Motif-aware Attribute Masking for Molecular Graph Pre-training

Eric Inae, Gang Liu, Meng Jiang

Abstract

Attribute reconstruction is used to predict node or edge features in the pre-training of graph neural networks. Given a large number of molecules, they learn to capture structural knowledge, which is transferable for various downstream property prediction tasks and vital in chemistry, biomedicine, and material science. Previous strategies that randomly select nodes to do attribute masking leverage the information of local neighbors. However, the over-reliance of these neighbors inhibits the model's ability to learn long-range dependencies from higher-level substructures, such as functional groups or chemical motifs. To explicitly measure and encourage the inter-motif knowledge transfer in pre-trained models, we define inter-motif node influence measures and propose a novel motif-aware attribute masking strategy to capture long-range inter-motif structures by leveraging the information of atoms in neighboring motifs. Once each graph is decomposed into disjoint motifs, the features for every node within a sample motif are masked and subsequently predicted using a graph decoder. We evaluate our approach on eleven molecular classification and regression datasets and demonstrate its advantages.

Benchmark Results

Tested on 9 datasets 31 baseline values reported too low

Tables show presented results in global context (best 3 papers + GCN + MLP).

MoleculeNet

Tox21

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	HIMP Hyb global	0.8740 ±0.0050	Graph convolutions that can finally model local structure
2	GINE w/ VN GNN global	0.8720 ±0.0030	Graph convolutions that can finally model local structure
3	NaiveGINe+ K=3 w/ VN GNN global	0.8700 ±0.0040	Graph convolutions that can finally model local structure
14	GCN GNN global	0.8400 ±0.0040	Graph convolutions that can finally model local structure
15	GROVER GNN baseline actual #15 vs paper #175 160 places worse	0.8380 ±0.0170 best reported	actual #15 vs paper #175 160 places worse paper 0.7630 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7630 ±0.0060 Δ at pub 0.0750
31	JOAO GNN baseline actual #31 vs paper #238 207 places worse	0.8230 ±0.0220 best reported	actual #31 vs paper #238 207 places worse paper 0.7480 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7480 ±0.0060 Δ at pub 0.0750
46	ContextPred GNN baseline actual #46 vs paper #269 223 places worse	0.8060 ±0.0120 best reported	actual #46 vs paper #269 223 places worse paper 0.7360 ±0.0030 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7360 ±0.0030 Δ at pub 0.0700
72	Uni-Mol GNN baseline actual #72 vs paper #92 20 places worse	0.7960 ±0.0050 best reported	actual #72 vs paper #92 20 places worse paper 0.7890 ±0.0040 best today: A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the Foundation Model Era (2026) Difference within reported uncertainty; not counted as underperformance. paper 0.7890 ±0.0040 Δ today 0.0070
71	GraphLoG GNN baseline actual #71 vs paper #193 122 places worse	0.7960 ±0.0250 best reported	actual #71 vs paper #193 122 places worse paper 0.7570 ±0.0050 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7570 ±0.0050 Δ at pub 0.0390
86	No pretrain GNN baseline actual #86 vs paper #242 156 places worse	0.7920 ±0.0199 best reported	actual #86 vs paper #242 156 places worse paper 0.7460 ±0.0040 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.7460 ±0.0040 Δ at pub 0.0460
112	MoAMa Hyb ★ proposed actual #112 vs paper #170 58 places worse	0.7829 ±0.0055 best reported	actual #112 vs paper #170 58 places worse paper 0.7650 ±0.0010 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 0.7650 ±0.0010 Δ today 0.0179
150	Mole-BERT GNN baseline actual #150 vs paper #160 10 places worse	0.7710 ±0.0040 best reported	actual #150 vs paper #160 10 places worse paper 0.7680 ±0.0050 best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7680 ±0.0050 Δ today 0.0030
160	D-SLA Hyb baseline actual #160 vs paper #160	0.7681 ±0.0052 best reported	actual #160 vs paper #160 paper 0.7680 ±0.0050 best at pub: Graph Self-supervised Learning with Accurate Discrepancy Learning (NeurIPS 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7680 ±0.0050 Δ at pub 0.0001
172	MGSSL GNN baseline actual #172 vs paper #219 47 places worse	0.7650 ±0.0030 best reported	actual #172 vs paper #219 47 places worse paper 0.7520 ±0.0060 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) paper 0.7520 ±0.0060 Δ at pub 0.0130
202	GraphMAE GNN baseline	0.7550 ±0.0060	This paper
229	AttrMask GNN baseline actual #229 vs paper #229	0.7510 ±0.0090	actual #229 vs paper #229 paper 0.7510 ±0.0090 prev best: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) Improvement within reported uncertainty; not highlighted. paper 0.7510 ±0.0090 Δ at pub 0.0030
251	MCM GNN baseline	0.7440 ±0.0010	This paper

BACE

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	ProtoMol Hyb global	0.9140 ±0.0030	ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning
2	MMSG GNN global	0.9080 ±0.0050	ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning
3	MemGNN code ↗ Hyb global	0.9070	Memory-Based Graph Networks code ↗
22	No pretrain GNN baseline actual #22 vs paper #263 241 places worse	0.8743 ±0.0163 best reported	actual #22 vs paper #263 241 places worse paper 0.7240 ±0.0380 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.7240 ±0.0380 Δ at pub 0.1503
43	GIN code ↗ GNN global	0.8646 ±0.0081	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
61	Uni-Mol GNN baseline actual #61 vs paper #214 153 places worse	0.8570 ±0.0020 best reported	actual #61 vs paper #214 153 places worse paper 0.7810 ±0.0140 best today: Invariant-Based Diagnostics for Graph Benchmarks (2026) paper 0.7810 ±0.0140 Δ today 0.0760
94	D-SLA Hyb baseline	0.8380 ±0.0100	This paper
98	GraphLoG GNN baseline actual #98 vs paper #98	0.8350 ±0.0120	actual #98 vs paper #98 paper 0.8350 ±0.0120 prev best: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) Improvement within reported uncertainty; not highlighted. paper 0.8350 ±0.0120 Δ at pub 0.0090
100	JOAO GNN baseline actual #100 vs paper #260 160 places worse	0.8340 ±0.0330 best reported	actual #100 vs paper #260 160 places worse paper 0.7320 ±0.0160 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7320 ±0.0160 Δ at pub 0.1020
105	GraphMAE GNN baseline	0.8310 ±0.0090	This paper
118	GROVER GNN baseline actual #118 vs paper #187 69 places worse	0.8260 ±0.0070 best reported	actual #118 vs paper #187 69 places worse paper 0.7950 ±0.0110 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) paper 0.7950 ±0.0110 Δ today 0.0310
130	MGSSL GNN baseline actual #130 vs paper #206 76 places worse	0.8203 ±0.0379 best reported	actual #130 vs paper #206 76 places worse paper 0.7880 ±0.0090 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7880 ±0.0090 Δ at pub 0.0090 Δ today 0.0323
143	Mole-BERT GNN baseline actual #143 vs paper #164 21 places worse	0.8140 ±0.0100 best reported	actual #143 vs paper #164 21 places worse paper 0.8080 ±0.0140 best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.8080 ±0.0140 Δ today 0.0060
147	MoAMa Hyb ★ proposed actual #147 vs paper #173 26 places worse	0.8132 ±0.0106 best reported	actual #147 vs paper #173 26 places worse paper 0.8010 ±0.0050 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8010 ±0.0050 Δ today 0.0122
182	AttrMask GNN baseline actual #182 vs paper #223 41 places worse	0.7970 ±0.0033 best reported	actual #182 vs paper #223 41 places worse paper 0.7780 ±0.0180 best at pub: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7780 ±0.0180 Δ at pub 0.0190 Δ today 0.0190
186	ContextPred GNN baseline actual #186 vs paper #206 20 places worse	0.7960 ±0.0120 best reported	actual #186 vs paper #206 20 places worse paper 0.7880 ±0.0120 best at pub: Graph Self-Contrast Representation Learning (2023) best today: Dual-perspective Cross Contrastive Learning in Graph Transformers (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7880 ±0.0120 Δ at pub 0.0080 Δ today 0.0080
188	MCM GNN baseline	0.7950 ±0.0130	This paper

BBBP

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	MAGPrompt+ Hyb global	0.9432 ±0.0238	MAGPrompt: Message-Adaptive Graph Prompt Tuning for Graph Neural Networks
2	CamS-LLaMA LLM 100.0M global	0.9420	Multi-scale Graph Autoregressive Modeling: Molecular Property Prediction via Next Token …
3	SE(3)-I MPPH Hyb global	0.9400 ±0.0210	SE(3)-Invariant Multiparameter Persistent Homology for Chiral-Sensitive Molecular Property …
18	No pretrain GNN baseline actual #18 vs paper #259 241 places worse	0.9223 ±0.0307 best reported	actual #18 vs paper #259 241 places worse paper 0.6570 ±0.0330 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.6570 ±0.0330 Δ at pub 0.2653
34	GIN code ↗ GNN global	0.8875 ±0.0049	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
37	GROVER GNN baseline actual #37 vs paper #201 164 places worse	0.8870 ±0.0060 best reported	actual #37 vs paper #201 164 places worse paper 0.6800 ±0.0150 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6800 ±0.0150 Δ at pub 0.2070
39	JOAO GNN baseline actual #39 vs paper #245 206 places worse	0.8790 ±0.0200 best reported	actual #39 vs paper #245 206 places worse paper 0.6640 ±0.0100 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6640 ±0.0100 Δ at pub 0.2150
40	ContextPred GNN baseline actual #40 vs paper #127 87 places worse	0.8770 ±0.0260 best reported	actual #40 vs paper #127 87 places worse paper 0.7060 ±0.0150 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7060 ±0.0150 Δ at pub 0.1710
51	MoAMa Hyb ★ proposed actual #51 vs paper #86 35 places worse	0.8589 ±0.0061 best reported	actual #51 vs paper #86 35 places worse paper 0.7280 ±0.0090 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 0.7280 ±0.0090 Δ today 0.1309
54	GraphLoG GNN baseline actual #54 vs paper #90 36 places worse	0.8460 ±0.0080 best reported	actual #54 vs paper #90 36 places worse paper 0.7250 ±0.0080 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7250 ±0.0080 Δ at pub 0.1210
61	MGSSL GNN baseline actual #61 vs paper #183 122 places worse	0.7952 ±0.0198 best reported	actual #61 vs paper #183 122 places worse paper 0.6880 ±0.0090 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 0.6880 ±0.0090 Δ at pub 0.0170 Δ today 0.1072
84	Uni-Mol GNN baseline actual #84 vs paper #267 183 places worse	0.7290 ±0.0060 best reported	actual #84 vs paper #267 183 places worse paper 0.6540 ±0.0100 best today: Invariant-Based Diagnostics for Graph Benchmarks (2026) paper 0.6540 ±0.0100 Δ today 0.0750
89	D-SLA Hyb baseline	0.7260 ±0.0080	This paper
96	Mole-BERT GNN baseline actual #96 vs paper #106 10 places worse	0.7230 ±0.0070 best reported	actual #96 vs paper #106 10 places worse paper 0.7190 ±0.0160 best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7190 ±0.0160 Δ today 0.0040
101	GraphMAE GNN baseline	0.7200 ±0.0060	This paper
113	MCM GNN baseline	0.7160 ±0.0060	This paper
141	AttrMask GNN baseline actual #141 vs paper #268 127 places worse	0.7020 ±0.0050 best reported	actual #141 vs paper #268 127 places worse paper 0.6520 ±0.0140 best at pub: 16 Pre-training Molecular Graph Representation with 3D Geometry (ICLR 2021) best today: Graph Generative Pre-trained Transformer (ICML 2025) paper 0.6520 ±0.0140 Δ at pub 0.0500 Δ today 0.0500

ClinTox

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	SE(3)-I MPPH Hyb global	0.9930 ±0.0040	SE(3)-Invariant Multiparameter Persistent Homology for Chiral-Sensitive Molecular Property …
2	KA-GCN code ↗ GNN global	0.9920 ±0.0050	A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the … code ↗
3	KA-GAT code ↗ GNN global	0.9910 ±0.0050	A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the … code ↗
17	Uni-Mol GNN baseline actual #17 vs paper #79 62 places worse	0.9190 ±0.0180 best reported	actual #17 vs paper #79 62 places worse paper 0.8120 ±0.0170 best today: A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the Foundation Model Era (2026) paper 0.8120 ±0.0170 Δ today 0.1070
19	GCN GNN global	0.9130 ±0.0170	OOD-GNN: Out-of-Distribution Generalized Graph Neural Network
28	No pretrain GNN baseline actual #28 vs paper #227 199 places worse	0.8970 ±0.0093 best reported	actual #28 vs paper #227 199 places worse paper 0.5840 ±0.0640 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.5840 ±0.0640 Δ at pub 0.3130
40	AttrMask GNN baseline actual #40 vs paper #145 105 places worse	0.8770 ±0.0119 best reported	actual #40 vs paper #145 105 places worse paper 0.7350 ±0.0430 best at pub: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) paper 0.7350 ±0.0430 Δ at pub 0.1420 Δ today 0.1420
56	MoAMa Hyb ★ proposed	0.8530 ±0.0220	This paper
71	GraphMAE GNN baseline	0.8230 ±0.0120	This paper
76	GROVER GNN baseline actual #76 vs paper #124 48 places worse	0.8170 ±0.0250 best reported	actual #76 vs paper #124 48 places worse paper 0.7540 ±0.0860 best at pub: Unified 2D and 3D Pre-Training of Molecular Representations (KDD 2022) best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7540 ±0.0860 Δ at pub 0.0240 Δ today 0.0630
79	JOAO GNN baseline actual #79 vs paper #195 116 places worse	0.8132 ±0.0249 best reported	actual #79 vs paper #195 116 places worse paper 0.6660 ±0.0310 best at pub: Graph Self-Contrast Representation Learning (2023) paper 0.6660 ±0.0310 Δ at pub 0.1472
85	MGSSL GNN baseline actual #85 vs paper #115 30 places worse	0.8070 ±0.0210 best reported	actual #85 vs paper #115 30 places worse paper 0.7710 ±0.0450 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7710 ±0.0450 Δ at pub 0.0360
87	D-SLA Hyb baseline actual #87 vs paper #87	0.8020 ±0.0150	actual #87 vs paper #87 paper 0.8020 ±0.0150 prev best: Graph Self-supervised Learning with Accurate Discrepancy Learning (NeurIPS 2022) Improvement within reported uncertainty; not highlighted. paper 0.8020 ±0.0150 Δ at pub 0.0003
88	Mole-BERT GNN baseline actual #88 vs paper #96 8 places worse	0.8010 ±0.0360 best reported	actual #88 vs paper #96 8 places worse paper 0.7890 ±0.0300 best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7890 ±0.0300 Δ today 0.0120
103	GraphLoG GNN baseline actual #103 vs paper #116 13 places worse	0.7872 ±0.0258 best reported	actual #103 vs paper #116 13 places worse paper 0.7670 ±0.0330 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Graph Self-Contrast Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7670 ±0.0330 Δ at pub 0.0202 Δ today 0.0202
115	ContextPred GNN baseline actual #115 vs paper #139 24 places worse	0.7720 ±0.0088 best reported	actual #115 vs paper #139 24 places worse paper 0.7400 ±0.0340 best at pub: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7400 ±0.0340 Δ at pub 0.0320 Δ today 0.0320
204	MCM GNN baseline	0.6470 ±0.0050	This paper

SIDER

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	KA-GAT code ↗ GNN global	0.8470 ±0.0020	A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the … code ↗
2	KA-GCN code ↗ GNN global	0.8420 ±0.0010	A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the … code ↗
3	GraphKAN code ↗ GNN global	0.8370 ±0.0010	A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the … code ↗
37	Uni-Mol GNN baseline actual #37 vs paper #216 179 places worse	0.6590 ±0.0130 best reported	actual #37 vs paper #216 179 places worse paper 0.5770 ±0.0030 best today: A Systematic Survey and Benchmark of Deep Learning for Molecular Property Prediction in the Foundation Model Era (2026) paper 0.5770 ±0.0030 Δ today 0.0820
52	GROVER GNN baseline actual #52 vs paper #223 171 places worse	0.6480 ±0.0060 best reported	actual #52 vs paper #223 171 places worse paper 0.5710 ±0.0160 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) paper 0.5710 ±0.0160 Δ at pub 0.0670 Δ today 0.0770
58	MoAMa Hyb ★ proposed	0.6460 ±0.0050	This paper
66	GIN code ↗ GNN global	0.6403 ±0.0104	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
68	JOAO GNN baseline actual #68 vs paper #169 101 places worse	0.6400 ±0.0100 best reported	actual #68 vs paper #169 101 places worse paper 0.6040 ±0.0150 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6040 ±0.0150 Δ at pub 0.0360
90	Mole-BERT GNN baseline	0.6280 ±0.0110	This paper
102	MCM GNN baseline	0.6230 ±0.0090	This paper
101	No pretrain GNN baseline actual #101 vs paper #212 111 places worse	0.6230 ±0.0114 best reported	actual #101 vs paper #212 111 places worse paper 0.5820 ±0.0170 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.5820 ±0.0170 Δ at pub 0.0410
103	ContextPred GNN baseline actual #103 vs paper #192 89 places worse	0.6220 ±0.0059 best reported	actual #103 vs paper #192 89 places worse paper 0.5970 ±0.0180 best at pub: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.5970 ±0.0180 Δ at pub 0.0250 Δ today 0.0250
116	MGSSL GNN baseline actual #116 vs paper #119 3 places worse	0.6180 ±0.0080 best reported	actual #116 vs paper #119 3 places worse paper 0.6160 ±0.0100 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.6160 ±0.0100 Δ at pub 0.0020
123	GraphLoG GNN baseline actual #123 vs paper #133 10 places worse	0.6150 ±0.0130 best reported	actual #123 vs paper #133 10 places worse paper 0.6120 ±0.0110 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.6120 ±0.0110 Δ at pub 0.0030
134	AttrMask GNN baseline actual #134 vs paper #162 28 places worse	0.6120 ±0.0012 best reported	actual #134 vs paper #162 28 places worse paper 0.6050 ±0.0090 best at pub: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) best today: A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.6050 ±0.0090 Δ at pub 0.0070 Δ today 0.0070
175	GraphMAE GNN baseline	0.6030 ±0.0110	This paper
176	D-SLA Hyb baseline actual #176 vs paper #176	0.6022 ±0.0113 best reported	actual #176 vs paper #176 paper 0.6020 ±0.0110 best at pub: Graph Self-supervised Learning with Accurate Discrepancy Learning (NeurIPS 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.6020 ±0.0110 Δ at pub 0.0002

ToxCast

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	GeomGCL GNN global	0.7630	GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction
2	LiGhT DeepGPT ext data Hyb 0.4M global	0.7570 ±0.0110	Deep Prompt Tuning for Graph Transformers
3	CoMPT GT global	0.7550	GeomGCL: Geometric Graph Contrastive Learning for Molecular Property Prediction
14	GCN GNN global	0.7350 ±0.0020	Graph convolutions that can finally model local structure
28	ContextPred* GNN baseline actual #28 vs paper #171 143 places worse	0.7140 ±0.0180 best reported	actual #28 vs paper #171 143 places worse paper 0.6260 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6260 ±0.0060 Δ at pub 0.0880
31	JOAO* GNN baseline actual #31 vs paper #167 136 places worse	0.7110 ±0.0140 best reported	actual #31 vs paper #167 136 places worse paper 0.6280 ±0.0070 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6280 ±0.0070 Δ at pub 0.0830
34	GROVER GNN baseline actual #34 vs paper #141 107 places worse	0.7080 ±0.0150 best reported	actual #34 vs paper #141 107 places worse paper 0.6340 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6340 ±0.0060 Δ at pub 0.0740
46	No Pretrain* GNN baseline actual #46 vs paper #185 139 places worse	0.6913 ±0.0055 best reported	actual #46 vs paper #185 139 places worse paper 0.6170 ±0.0050 best at pub: Does GNN Pretraining Help Molecular Representation? (NeurIPS 2022) paper 0.6170 ±0.0050 Δ at pub 0.0743
58	MoAMa Hyb ★ proposed actual #58 vs paper #162 104 places worse	0.6801 ±0.0107 best reported	actual #58 vs paper #162 104 places worse paper 0.6300 ±0.0040 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 0.6300 ±0.0040 Δ today 0.0501
64	GraphLoG* GNN baseline actual #64 vs paper #137 73 places worse	0.6770 ±0.0190 best reported	actual #64 vs paper #137 73 places worse paper 0.6350 ±0.0070 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6350 ±0.0070 Δ at pub 0.0420
114	Mole-BERT* LLM baseline actual #114 vs paper #121 7 places worse	0.6450 ±0.0040 best reported	actual #114 vs paper #121 7 places worse paper 0.6430 ±0.0020 best today: UniMAP: Universal SMILES-Graph Representation Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.6430 ±0.0020 Δ today 0.0020
124	D-SLA* GNN baseline actual #124 vs paper #125 1 place worse	0.6424 ±0.0050 best reported	actual #124 vs paper #125 1 place worse paper 0.6420 ±0.0050 best at pub: Graph Self-supervised Learning with Accurate Discrepancy Learning (NeurIPS 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.6420 ±0.0050 Δ at pub 0.0004
129	MGSSL* GNN baseline actual #129 vs paper #149 20 places worse	0.6410 ±0.0070 best reported	actual #129 vs paper #149 20 places worse paper 0.6330 ±0.0050 best at pub: Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.6330 ±0.0050 Δ at pub 0.0080
128	GraphMAE* GNN baseline	0.6410 ±0.0030	This paper
150	AttrMask* GNN baseline actual #150 vs paper #150	0.6330 ±0.0060	actual #150 vs paper #150 paper 0.6330 ±0.0060 prev best: A Group Symmetric Stochastic Differential Equation Model for Molecule Multi-modal Pretraining (ICML 2023) Improvement within reported uncertainty; not highlighted. paper 0.6330 ±0.0060 Δ at pub 0.0040
170	Uni-Mol GNN baseline	0.6270 ±0.0040	This paper
194	MCM GNN baseline	0.6100 ±0.0040	This paper

ESOL

Global leaderboard →

#	Model	RMSE ↓	Paper
1	Graphormer-SPIS code ↗ GT global	0.4840 ±0.0050	On Structural Expressive Power of Graph Transformers code ↗
2	Graphormer-SPD code ↗ GT global	0.4920 ±0.0040	On Structural Expressive Power of Graph Transformers code ↗
3	conan-FGW global	0.5140 ±0.0190	Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks
7	GCN code ↗ GNN global	0.5640 ±0.0150	On Structural Expressive Power of Graph Transformers code ↗
103	AttrMask ext data GNN baseline actual #103 vs paper #127 24 places worse	1.1120 ±0.0480 best reported	actual #103 vs paper #127 24 places worse paper 1.4860 ±0.1200 best at pub: Improving Molecular Pretraining with Complementary Featurizations (2022) paper 1.4860 ±0.1200 Δ at pub 0.3740
105	JOAO ext data GNN baseline actual #105 vs paper #136 31 places worse	1.1200 best reported	actual #105 vs paper #136 31 places worse paper 1.6400 ±0.0900 best at pub: 16 Pre-training Molecular Graph Representation with 3D Geometry (ICLR 2021) best today: 3D Molecular Geometry Analysis with 2D Graphs (2023) paper 1.6400 ±0.0900 Δ at pub 0.5200 Δ today 0.5200
107	MoAMa Hyb ★ proposed actual #107 vs paper #124 17 places worse	1.1250 ±0.0290 best reported	actual #107 vs paper #124 17 places worse paper 1.3940 ±0.1200 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 1.3940 ±0.1200 Δ today 0.2690
113	MGSSL ext data GNN baseline actual #113 vs paper #126 13 places worse	1.1790 ±0.0080 best reported	actual #113 vs paper #126 13 places worse paper 1.4380 ±0.1200 best today: Atom-Motif Contrastive Transformer for Molecular Property Prediction (2023) paper 1.4380 ±0.1200 Δ today 0.2590
114	ContextPred ext data GNN baseline actual #114 vs paper #135 21 places worse	1.1960 ±0.0370 best reported	actual #114 vs paper #135 21 places worse paper 1.5970 ±0.1400 best at pub: Improving Molecular Pretraining with Complementary Featurizations (2022) paper 1.5970 ±0.1400 Δ at pub 0.4010
130	GraphLoG ext data GNN baseline	1.5070 ±0.1300	This paper
131	GraphMAE ext data GNN baseline	1.5100 ±0.0900	This paper
134	D-SLA ext data Hyb baseline	1.5380 ±0.1300	This paper
135	MCM ext data GNN baseline	1.5530 ±0.1100	This paper
136	No pretrain GNN baseline	1.6050 ±0.0800	This paper
137	Uni-Mol ext data GNN baseline	1.7170 ±0.0800	This paper
139	Mole-BERT ext data GNN baseline	1.7420 ±0.0500	This paper
152	GROVER GNN baseline	3.1070 ±1.3900	This paper

FreeSolv

Global leaderboard →

#	Model	RMSE ↓	Paper
1	SE(3)-I MPPH Hyb global	0.3540 ±0.0530	SE(3)-Invariant Multiparameter Persistent Homology for Chiral-Sensitive Molecular Property …
2	ECFP-4 + RF RF global	0.5600 ±0.0660	SE(3)-Invariant Multiparameter Persistent Homology for Chiral-Sensitive Molecular Property …
3	GPSE + GPS code ↗ ext data Hyb global	0.6820	Graph Positional and Structural Encoder code ↗
54	MoAMa Hyb ★ proposed actual #54 vs paper #80 26 places worse	2.0720 ±0.0530 best reported	actual #54 vs paper #80 26 places worse paper 2.7000 ±0.1100 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 2.7000 ±0.1100 Δ today 0.6280
67	GraphLoG GNN baseline actual #67 vs paper #107 40 places worse	2.3350 ±0.0520 best reported	actual #67 vs paper #107 40 places worse paper 3.4670 ±0.0600 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 3.4670 ±0.0600 Δ today 1.1320
69	GIN code ↗ GNN global	2.3460 ±0.1220	ChemRL-GEM: Geometry Enhanced Molecular Representation Learning for Property Prediction code ↗
83	GraphMAE GNN baseline actual #83 vs paper #98 15 places worse	2.7400 ±0.2330 best reported	actual #83 vs paper #98 15 places worse paper 3.0600 ±0.1500 best today: Frequency-Corrupt Based Graph Self-Supervised Learning (WWW 2026) Difference within reported uncertainty; not counted as underperformance. paper 3.0600 ±0.1500 Δ today 0.3200
87	Uni-Mol GNN baseline	2.8110 ±0.1600	This paper
90	MGSSL GNN baseline	2.8400 ±0.1900	This paper
93	D-SLA Hyb baseline	2.9200 ±0.0500	This paper
94	AttrMask GNN baseline	2.9340 ±0.1300	This paper
97	MCM GNN baseline	3.0510 ±0.1100	This paper
98	ContextPred GNN baseline	3.0570 ±0.1700	This paper
104	JOAO GNN baseline actual #104 vs paper #107 3 places worse	3.2820 ±0.0020 best reported	actual #104 vs paper #107 3 places worse paper 3.4670 ±0.0600 best today: Hierarchical Molecular Representation Learning via Fragment-Based Self-Supervised Embedding Prediction (2026) paper 3.4670 ±0.0600 Δ today 0.1850
109	No pretrain GNN baseline	3.5890 ±0.3300	This paper
115	Mole-BERT GNN baseline	4.0290 ±0.0800	This paper
126	GROVER GNN baseline	7.9390 ±3.1700	This paper

Lipophilicity

Global leaderboard →

#	Model	RMSE ↓	Paper
1	MemGNN code ↗ Hyb global	0.5560 ±0.0230	Memory-Based Graph Networks code ↗
2	ProtoS-L2 Hyb 0.1M global	0.5800 ±0.0160	Optimal Transport Graph Neural Networks
3	ProtoW-L2 Hyb global	0.6040 ±0.0140	Optimal Transport Graph Neural Networks
7	GCN code ↗ GNN global	0.6550 ±0.0360	Memory-Based Graph Networks code ↗
17	ContextPred GNN baseline actual #17 vs paper #64 47 places worse	0.7020 ±0.0200 best reported	actual #17 vs paper #64 47 places worse paper 0.8350 ±0.0100 best at pub: Improving Molecular Pretraining with Complementary Featurizations (2022) paper 0.8350 ±0.0100 Δ at pub 0.1330
21	JOAO GNN baseline actual #21 vs paper #91 70 places worse	0.7080 best reported	actual #21 vs paper #91 70 places worse paper 1.0400 ±0.0500 best at pub: Improving Molecular Pretraining with Complementary Featurizations (2022) best today: 3D Molecular Geometry Analysis with 2D Graphs (2023) paper 1.0400 ±0.0500 Δ at pub 0.3320 Δ today 0.3320
31	AttrMask GNN baseline actual #31 vs paper #60 29 places worse	0.7300 ±0.0040 best reported	actual #31 vs paper #60 29 places worse paper 0.8170 ±0.0100 best at pub: Improving Molecular Pretraining with Complementary Featurizations (2022) paper 0.8170 ±0.0100 Δ at pub 0.0870
48	MoAMa Hyb ★ proposed	0.7720 ±0.0300	This paper
54	D-SLA Hyb baseline	0.7900 ±0.0900	This paper
60	MCM GNN baseline	0.8150 ±0.1100	This paper
62	Mole-BERT GNN baseline	0.8210 ±0.0200	This paper
64	No pretrain GNN baseline	0.8310 ±0.0200	This paper
70	GraphMAE GNN baseline	0.8520 ±0.0200	This paper
73	MGSSL GNN baseline	0.8690 ±0.0600	This paper
82	GraphLoG GNN baseline actual #82 vs paper #91 9 places worse	0.9320 ±0.0520 best reported	actual #82 vs paper #91 9 places worse paper 1.0400 ±0.0500 best today: Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2024) paper 1.0400 ±0.0500 Δ today 0.1080
96	Uni-Mol GNN baseline	1.1320 ±0.0500	This paper
103	GROVER GNN baseline	1.3770 ±0.0600	This paper

SOTA performance (not significantly different from the best) Proposed method (this paper) Baseline / comparison (this paper) Global context (other papers) Best at pub — reported result worse than the best known when published (best reported in green) Best today — OK at publication; today's leaderboard best is elsewhere