Efficient and Robust Continual Graph Learning for Graph Classification in Biology

Ding Zhang, Jane Downer, Can Chen, Ren Wang

Introduction

In biological research, such as drug discovery or enzyme function prediction, data is often generated in streams. Standard Graph Neural Networks (GNNs) trained on these streams suffer from catastrophic forgetting, where new task acquisition degrades performance on previously learned tasks. While continual learning (CL) is well-studied in computer vision, its application to graph classification presents unique hurdles: memory explosion due to graph storage and vulnerability to backdoor attacks where subtle structural triggers can compromise model integrity.

Methodology

The paper introduces PSCGL (Perturbed and Sparsified Continual Graph Learning), a framework designed to address these challenges through three core components:

  1. Perturbed Graph Sampling: Instead of simple random or confidence-based sampling, PSCGL selects representative graphs for the memory buffer by evaluating their stability under feature perturbations. This ensures that the stored samples are robust representatives of their respective class distributions.
  2. Motif-based Graph Sparsification: To mitigate the memory overhead of storing entire graphs, PSCGL prunes nodes that do not participate in critical biological motifs, specifically 3-node cliques (triangles). This "backbone" extraction reduces storage requirements and computational cost.
  3. Consistency Training: The model is trained to maintain invariant representations between original and perturbed graph versions, enhancing its robustness to noise.
Overview of the PSCGL framework, illustrating the perturbed sampling, motif-based sparsification, and consistency training stages across sequential tasks.

Experimental Evaluation

The framework was evaluated on two prominent biological benchmarks, Enzymes and Aromaticity, using a class-incremental setup where the model learns two classes at a time.

  • Superior Performance: On the Enzymes dataset, PSCGL achieved a 32% increase in Average Performance (AP) compared to standard random-sampling replay baselines.
  • Reduced Forgetting: The Average Forgetting (AF) rate was significantly lowered, indicating stable performance across long task sequences (e.g., maintaining accuracy on Task 1 after learning Task 10).
  • Attack Defense: Motif-based sparsification proved highly effective against backdoor attacks, reducing the Attack Success Rate (ASR) from roughly 72% to as low as 9% on the Enzymes dataset.

Key Insights

  • Sparsification as a Shield: A key insight is that biological motifs serve as a natural defense mechanism. Since attackers often place triggers in non-essential structural regions, motif-based pruning effectively "cleans" the data before it enters the memory buffer.
  • Stability over Confidence: PSCGL argues that confidence at a single point can be misleading. By testing confidence under perturbation, the framework identifies graphs that sit in the "center" of a class distribution.
  • Efficiency Trade-offs: The study demonstrates a viable efficiency-accuracy trade-off, where up to 50% of nodes can be pruned without significant performance loss, proving much of a graph's connectivity is redundant for classification.

Conclusion

PSCGL provides a robust and efficient solution for continual graph learning in biological contexts. While powerful, the authors acknowledge an accuracy-efficiency trade-off as sparsification ratios increase. Future work will explore more complex motifs (e.g., 4-node tetrads) and adaptive sparsification techniques to further optimize the balance between structural integrity and computational efficiency.

Summary updated

BibTeX
@article{zhang2025efficient,
  author = {Ding Zhang and Jane Downer and Can Chen and Ren Wang},
  title = {{Efficient and Robust Continual Graph Learning for Graph Classification in Biology}},
  journal = {IEEE Trans. Signal Inf. Process. Over Networks},
  year = {2025},
  doi = {10.1109/tsipn.2025.3592321},
}
Figure