Simple GNN Regularisation for 3D Molecular Property Prediction & Beyond

TL;DR

This work introduces Noisy Nodes, a lightweight regularisation strategy for GNNs designed to tackle oversmoothing by enforcing node-level diversity through input noise and a denoising objective. The approach is validated across 3D molecular property prediction (OC20 and QM9) and non-spatial graph benchmarks, yielding state-of-the-art results and enabling deeper or shared-weight architectures to perform competitively. Key contributions include a simple, broadly applicable regulariser that improves both depth-enabled and standard GNNs, along with extensive ablations and reproducibility details. The findings suggest Noisy Nodes as a practical, complementary tool for enhancing GNN representations in chemistry, materials science, and beyond.

Abstract

In this paper we show that simple noise regularisation can be an effective way to address GNN oversmoothing. First we argue that regularisers addressing oversmoothing should both penalise node latent similarity and encourage meaningful node representations. From this observation we derive "Noisy Nodes", a simple technique in which we corrupt the input graph with noise, and add a noise correcting node-level loss. The diverse node level loss encourages latent node diversity, and the denoising objective encourages graph manifold learning. Our regulariser applies well-studied methods in simple, straightforward ways which allow even generic architectures to overcome oversmoothing and achieve state of the art results on quantum chemistry tasks, and improve results significantly on Open Graph Benchmark (OGB) datasets. Our results suggest Noisy Nodes can serve as a complementary building block in the GNN toolkit.

Figures (9)

• Figure 1: Noisy Node mechanics during training. Input positions are corrupted with noise $\sigma$, and the training objective is the node-level difference between target positions and the noisy inputs.
• Figure 2: Per layer node latent diversity, measured by MAD on a 16 layer MPNN trained on OGBG-MOLPCBA. Noisy Nodes maintains a higher level of diversity throughout the network than competing methods.
• Figure 3: Validation curves, OC20 IS2RE ID. A) Without any node targets our model has poor performance and realises no benefit from depth. B) After adding a position node loss, performance improves as depth increases. C) As we add Noisy Nodes and parameters the model achieves SOTA, even with 3 layers, and stops overfitting. D) Adding Noisy Nodes allows a model with even fully shared weights to achieve SOTA.
• Figure 4: Adding Noisy Nodes with random flipping of input categories improves the performance of MPNNs, and the effect is accentuated with depth.
• Figure 5: Validation curve comparing with and without noisy nodes. Using Noisy Nodes leads to a consistent improvement.