV dataset.
• Figure 3: Visualization results of baseline methods and ProtoMol on the BBBP dataset.
• Figure 4: Performance of different backbones on BBBP and Tox21 (classification) and on FreeSolv and Lipophilicity (regression): (a) graph encoder backbones; (b) textual encoder backbones.
• Figure 5: Hyperparameter sensitivity analysis of $\lambda_\text{align}$, $\lambda_\text{pred}$, $\lambda_\text{proto}$, number of prototype $N$, and number of $K$ entries on the BACE and HIV datasets for classification tasks, and on the ESOL and FreeSolv datasets for regression tasks.