Efficient and Robust Continual Graph Learning for Graph Classification in Biology

Ding Zhang, Jane Downer, Can Chen, Ren Wang

TL;DR

The paper introduces PSCGL (Perturbed and Sparsified Continual Graph Learning), a framework designed to help Graph Neural Networks (GNNs) learn biological tasks sequentially without forgetting previous ones. By combining a "perturbed sampling" strategy to pick the best data for memory replay and a "motif-based sparsification" technique to prune less important nodes, the authors achieve state-of-the-art performance while simultaneously defending against malicious backdoor attacks.

What Problem Does This Paper Address?

In biological research, data often arrives in "streams"—for example, discovering new enzyme functions or molecular properties over time. Standard GNNs suffer from catastrophic forgetting: when trained on a new task, they quickly lose the ability to perform well on previous ones.

While "continual learning" has been studied for images and text, applying it to graph classification (where an entire molecule or protein is a single data point) is difficult. Existing methods face two major hurdles:

  1. Efficiency: Storing old graphs in a "memory buffer" for retraining leads to a "memory explosion" as the number of tasks grows.
  2. Robustness: Biological models are sensitive. Malicious "backdoor attacks"—where a tiny, hidden trigger in the graph structure misleads the model—can have severe consequences in drug discovery or medical diagnosis.

Methods

The authors propose the PSCGL framework, which uses three main components to tackle these issues:

  • Perturbed Graph Sampling: To prevent forgetting, models often keep a "buffer" of old examples to replay later. PSCGL selects these examples by applying small "perturbations" (noise) to node features. It chooses graphs that maintain high model confidence even when noisy. This ensures the stored samples are "representative" and "stable" within their data neighborhood.
  • Motif-based Graph Sparsification: To save space, PSCGL prunes nodes that don't participate in "motifs" (specifically 3-node cliques or triangles). In biology, these motifs often represent critical functional groups. By keeping only the structural "backbone," the model reduces storage needs and computational overhead.
  • Consistency Training: During training, the model is forced to produce similar internal representations (embeddings) for both the original graph and its perturbed version. This "consistency" makes the model more robust to noise and helps it learn more nuanced, invariant patterns.

Datasets

The study evaluates PSCGL on two prominent biological benchmarks:

  • Enzymes: A dataset of 600 protein graphs where the goal is to classify them into one of six enzyme types.
  • Aromaticity: A larger dataset of 3,945 molecular graphs (filtered to 34 classes) where the model predicts the number of aromatic atoms in a molecule.

The experiments use a class-incremental setup, where the model learns two classes at a time and must eventually distinguish between all classes encountered so far without being told which "task" a test sample belongs to.

Key Insights and Ideas

  • Sparsification as a Shield: One of the most clever insights is that motif-based sparsification naturally defends against backdoor attacks. Attackers usually hide their "triggers" in less critical parts of the graph. By pruning nodes that don't contribute to fundamental biological motifs (triangles), PSCGL often accidentally "deletes" the attacker's trigger before it can reach the memory buffer.
  • Stability over Confidence: Most sampling methods just pick the graphs the model is most "confident" about. PSCGL argues that confidence at a single point can be a fluke of the model's complex landscape. By testing confidence under perturbation, they identify graphs that sit in the "center" of a class distribution.
  • The Efficiency Trade-off: The authors show that you can prune up to 50% of a graph's nodes (sparsification ratio $r=0.5$) and still outperform standard non-replay baselines, proving that much of a graph's connectivity is "redundant" for the task of classification.

Key Results

  • Superior Performance: On the Enzymes dataset, PSCGL achieved a 32% increase in Average Performance (AP) compared to a standard random-sampling replay baseline.
  • Reduced Forgetting: PSCGL significantly lowered the "Average Forgetting" (AF) rate, maintaining high accuracy on Task 1 even after learning Task 10.
  • Attack Defense: In experiments with poisoned data, PSCGL dropped the Attack Success Rate (ASR) from roughly 72% (no sparsification) to as low as 9% (with high sparsification) on the Enzymes dataset.
  • Scalability: Even at high levels of sparsification ($r=0.5$), PSCGL remained competitive with other memory-heavy methods, making it a viable choice for resource-constrained environments.

Limitations and Future Work

The authors acknowledge that while motif-based sparsification is powerful, there is a clear accuracy-efficiency trade-off. As the sparsification ratio increases beyond a certain point, the model eventually loses critical structural information, leading to performance degradation.

Future directions suggested by the work include:

  • Exploring more complex motifs beyond simple triangles (e.g., 4-node tetrads or specific chemical rings).
  • Testing the framework on non-biological graph datasets (like social networks) to see if the motif-based logic holds.
  • Developing more sophisticated "adaptive" sparsification that decides how much to prune based on the specific complexity of each task.

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