Source-linked AI summary
GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction
Shuangli Li, Jingbo Zhou, Tong Xu, Dejing Dou, Hui Xiong
TL;DR
Molecular property prediction is limited by scarce labeled data, while common graph augmentations can alter molecular structure and geometric information is underused. GeomGCL addresses this gap by jointly learning from 2D and 3D molecular views with adaptive message passing and geometric contrastive supervision. On seven molecular datasets, it outperforms state-of-the-art GNN and graph contrastive learning methods.
Problem
Scarce labels and structure-altering graph augmentations limit molecular property prediction, while molecular geometric information remains underexplored.
Method
GeomGCL uses dual-channel geometric message passing and contrastive learning to jointly exploit 2D and 3D molecular views without random augmentation.
Results
GeomGCL outperforms state-of-the-art GNN and graph contrastive learning methods on seven molecular datasets.
Takeaways & Limitations
Contrasting consistent 2D and 3D molecular structures provides a geometry-based route to molecular representation learning without corrupting chemical structure.
Abstract
from arXiv · showhide
Recently many efforts have been devoted to applying graph neural networks (GNNs) to molecular property prediction which is a fundamental task for computational drug and material discovery. One of major obstacles to hinder the successful prediction of molecule property by GNNs is the scarcity of labeled data. Though graph contrastive learning (GCL) methods have achieved extraordinary performance with insufficient labeled data, most focused on designing data augmentation schemes for general graphs. However, the fundamental property of a molecule could be altered with the augmentation method (like random perturbation) on molecular graphs. Whereas, the critical geometric information of molecules remains rarely explored under the current GNN and GCL architectures. To this end, we propose a novel graph contrastive learning method utilizing the geometry of the molecule across 2D and 3D views, which is named GeomGCL. Specifically, we first devise a dual-view geometric message passing network (GeomMPNN) to adaptively leverage the rich information of both 2D and 3D graphs of a molecule. The incorporation of geometric properties at different levels can greatly facilitate the molecular representation learning. Then a novel geometric graph contrastive scheme is designed to make both geometric views collaboratively supervise each other to improve the generalization ability of GeomMPNN. We evaluate GeomGCL on various downstream property prediction tasks via a finetune process. Experimental results on seven real-life molecular datasets demonstrate the effectiveness of our proposed GeomGCL against state-of-the-art baselines.
1 Introduction
Molecular property prediction is important for drug and material discovery, but scarce labels and structure-altering graph augmentations limit existing approaches. GeomGCL addresses these issues by contrasting molecular 2D and 3D geometric views while preserving structure, and outperforms state-of-the-art methods on seven datasets.
- Molecular property prediction supports selecting chemical molecules with desired characteristics for computational drug and material discovery.
- Scarce labeled data, whose collection can be costly, hinders molecular property prediction with GNNs.
- Existing molecular GCL augmentations can alter molecular structure and semantics by dropping atoms, perturbing edges, or masking attributes.
- GeomGCL contrasts complementary 2D chemical and 3D geometric views without changing chemical semantics.
- GeomMPNN adaptively processes both views, while a spatial regularizer preserves relative geometric relations and improves performance.
- Experiments on seven molecular datasets show that GeomGCL outperforms state-of-the-art GNN and graph contrastive learning methods.
2 Related Work
Related work spans molecular representation learning with GNNs, geometric graph learning, and graph contrastive learning. Existing geometric methods generally use a single view or target small molecules, motivating adaptive joint use of 2D and 3D information.
- Earlier molecular methods used descriptors or fingerprints, whereas GNNs learn representations from graph structures.
- Geometric molecular models incorporate spatial information and may enforce translation and rotation invariance through geometric kernels or node-edge interactions.
- Existing efficient geometric methods often target small molecules and do not jointly incorporate 2D and 3D geometric information.
- Graph contrastive learning performs well in many applications, but molecular graph models have received comparatively little attention.
3 Preliminaries
The preliminaries represent a molecule with atoms, bonds, and coordinates, then construct 2D and 3D view graphs. The learned geometric encoder produces a label-free molecular representation for downstream property prediction.
- A geometric molecular graph is G = (V, E, C), with atom set V, bond set E, and coordinate matrix C in spatial dimension d ∈ {2, 3}.
- The 2D view graph uses primary covalent bonds and two-dimensional atom coordinates.
- The 3D view graph uses generated three-dimensional coordinates and edges connecting atoms within cutoff distance dθ.
- Given atom and bond features, the objective is to learn a molecular representation vector without labels and fine-tune it for downstream property prediction.
4 Model Framework
GeomGCL derives 2D and 3D graphs from a molecule and uses a dual-channel GeomMPNN to adaptively learn representations from both geometric views. Because 3D coordinates can vary across coordinate systems, the framework represents stable distances and angles with geometric embeddings.
- GeomGCL derives 2D and 3D view graphs from a molecule in SMILES format before applying its dual-channel GeomMPNN.
- The framework addresses changeable coordinate systems by computing definite distance and angle factors for geometric representation.
4.1 Geometric Embedding
GeomGCL converts molecular distances and angles into geometric embeddings that remain consistent across coordinate systems. It encodes local 2D geometry and global 3D geometry using radial basis functions.
- Geometric factor encoding: Coordinate-invariant geometric factors—distances and angles—are encoded with radial basis functions into geometric embeddings.The method avoids directly using changeable, inconsistent coordinates.
- 2D geometric view: Local distance l and 2D angle φ capture covalent-bond geometry and chemical information.
- 3D geometric view: Global distance r supplies non-local molecular correlations, while 3D angle θ represents the spatial distribution of global connections.
- Geometric factor encoding: Several RBF layers encode the diverse geometric factors into dual-level representations.
4.2 Adaptive Geometric Message Passing
GeomMPNN adaptively learns molecular topology and geometry through dual 2D and 3D channels. Its node–edge interactions use distance- and angle-aware message passing before attentive graph pooling.
- Architecture: GeomMPNN uses Node→Edge, Edge→Edge, and Edge→Node stages followed by Node→Graph attentive pooling.Both channels follow this architecture while learning 2D and 3D geometric factors.
- Node→Edge Message Passing: Node→Edge updates integrate node embeddings with chemical bond features for 2D edges and global distance embeddings for 3D edges.The resulting edge embeddings combine chemical semantic and geometric information.
- Edge→Edge Message Passing: Edge→Edge propagation captures geometric distributions through 2D angle-aware aggregation and 3D hierarchical aggregation across angle domains.The 3D process contains local and global stages, with max pooling extracting high-level spatial distribution information.
- Edge→Edge Message Passing: The 3D channel divides neighboring edges into angle domains and aggregates them locally before global max pooling.
- Edge→Node Message Passing: Edge→Node propagation incorporates local and global distance embeddings while aggregating neighboring edges in 2D and distance domains in 3D.
- Node→Graph Attentive Pooling: After stacked message-passing layers, node representations reflect topology and molecular geometry before attentive pooling produces graph-level 2D and 3D representations.
4.3 Geometric Contrastive Optimization
GeomGCL uses correlated 2D and 3D molecular views for contrastive learning, avoiding additional fake samples while encouraging complementary representations. Its optimization combines contrastive learning with spatial regularization across adjacent angle domains.
- Cross-view contrastive learning: Correlated 2D and 3D molecular views supervise each other without constructing additional fake samples.The method maps both views into a contrastive-learning space and treats corresponding embeddings as positive pairs.
- Spatial regularization: The spatial regularizer encourages transformation matrices of adjacent angle domains to remain similar.This constraint is introduced to reflect local spatial correlations across 3D geometric domains.
- Joint optimization: The final optimization combines the spatial and contrastive losses, with λ controlling the spatial regularizer’s importance.The trade-off parameter guides how strongly spatial regularization contributes to representation learning.
4.4 Downstream Inference
After geometric contrastive pretraining, GeomGCL combines the learned 2D and 3D representations for downstream molecular property prediction. Classification uses cross-entropy loss, while regression uses L1 loss.
- Downstream prediction: Well-trained 2D and 3D geometric MPNNs provide representations that are combined for downstream molecular property prediction.The downstream prediction stage is performed through finetuning.
- Task-specific losses: Classification tasks use cross-entropy loss, whereas regression tasks use L1 loss.The spatial regularizer is also adopted during downstream training.
- Task-specific losses: The prediction compares the predicted value ŷ with the measured true value y for a specific molecular property.These quantities define the output and target used by the prediction objective.
5 Experiments
GeomGCL is evaluated on seven MoleculeNet datasets spanning physiology classification and physical-chemistry regression tasks. The study compares it with message-passing, geometry-learning, and graph-contrastive baselines using cross-validation and standard task metrics.
- Datasets: Seven MoleculeNet datasets cover four physiology classification tasks and three physical-chemistry regression tasks.The classification datasets are ClinTox, Sider, Tox21, and ToxCast; the regression datasets are ESOL, FreeSolv, and Lipophilicity.
- Baselines: The baselines include molecular message-passing, geometry-learning GNN, and graph contrastive learning methods.This comparison evaluates GeomGCL against multiple categories of molecular representation-learning approaches.
- Evaluation protocol: All methods are evaluated with 10-fold cross-validation to report averaged experimental results.The datasets are randomly split into training, validation, and testing sets with a 0.8/0.1/0.1 ratio.
- Evaluation protocol: Classification performance is measured with ROC-AUC, while regression performance is measured with RMSE.Validation data supports early stopping and model selection.
- Implementation: 3D molecular structures are generated 50 times using stochastic MMFF optimization implemented in RDKit.Training uses Adam with learning rate 1e-3, batch sizes 256 for contrastive learning and 32 for finetuning, and τ = 0.5.
5.2 Performance Evaluation
GeomGCL outperforms the evaluated baselines across classification and regression tasks, while ablations show that combining 2D and 3D geometric views and using spatial regularization improves performance.
- GeomGCL significantly outperforms all baselines on both graph classification and regression tasks.The evaluation covers seven molecular graph datasets, with Cls.Ave and Reg.Ave denoting average classification and regression results.
- Ablation Study: GeomGCL achieves the best performance among variants, supporting contrastive and synergistic learning of 2D-3D geometric structures.Variants isolate 2D message passing, 3D message passing, dual-channel message passing without contrastive learning, and removal of the spatial regularizer.
- Ablation Study: Removing or decoupling geometric components yields a significant performance drop, while the spatial regularizer helps discriminate relative correlations across angle domains.GeomMPNN performs worse than GeomGCL, supporting the benefit of the geometric contrastive learning scheme.
- Parameter Analysis: Performance first improves and then slightly declines as the spatial-regularizer coefficient increases, indicating an appropriate trade-off weight is beneficial.Across the tested coefficients, GeomGCL remains stable and better than baseline methods.
6 Conclusion
The paper presents GeomGCL as a bridge between geometric structure learning and graph contrastive learning for molecular representation learning. Its dual-channel network captures distance and angle information from 2D and 3D views, while geometry-based contrastive learning enhances representations and performs effectively on downstream property prediction.
- Its dual-channel message passing networks capture distance and angle information under both 2D and 3D molecular views.
- A geometry-based contrastive learning strategy with a spatial regularizer enhances molecular representation learning.
- Experiments on downstream property prediction tasks demonstrate the effectiveness of GeomGCL.