[Paper Review] Multi-View Graph Neural Networks for Molecular Property Prediction

Hehuan Ma, Yatao Bian+6 more|arXiv (Cornell University)|May 17, 2020

Machine Learning in Materials Science53 references35 citations

TL;DR

Introduces MV-GNN and MV-GNN^cross, dual-view GNNs that leverage both node (atom) and edge (bond) information with a shared self-attentive readout and disagreement loss, plus cross-dependent message passing to boost expressiveness and interpretability for molecular property prediction.

ABSTRACT

The crux of molecular property prediction is to generate meaningful representations of the molecules. One promising route is to exploit the molecular graph structure through Graph Neural Networks (GNNs). It is well known that both atoms and bonds significantly affect the chemical properties of a molecule, so an expressive model shall be able to exploit both node (atom) and edge (bond) information simultaneously. Guided by this observation, we present Multi-View Graph Neural Network (MV-GNN), a multi-view message passing architecture to enable more accurate predictions of molecular properties. In MV-GNN, we introduce a shared self-attentive readout component and disagreement loss to stabilize the training process. This readout component also renders the whole architecture interpretable. We further boost the expressive power of MV-GNN by proposing a cross-dependent message passing scheme that enhances information communication of the two views, which results in the MV-GNN^cross variant. Lastly, we theoretically justify the expressiveness of the two proposed models in terms of distinguishing non-isomorphism graphs. Extensive experiments demonstrate that MV-GNN models achieve remarkably superior performance over the state-of-the-art models on a variety of challenging benchmarks. Meanwhile, visualization results of the node importance are consistent with prior knowledge, which confirms the interpretability power of MV-GNN models.

Motivation & Objective

• Motivate accurate molecular property prediction by jointly leveraging atom and bond information via a dual-view GNN framework.
• Develop a shared self-attentive readout to generate interpretable graph embeddings and stabilize training with a disagreement loss.
• Propose a cross-dependent message passing scheme to enhance information exchange between the two views.
• Provide theoretical analysis of expressive power relative to GIN and WL tests.
• Demonstrate empirical superiority on 11 benchmark datasets with robust performance and interpretability via attention visualizations.

Proposed method

• Two concurrent encoders: a Node-GNN (atom view) and an Edge-GNN (bond view) that produce node/edge embeddings.
• A shared self-attentive readout to compute a graph embedding from both views and enable interpretation of atom/functional-group importance.
• A disagreement loss that minimizes the prediction gap between the two views to stabilize training.
• (MV-GNN^cross) Cross-dependent message passing that allows one view to inform the other during each propagation step, enhancing information flow.
• Theoretical results showing MV-GNN is at least as powerful as GIN and MV-GNN^cross is strictly more powerful than GIN.
• An end-to-end loss combining prediction losses from both views with the disagreement term: L = L_pred + λ L_dis.

Experimental results

Research questions

• RQ1 Can a dual-view GNN architecture that separately processes node and edge information improve molecular property prediction over single-view models?
• RQ2 Does a shared self-attentive readout provide interpretable, atom-level explanations for predictions?
• RQ3 Does enforcing agreement between the two views via a disagreement loss stabilize training and improve robustness?
• RQ4 Can cross-dependent message passing further enhance expressive power and predictive performance beyond MV-GNN?
• RQ5 What is the theoretical expressive power of MV-GNN and MV-GNN^cross relative to GIN/WL tests?

Key findings

• MV-GNN and MV-GNN^cross outperform state-of-the-art baselines across 11 benchmarks in classification (AUC-ROC) and regression tasks.
• MV-GNN provides about 1.15 percentage point average AUC improvement over SOTA baselines; MV-GNN^cross attains about 1.65 percentage point improvement on average.
• MV-GNN^cross achieves higher robustness (lower standard deviations) and, in some cases, superior performance to MV-GNN.
• Cross-dependent message passing (MV-GNN^cross) reduces model size significantly and yields stronger expressiveness than vanilla MV-GNN.
• Theoretical results: MV-GNN is at least as expressive as GIN; MV-GNN^cross is strictly more expressive than GIN.
• Qualitative case studies show interpretable atom-level attention aligning with known toxicophores (e.g., trifluoromethyl and cyanide groups).