OGB
ogbg-molpcba
Structure
437,929Graphs
26Nodes
28Edges
128Classes
9Node Feature Dim
3Edge Feature Dim
Coverage
172Results
1Variants
30Papers
- Latent-Hysteresis Graph ODEs: Modeling Coupled Topology-Feature Evolution via Continuous Phase Transitions Latent-Hysteresis Graph ODEs: Modeling Coupled Topology-Feature Evolution via Continuous Phase Transitions
- SigGate-GT: Taming Over-Smoothing in Graph Transformers via Sigmoid-Gated Attention SigGate-GT: Taming Over-Smoothing in Graph Transformers via Sigmoid-Gated Attention
- When does global attention help: a unified empirical study on atomistic graph learning When does global attention help: a unified empirical study on atomistic graph learning
- Adaptive Canonicalization with Application to Invariant Anisotropic Geometric Networks Adaptive Canonicalization with Application to Invariant Anisotropic Geometric Networks
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization
- WWW 2025: Graph Self-Supervised Learning with Learnable Structural and Positional Encodings Graph Self-Supervised Learning with Learnable Structural and Positional Encodings WWW 2025
- AAAI 2024: A Scalable and Effective Alternative to Graph Transformers A Scalable and Effective Alternative to Graph Transformers AAAI 2024
- Automatic Graph Topology-Aware Transformer Automatic Graph Topology-Aware Transformer
- ICLR 2024: Graph Parsing Networks Graph Parsing Networks ICLR 2024
- Two Trades are not Baffled: Condensing Graph via Crafting Rational Gradient Matching Two Trades are not Baffled: Condensing Graph via Crafting Rational Gradient Matching
- Deep Prompt Tuning for Graph Transformers Deep Prompt Tuning for Graph Transformers
- Rethinking the Power of Graph Canonization in Graph Representation Learning with Stability Rethinking the Power of Graph Canonization in Graph Representation Learning with Stability
- NeurIPS 2023: Fused Gromov-Wasserstein Graph Mixup for Graph-level Classifications Fused Gromov-Wasserstein Graph Mixup for Graph-level Classifications NeurIPS 2023
- ICML 2023: Towards Better Graph Representation Learning with Parameterized Decomposition & Filtering Towards Better Graph Representation Learning with Parameterized Decomposition & Filtering ICML 2023
- NeurIPS 2023: Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling Graph Mixture of Experts: Learning on Large-Scale Graphs with Explicit Diversity Modeling NeurIPS 2023
- ICML 2023: On the Connection Between MPNN and Graph Transformer On the Connection Between MPNN and Graph Transformer ICML 2023
- LoG 2022: You Can Have Better Graph Neural Networks by Not Training Weights at All: Finding Untrained GNNs Tickets You Can Have Better Graph Neural Networks by Not Training Weights at All: Finding Untrained GNNs Tickets LoG 2022
- LoG 2022: Expander Graph Propagation Expander Graph Propagation LoG 2022
- Motif-based Graph Representation Learning with Application to Chemical Molecules Motif-based Graph Representation Learning with Application to Chemical Molecules
- KDD 2022: Unified 2D and 3D Pre-Training of Molecular Representations Unified 2D and 3D Pre-Training of Molecular Representations KDD 2022
- NeurIPS 2022: Affinity-Aware Graph Networks Affinity-Aware Graph Networks NeurIPS 2022
- ICLR 2022: Agent-based Graph Neural Networks Agent-based Graph Neural Networks ICLR 2022
- KDD 2022: Condensing Graphs via One-Step Gradient Matching Condensing Graphs via One-Step Gradient Matching KDD 2022
- KDD 2022: Compressing Deep Graph Neural Networks via Adversarial Knowledge Distillation Compressing Deep Graph Neural Networks via Adversarial Knowledge Distillation KDD 2022
- Transformer for Graphs: An Overview from Architecture Perspective Transformer for Graphs: An Overview from Architecture Perspective
- ICML 2022: Molecular Representation Learning via Heterogeneous Motif Graph Neural Networks Molecular Representation Learning via Heterogeneous Motif Graph Neural Networks ICML 2022
- GRPE: Relative Positional Encoding for Graph Transformer GRPE: Relative Positional Encoding for Graph Transformer
- ICLR 2021: Graph Neural Networks with Learnable Structural and Positional Representations Graph Neural Networks with Learnable Structural and Positional Representations ICLR 2021
- NeurIPS 2021: Reconstruction for Powerful Graph Representations Reconstruction for Powerful Graph Representations NeurIPS 2021
- Parameterized Hypercomplex Graph Neural Networks for Graph Classification Parameterized Hypercomplex Graph Neural Networks for Graph Classification
128-task multi-label graph classification Predict activity separately for 128 PubChem BioAssay tasks. Inputs use the standard OGB nine-field categorical atom representation and three-field categorical bond representation generated from SMILES. Evaluated by AP.
Dataset details
Task
Predict activity separately for 128 PubChem BioAssay tasks. Labels are binary where measured, but many molecule-task pairs are unassayed and stored as missing rather than negative.
Data
OGB adopts 437,929 RDKit-processable compounds from MoleculeNet’s PCBA collection. Graphs represent atoms and covalent bonds; assay outcomes originate from PubChem’s heterogeneous high-throughput screening records.
Features
Inputs use the standard OGB nine-field categorical atom representation and three-field categorical bond representation generated from SMILES. Assay identifiers and labels are targets, not molecular input features.
Splits and evaluation
Molecules are separated by two-dimensional scaffold into the official train, validation, and test sets. Average precision is calculated for each evaluable assay and averaged across the 128 tasks, ignoring missing labels.
Quirks and pitfalls
Only about 1.4% of assigned labels are positive, so accuracy and unweighted treatment of missing labels are misleading. The assays measure different biological endpoints and have different coverage; a missing label means unknown, not inactive. Random splits and per-task metrics are not directly comparable to the official mean AP.
Sources
OGB scaffold split; AP is averaged across all 128 tasks.
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | AP ↑ | Accuracy | ROC-AUC | Paper |
|---|---|---|---|---|---|
| 1 | GRPE-Large | 0.3150
±
0.0010 | — | — | |
| 2 | Graphormer | 0.3140 | — | — | |
| 3 | Graphormer-FLAG | 0.3139
±
0.0032 | — | — | |
| 4 | Graphormer FT | 0.3130
±
0.0030 | — | — | |
| 5 | UnifiedMolPretrain | 0.3125 | — | — | |
| 6 | GRPE-Standard | 0.3077
±
0.0007 | — | — | |
| 7 | GSO | 0.3031
±
0.0026 | — | — | |
| 8 | GraphGPS FT | 0.3010
±
0.0130 | — | — | |
| 9 | CRaWl | 0.2986
±
0.0025 | — | — | |
| 10 | SigGate-GT | 0.2984
±
0.0031 | — | — | |
| 11 | GINE-AP | 0.2979
±
0.0030 | — | — | |
| 12 | GINE | 0.2979
±
0.0030 | — | — | |
| 13 | GraphGPS DeepGPT | 0.2970
±
0.0200 | — | — | |
| 14 | A-NLSF | 0.2968
±
0.0022 | — | — | |
| 15 | GECO | 0.2961
±
0.0008 | — | — | |
| 16 | EGT-Large | 0.2961
±
0.0024 | — | — | |
| 17 | LiGhT FT | 0.2950
±
0.0060 | — | — | |
| 18 | PHC-GN | 0.2947
±
0.0026 | — | — | |
| 19 | PHC-2 | 0.2947
±
0.0026 | — | — | |
| 20 | PHC-GNN | 0.2947
±
0.0026 | — | — | |
| 21 | PHC-2-C | 0.2941
±
0.0015 | — | — | |
| 22 | PHC-3 | 0.2935
±
0.0028 | — | — | |
| 23 | PHC-4 | 0.2930
±
0.0016 | — | — | |
| 24 | Subgraphormer | 0.2930
±
0.0040 | — | — | |
| 25 | PHC-4-Q | 0.2921
±
0.0023 | — | — | |
| 26 | Exphormer | 0.2920
±
0.0030 | — | — | |
| 27 | PHC-1 | 0.2917
±
0.0016 | — | — | |
| 28 | PHC-5 | 0.2913
±
0.0024 | — | — | |
| 29 | GPS | 0.2907
±
0.0028 | — | — | |
| 30 | GraphGPS | 0.2907
±
0.0028 | — | — | |
| 31 | GIN | 0.2902
±
0.0017 | — | — | |
| 32 | Graphormer DeepGPT | 0.2890
±
0.0090 | — | — | |
| 33 | DGN | 0.2885
±
0.0030 | — | — | |
| 34 | HM-GNN | 0.2870
±
0.0026 | — | — | |
| 35 | DeeperG | 0.2842
±
0.0043 | — | — | |
| 36 | DeeperGCN+FLAG | 0.2842
±
0.0043 | — | — | |
| 37 | PNA-LSPE | 0.2840
±
0.0021 | — | — | |
| 38 | PNA | 0.2838
±
0.0035 | — | — | |
| 39 | GIN+FLAG | 0.2834
±
0.0038 | — | — | |
| 40 | MPNN + Noisy Nodes | 0.2832
±
0.0013 | — | — | |
| 41 | MPNN + VN + NoPE | 0.2823
±
0.0026 | — | — | |
| 42 | MPNN | 0.2791
±
0.0022 | — | — | |
| 43 | MPNN + Random Features + Noisy Nodes | 0.2790
±
0.0018 | — | — | |
| 44 | OAP+GatedGCN | 0.2783
±
0.0024 | — | — | |
| 45 | ESAN (GIN + ED) | 0.2782
±
0.0036 | — | — | |
| 46 | GN-GIN | 0.2781
±
0.0043 | — | — | |
| 47 | DeeperGCN | 0.2781
±
0.0038 | — | — | |
| 48 | HGODE | 0.2780
±
0.0030 | — | — | |
| 49 | SAN | 0.2765
±
0.0042 | — | — | |
| 50 | GraphTrans | 0.2761
±
0.0029 | — | — |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time