Research map / Graph learning / GNN architectures
On the Scalability of GNNs for Molecular Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies how message-passing GNNs, graph Transformers and hybrid GPS++ scale on the LargeMix collection of 2D molecular graphs under supervised multi-task pretraining, then probes/finetunes them on 38 downstream tasks and builds MolGPS.
Problem
GNNs for molecular property prediction have not shown the benefits of scale seen in language and image models, due to sparse-operation inefficiency, limited supervised data, and unclear architecture choice.
Approach
Systematically scales width, depth, number of molecules, labels and pretraining diversity of MPNN++, Transformer and GPS++ under supervised multi-task pretraining on LargeMix, adds phenomics data, and combines multi-layer fingerprints from several pretrained models for probing (MolGPS).
Key results
- MolGPS outperforms MolE on 21/22 TDC tasks and sets SOTA on 11/22 TDC tasks.
- MolGPS achieves state-of-the-art on 26 of 38 downstream tasks across TDC, Polaris and MoleculeNet.
- Scaling MPNN++ with phenomics pretraining to 3B parameters keeps a positive scaling trend, while the model without phenomics stopped scaling beyond 1B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase GNN depth | molecular GNNs · pretraining test sets | test error: 12.5% | Helped |
| increase depth beyond around 8-16 layers | molecular GNNs · quantum datasets | test performance | No clear effect |
| increase training molecules from 12.5% to 100% | GPS++ · pretraining test sets | model performance: 33.33% | Helped |
| increase training label fraction | pretrained GNNs · Polaris | Spearman correlation: 0.57 | Helped |
| increase training label fraction | pretrained GNNs · TDC | Spearman correlation: 0.54 | Helped |
| remove PCBA_1328 from pretraining (ablation) | pretrained GNNs · downstream tasks | performance | Hurt |
| add L1000_* pretraining datasets (ablation) | pretrained GNNs · some downstream tasks | performance | Hurt |
| probe from the first layer of the PCBA_1328 task head instead of the graph output network (ablation) | pretrained GNNs · some downstream tasks | performance | Helped |
| probe from PCQM4M_G25 task-head layers instead of the graph output network (ablation) | pretrained GNNs · downstream tasks | performance | Hurt |
| combine fingerprints from multiple layers and pretrained models | MolGPS probing · downstream tasks | performance | Helped |
| remove Phenomics pretraining (ablation) | 1B MolGPS · MoleculeNet test | AUROC: 0.806 vs 0.828; 0.802 vs 0.8; 0.797 vs 0.809; 0.649 vs 0.67 | Mixed |
| add Phenomics pretraining (ablation) | 1B MolGPS · TDC benchmark collection | normalized performance: 1.222 -> 1.305 | Helped |
Surprising findings
- Transformers and deeper MPNNs show degradation with depth on some tasks, and molecule-count scaling correlates only weakly with downstream performance.
- Pretraining without the L1000 datasets improved performance across all scales, as L1000 tasks hurt some downstream tasks.
Limitations
- Surprisingly could not scale MPNN++ to 3B without phenomics data, attributed to pretraining data mix limits.
- Scaling analysis restricted to parameters and molecule counts during pretraining and finetuning; aggregation functions and other factors unexplored.
- Potential data leakage between PCBA_1328 pretraining and TDC CYP*_Veith tasks.
Experimental setup
Datasets: LargeMix (L1000_VCAP, L1000_MCF7, PCBA_1328, PCQM4M_G25, PCQM4M_N4) phenomics imaging data TDC ADMET tasks Polaris benchmarks MoleculeNet subsets (BACE, BBBP, ClinTox, SIDER)
Benchmarks: TDC Polaris MoleculeNet
Metrics: AUROC AUPRC MAE Spearman correlation Pearson R2 average standardized score
Baselines: MolE GraphMVP MapLight+GNN TDC leaderboard SOTA Chemprop-RDKit
Builds on: GPS (General, Powerful, Scalable Graph Transformer) MPNN muP Graphium
Closest papers by meaning
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.89)
- Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.89)
- Exploring Molecular Pretraining Model at Scale (NeurIPS 2024 · similarity 0.89)
- Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks (ICML 2025 · similarity 0.89)
- Expressivity and Generalization: Fragment-Biases for Molecular GNNs (ICML 2024 · similarity 0.88)
- A Simple and Scalable Representation for Graph Generation (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.