PACIA: Parameter-Efficient Adapter for Few-Shot Molecular Property Prediction

TL;DR

The paper tackles few-shot molecular property prediction under data scarcity by introducing PACIA, a parameter-efficient GNN adapter designed to modulate both node embeddings and propagation depth. By employing a unified hypernetwork-based adapter, PACIA enables hierarchical adaptation: task-level adaptation of the encoder and query-level adaptation of the predictor, with a small adaptive parameter budget relative to the main network. Empirical results on MoleculeNet and FS-Mol show state-of-the-art performance and favorable speed compared with gradient-based meta-learning methods, with ablations confirming the necessity of both adaptation levels and the effectiveness of depth modulation. The work demonstrates that amortized, hypernetwork-driven modulation can significantly improve generalization in few-shot MPP while avoiding overfitting. Overall, PACIA offers a practical, scalable approach to few-shot molecular prediction with explicit control over adaptation complexity and query-specific refinements.

Abstract

Molecular property prediction (MPP) plays a crucial role in biomedical applications, but it often encounters challenges due to a scarcity of labeled data. Existing works commonly adopt gradient-based strategy to update a large amount of parameters for task-level adaptation. However, the increase of adaptive parameters can lead to overfitting and poor performance. Observing that graph neural network (GNN) performs well as both encoder and predictor, we propose PACIA, a parameter-efficient GNN adapter for few-shot MPP. We design a unified adapter to generate a few adaptive parameters to modulate the message passing process of GNN. We then adopt a hierarchical adaptation mechanism to adapt the encoder at task-level and the predictor at query-level by the unified GNN adapter. Extensive results show that PACIA obtains the state-of-the-art performance in few-shot MPP problems, and our proposed hierarchical adaptation mechanism is rational and effective.

Figures (7)

• Figure 1: Illustration of the proposed PACIA. (a). The adaptation ratio ($\frac{|adaptive~params|}{|total~params|}$) of different methods and their testing performance on 1-shot tasks of Tox21. (b). Under the encoder-predictor framework, PACIA uses a unified GNN adapter, where hypernetworks are used to generate adaptive parameters, to modulate node embeddings and propagation depths in both GNN encoder and predictor.
• Figure 2: Ablation study on 10-shot tasks of Tox21.
• Figure 3: Molecular representation visualization for 10-shot case in task SR-p53 of Tox21.
• Figure 4: Illustration of query-level adaptation. (a), Molecular graphs of support molecules $x_1,~x_0$ and query molecules $q_1,~q_2,~q_3$. (b), Cosine similarities between query molecules and support molecules, and propagation depth taken to classify each query molecule.
• Figure 5: Testing ROC-AUC of PACIA and GIN on SIDER, with different number of labeled samples (shot) per task.