AdaptMol: Adaptive Fusion from Sequence String to Topological Structure for Few-shot Drug Discovery

TL;DR

AdaptMol tackles few-shot molecular property prediction by integrating local graph topology with global SMILES semantics through an Adaptive Multi-level Attention module. It employs a Prototypical Network framework with dual encoders (graph and SMILES) and AMA that adaptively weights modalities at local and global levels, enabling robust representations under scarce labeled data. The authors also derive molecular rationales via Monte Carlo Tree Search to improve interpretability of predictions. Empirically, AdaptMol achieves state-of-the-art ROC-AUC on MoleculeNet benchmarks in both $5$-shot and $10$-shot settings and shows cross-domain generalization on the TDC dataset, underscoring its practical impact for data-limited drug discovery.

Abstract

Accurate molecular property prediction (MPP) is a critical step in modern drug development. However, the scarcity of experimental validation data poses a significant challenge to AI-driven research paradigms. Under few-shot learning scenarios, the quality of molecular representations directly dictates the theoretical upper limit of model performance. We present AdaptMol, a prototypical network integrating Adaptive multimodal fusion for Molecular representation. This framework employs a dual-level attention mechanism to dynamically integrate global and local molecular features derived from two modalities: SMILES sequences and molecular graphs. (1) At the local level, structural features such as atomic interactions and substructures are extracted from molecular graphs, emphasizing fine-grained topological information; (2) At the global level, the SMILES sequence provides a holistic representation of the molecule. To validate the necessity of multimodal adaptive fusion, we propose an interpretable approach based on identifying molecular active substructures to demonstrate that multimodal adaptive fusion can efficiently represent molecules. Extensive experiments on three commonly used benchmarks under 5-shot and 10-shot settings demonstrate that AdaptMol achieves state-of-the-art performance in most cases. The rationale-extracted method guides the fusion of two modalities and highlights the importance of both modalities.

Figures (5)

• Figure 1: (a) Overview of the proposed AdaptMol framework, where we plot a 2-way 2-shot task. AdaptMol is optimized over training tasks. Within each task $T_t$, the support set obtains prototypes for each class, while the query set optimizes the two molecular encoders and AMA module. During the testing phase, molecule in the query set is represented by the encoders and AMA module, used to compute similarity with the prototypes, leading to the final prediction. (b) The molecular graph encoder, generating molecular representations from the molecular graph. (c) Molecular SMILES encoder, using a large language model to capture the semantic and contextual information of molecular sequences. (d) The overall framework of the proposed AMA. The representation of all nodes within a molecule is sequentially processed through adaptive attention modules from local and global level, resulting in the final features-refined molecular representation.
• Figure 2: Illustration of the Monte Carlo Tree Search (MCTS) method for deriving chemical structure rationales (graph substructures) associated with high predicted molecular activity.
• Figure 3: Using AdaptMol as the scorer for (a) and (b), and single GIN as the scorer for (c) and (d), Monte Carlo Tree Search (MCTS) was employed to extract molecular rationales, which were highlighted within the original molecules. The associated scores for these rationales are presented beneath the figure.
• Figure 4: (a) ROC-AUC performance from Tox21 datasets. (b) ROC-AUC performance from SIDER datasets.
• Figure 5: The additional performance of all compared methods on three tasks with a support set of size 10 on the MoleculeNet benchmark. Each colored sector corresponds to a specific method, where the length of the sector reflects its performance based on F1-score and PRAUC (%). Starting from the horizontal right-pointing arrow, the methods are listed in the legend in a counterclockwise direction. Our AdaptMol corresponds to the last dark blue sector.