Graph Self-Contrast Representation Learning

Minjie Chen, Yao Cheng, Ye Wang, Xiangci Li, Ming Gao

Abstract

Graph contrastive learning (GCL) has recently emerged as a promising approach for graph representation learning. Some existing methods adopt the 1-vs-$K$ scheme to construct one positive and $K$ negative samples for each graph, but it is difficult to set $K$. For those methods that do not use negative samples, it is often necessary to add additional strategies to avoid model collapse, which could only alleviate the problem to some extent. All these drawbacks will undoubtedly have an adverse impact on the generalizability and efficiency of the model. In this paper, to address these issues, we propose a novel graph self-contrast framework GraphSC, which only uses one positive and one negative sample, and chooses triplet loss as the objective. Specifically, self-contrast has two implications. First, GraphSC generates both positive and negative views of a graph sample from the graph itself via graph augmentation functions of various intensities, and use them for self-contrast. Second, GraphSC uses Hilbert-Schmidt Independence Criterion (HSIC) to factorize the representations into multiple factors and proposes a masked self-contrast mechanism to better separate positive and negative samples. Further, Since the triplet loss only optimizes the relative distance between the anchor and its positive/negative samples, it is difficult to ensure the absolute distance between the anchor and positive sample. Therefore, we explicitly reduced the absolute distance between the anchor and positive sample to accelerate convergence. Finally, we conduct extensive experiments to evaluate the performance of GraphSC against 19 other state-of-the-art methods in both unsupervised and transfer learning settings.

Benchmark Results

Tested on 10 datasets 38 baseline values reported too low

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

TU Dortmund

MUTAG

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	MSH-GNN GNN global	0.9910 ±0.0030	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
2	ID-GNN+BEC Hyb global	0.9830 ±0.0360	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
3	SupCosine Hyb global	0.9830 ±0.0250	Supervised Contrastive Learning with Structure Inference for Graph Classification
39	GIN GNN global	0.9330 ±0.0290	ProtGNN: Towards Self-Explaining Graph Neural Networks
159	WL Trad baseline actual #159 vs paper #596 437 places worse	0.9040 ±0.0570 best reported	actual #159 vs paper #596 437 places worse paper 0.8072 ±0.0300 best at pub: How Powerful are $K$-hop Message Passing Graph Neural Networks (NeurIPS 2022) paper 0.8072 ±0.0300 Δ at pub 0.0968
178	AD-GCL GNN baseline actual #178 vs paper #294 116 places worse	0.9028 ±0.0401 best reported	actual #178 vs paper #294 116 places worse paper 0.8874 ±0.0185 best at pub: Graph Self-Supervised Learning: A Survey (2021) best today: Unifying Graph Contrastive Learning via Graph Message Augmentation (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8874 ±0.0185 Δ at pub 0.0096 Δ today 0.0154
180	LaGraph GNN baseline	0.9020 ±0.0110	This paper
222	JOAOv2 GNN baseline actual #222 vs paper #353 131 places worse	0.8967 ±0.0524 best reported	actual #222 vs paper #353 131 places worse paper 0.8767 ±0.0079 best at pub: Graph Contrastive Learning Automated (ICML 2021) best today: Unifying Graph Contrastive Learning via Graph Message Augmentation (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8767 ±0.0079 Δ today 0.0200
257	InfoGraph GNN baseline actual #257 vs paper #272 15 places worse	0.8930 ±0.0514 best reported	actual #257 vs paper #272 15 places worse paper 0.8901 ±0.0113 best at pub: Improving Expressivity of GNNs with Subgraph-specific Factor Embedded Normalization (KDD 2023) Difference within reported uncertainty; not counted as underperformance. paper 0.8901 ±0.0113 Δ at pub 0.0029
269	GraphSC GNN ★ proposed	0.8919 ±0.0183	This paper
272	SimGRACE GNN baseline actual #272 vs paper #272	0.8910 ±0.0140 best reported	actual #272 vs paper #272 paper 0.8901 ±0.0131 best at pub: SimGRACE: A Simple Framework for Graph Contrastive Learning without Data Augmentation (WWW 2022) best today: GPS: graph contrastive learning via multi-scale augmented views from adversarial pooling (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8901 ±0.0131 Δ today 0.0009
330	GraphSC-MSE GNN ★ proposed	0.8804 ±0.0156	This paper
351	JOAO GNN baseline actual #351 vs paper #370 19 places worse	0.8770 ±0.0080 best reported	actual #351 vs paper #370 19 places worse paper 0.8735 ±0.0102 best at pub: Graph Self-Supervised Learning: A Survey (2021) best today: Learning Graph Augmentations to Learn Graph Representations (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.8735 ±0.0102 Δ at pub 0.0035 Δ today 0.0035
354	GraphCL GNN baseline actual #354 vs paper #396 42 places worse	0.8766 ±0.0103 best reported	actual #354 vs paper #396 42 places worse paper 0.8680 ±0.0134 best at pub: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.8680 ±0.0134 Δ at pub 0.0086
355	RGCL GNN baseline	0.8766 ±0.0101	This paper
365	DGK SVM baseline	0.8744 ±0.0272	This paper
403	MLP code ↗ Trad global	0.8660 ±0.0495	Breaking the Limits of Message Passing Graph Neural Networks code ↗
407	graph2vec Walk baseline actual #407 vs paper #517 110 places worse	0.8644 ±0.0125 best reported	actual #407 vs paper #517 110 places worse paper 0.8315 ±0.0925 best at pub: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.8315 ±0.0925 Δ at pub 0.0329
444	GL Trad baseline actual #444 vs paper #577 133 places worse	0.8520 ±0.0090 best reported	actual #444 vs paper #577 133 places worse paper 0.8166 ±0.0211 best at pub: On Valid Optimal Assignment Kernels and Applications to Graph Classification (NeurIPS 2016) paper 0.8166 ±0.0211 Δ at pub 0.0354
472	Sub2Vec Walk baseline actual #472 vs paper #786 314 places worse	0.8447 ±0.0037 best reported	actual #472 vs paper #786 314 places worse paper 0.6105 ±0.1580 best at pub: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.6105 ±0.1580 Δ at pub 0.2342

NCI1

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	WL-MLP GT global	0.9954	1-WL Expressiveness Is (Almost) All You Need
2	GIN GNN global	0.9934	1-WL Expressiveness Is (Almost) All You Need
3	GCN GNN global	0.9727	1-WL Expressiveness Is (Almost) All You Need
12	WL Trad baseline actual #12 vs paper #223 211 places worse	0.8590 ±0.0010 best reported	actual #12 vs paper #223 211 places worse paper 0.8001 ±0.0050 best at pub: Propagation Kernels (2014) paper 0.8001 ±0.0050 Δ at pub 0.0589
104	GraphCL GNN baseline actual #104 vs paper #325 221 places worse	0.8250 ±0.0013 best reported	actual #104 vs paper #325 221 places worse paper 0.7787 ±0.0041 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.7787 ±0.0041 Δ at pub 0.0463
131	GraphSAGE code ↗ GNN global	0.8200 ±0.0100	Weisfeiler and Leman Go Infinite: Spectral and Combinatorial Pre-Colorings code ↗
141	SimGRACE GNN baseline actual #141 vs paper #258 117 places worse	0.8180 ±0.0015 best reported	actual #141 vs paper #258 117 places worse paper 0.7912 ±0.0044 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.7912 ±0.0044 Δ at pub 0.0268
146	JOAOv2 GNN baseline actual #146 vs paper #303 157 places worse	0.8170 ±0.0026 best reported	actual #146 vs paper #303 157 places worse paper 0.7836 ±0.0053 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.7836 ±0.0053 Δ at pub 0.0334
176	GraphSC GNN ★ proposed	0.8112 ±0.0040	This paper
207	GraphSC-MSE GNN ★ proposed	0.8039 ±0.0062	This paper
210	DGK SVM baseline	0.8031 ±0.0046	This paper
227	LaGraph GNN baseline	0.7990 ±0.0050	This paper
301	JOAO GNN baseline actual #301 vs paper #317 16 places worse	0.7840 ±0.0050 best reported	actual #301 vs paper #317 16 places worse paper 0.7807 ±0.0047 best at pub: Omni-Granular Ego-Semantic Propagation for Self-Supervised Graph Representation Learning (ICML 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7807 ±0.0047 Δ at pub 0.0033
315	RGCL GNN baseline	0.7814 ±0.0108	This paper
338	InfoGraph GNN baseline actual #338 vs paper #386 48 places worse	0.7750 ±0.0074 best reported	actual #338 vs paper #386 48 places worse paper 0.7620 ±0.0106 best at pub: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7620 ±0.0106 Δ at pub 0.0130
406	graph2vec Walk baseline actual #406 vs paper #446 40 places worse	0.7540 ±0.0120 best reported	actual #406 vs paper #446 40 places worse paper 0.7322 ±0.0181 best at pub: InfoGCL: Information-Aware Graph Contrastive Learning (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7322 ±0.0181 Δ at pub 0.0218
409	AD-GCL GNN baseline actual #409 vs paper #439 30 places worse	0.7518 ±0.0031 best reported	actual #409 vs paper #439 30 places worse paper 0.7391 ±0.0077 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) best today: GIMM: InfoMin-Max for Automated Graph Contrastive Learning (2023) paper 0.7391 ±0.0077 Δ at pub 0.0127 Δ today 0.0127
557	Sub2Vec Walk baseline actual #557 vs paper #575 18 places worse	0.5976 ±0.0035 best reported	actual #557 vs paper #575 18 places worse paper 0.5284 ±0.0147 best at pub: Graph Masked Autoencoders with Transformers (2022) paper 0.5284 ±0.0147 Δ at pub 0.0692

COLLAB

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	WL-MLP GT global	0.9792	1-WL Expressiveness Is (Almost) All You Need
2	MSH-GNN GNN global	0.9640 ±0.0070	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
3	CoCN code ↗ GNN global	0.8722 ±0.0013	Scalable Graph Compressed Convolutions code ↗
23	GCN code ↗ GNN global	0.8260 ±0.0220	Scalable Expressiveness through Preprocessed Graph Perturbations code ↗
32	JOAOv2 GNN baseline actual #32 vs paper #305 273 places worse	0.8216 ±0.0017 best reported	actual #32 vs paper #305 273 places worse paper 0.6933 ±0.0034 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.6933 ±0.0034 Δ at pub 0.1283
37	SimGRACE GNN baseline actual #37 vs paper #286 249 places worse	0.8188 ±0.0023 best reported	actual #37 vs paper #286 249 places worse paper 0.7172 ±0.0082 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.7172 ±0.0082 Δ at pub 0.1016
80	GraphCL GNN baseline actual #80 vs paper #290 210 places worse	0.8108 ±0.0017 best reported	actual #80 vs paper #290 210 places worse paper 0.7136 ±0.0115 best at pub: Label-invariant Augmentation for Semi-Supervised Graph Classification (NeurIPS 2022) paper 0.7136 ±0.0115 Δ at pub 0.0972
146	InfoGraph GNN baseline actual #146 vs paper #295 149 places worse	0.7940 best reported	actual #146 vs paper #295 149 places worse paper 0.7065 ±0.0113 best at pub: CogDL: A Comprehensive Library for Graph Deep Learning (WWW 2021) paper 0.7065 ±0.0113 Δ at pub 0.0875
151	MLP Trad global	0.7920 ±0.0240	Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors
162	GraphSC GNN ★ proposed	0.7890 ±0.0032	This paper
188	GraphSC-MSE GNN ★ proposed	0.7764 ±0.0026	This paper
189	LaGraph GNN baseline	0.7760 ±0.0020	This paper
233	AD-GCL GNN baseline actual #233 vs paper #284 51 places worse	0.7582 ±0.0026 best reported	actual #233 vs paper #284 51 places worse paper 0.7202 ±0.0056 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) best today: Unifying Graph Contrastive Learning via Graph Message Augmentation (2024) paper 0.7202 ±0.0056 Δ at pub 0.0380 Δ today 0.0380
248	JOAO GNN baseline actual #248 vs paper #305 57 places worse	0.7535 ±0.0093 best reported	actual #248 vs paper #305 57 places worse paper 0.6950 ±0.0036 best at pub: Spectral Augmentation for Self-Supervised Learning on Graphs (ICLR 2022) paper 0.6950 ±0.0036 Δ at pub 0.0585
271	RGCL GNN baseline actual #271 vs paper #293 22 places worse	0.7348 ±0.0093 best reported	actual #271 vs paper #293 22 places worse paper 0.7092 ±0.0065 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Graph Contrastive Learning with Cohesive Subgraph Awareness (WWW 2024) paper 0.7092 ±0.0065 Δ today 0.0256

RDT-B

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	OEPG Hyb global	0.9630 ±0.0090	Omni-Granular Ego-Semantic Propagation for Self-Supervised Graph Representation Learning
2	MSH-GNN GNN global	0.9580 ±0.0090	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
3	GNN-AK+ code ↗ GNN global	0.9480 ±0.0080	From Stars to Subgraphs: Uplifting Any GNN with Local Structure Awareness code ↗
13	GIN GNN global	0.9240 ±0.0250	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
23	GraphCL GNN baseline actual #23 vs paper #43 20 places worse	0.9169 ±0.0070 best reported	actual #23 vs paper #43 20 places worse paper 0.8953 ±0.0084 best at pub: Graph Contrastive Learning Automated (ICML 2021) best today: Graph Contrastive Learning with Cohesive Subgraph Awareness (WWW 2024) paper 0.8953 ±0.0084 Δ today 0.0216
24	JOAO GNN baseline actual #24 vs paper #63 39 places worse	0.9166 ±0.0059 best reported	actual #24 vs paper #63 39 places worse paper 0.8529 ±0.0135 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) best today: Graph Contrastive Learning with Cohesive Subgraph Awareness (WWW 2024) paper 0.8529 ±0.0135 Δ at pub 0.0285 Δ today 0.0637
26	RGCL GNN baseline actual #26 vs paper #33 7 places worse	0.9138 ±0.0040 best reported	actual #26 vs paper #33 7 places worse paper 0.9034 ±0.0058 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Graph Contrastive Learning with Cohesive Subgraph Awareness (WWW 2024) paper 0.9034 ±0.0058 Δ today 0.0104
28	GraphSC GNN ★ proposed	0.9108 ±0.0056	This paper
33	LaGraph GNN baseline	0.9040 ±0.0080	This paper
34	GraphSC-MSE GNN ★ proposed	0.9026 ±0.0080	This paper
37	AD-GCL GNN baseline actual #37 vs paper #37	0.9010 ±0.0015 best reported	actual #37 vs paper #37 paper 0.9007 ±0.0085 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.9007 ±0.0085 Δ at pub 0.0003
44	SimGRACE GNN baseline	0.8951 ±0.0089	This paper
46	InfoGraph GNN baseline actual #46 vs paper #68 22 places worse	0.8939 ±0.0181 best reported	actual #46 vs paper #68 22 places worse paper 0.8250 ±0.0142 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) best today: Graph Contrastive Learning with Cohesive Subgraph Awareness (WWW 2024) paper 0.8250 ±0.0142 Δ at pub 0.0616 Δ today 0.0689
49	JOAOv2 GNN baseline actual #49 vs paper #58 9 places worse	0.8879 ±0.0065 best reported	actual #49 vs paper #58 9 places worse paper 0.8642 ±0.0145 best at pub: Graph Contrastive Learning with Personalized Augmentation (2022) paper 0.8642 ±0.0145 Δ at pub 0.0237
73	DGK SVM baseline	0.7804 ±0.0039	This paper
75	GL Trad baseline	0.7734 ±0.0018	This paper
80	graph2vec Walk baseline actual #80 vs paper #80	0.7580 ±0.0100 best reported	actual #80 vs paper #80 paper 0.7578 ±0.0103 best at pub: Prototypical Graph Contrastive Learning (2021) best today: Self-Supervised Representation Learning via Latent Graph Prediction (ICML 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7578 ±0.0103 Δ at pub 0.0002 Δ today 0.0002
83	Sub2Vec Walk baseline actual #83 vs paper #83	0.7150 ±0.0040 best reported	actual #83 vs paper #83 paper 0.7148 ±0.0041 best at pub: Self-Supervised Representation Learning via Latent Graph Prediction (ICML 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7148 ±0.0041 Δ at pub 0.0002
85	WL Trad baseline	0.6882 ±0.0041	This paper

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
31	JOAO ext data GNN baseline actual #31 vs paper #232 201 places worse	0.8230 ±0.0220 best reported	actual #31 vs paper #232 201 places worse paper 0.7498 ±0.0029 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7498 ±0.0029 Δ at pub 0.0732
35	EdgePred ext data GNN baseline actual #35 vs paper #181 146 places worse	0.8180 ±0.0250 best reported	actual #35 vs paper #181 146 places worse paper 0.7600 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7600 ±0.0060 Δ at pub 0.0580
46	ContextPred ext data GNN baseline actual #46 vs paper #193 147 places worse	0.8060 ±0.0120 best reported	actual #46 vs paper #193 147 places worse paper 0.7570 ±0.0070 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7570 ±0.0070 Δ at pub 0.0490
47	GraphCL ext data GNN baseline actual #47 vs paper #266 219 places worse	0.8050 ±0.0170 best reported	actual #47 vs paper #266 219 places worse paper 0.7387 ±0.0066 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7387 ±0.0066 Δ at pub 0.0663
71	GraphLoG ext data GNN baseline actual #71 vs paper #242 171 places worse	0.7960 ±0.0250 best reported	actual #71 vs paper #242 171 places worse paper 0.7465 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7465 ±0.0060 Δ at pub 0.0495
163	AttrMasking ext data GNN baseline	0.7670 ±0.0040	This paper
198	GraphSC ext data GNN ★ proposed	0.7558 ±0.0056	This paper
213	DGI ext data GNN baseline	0.7530 ±0.0050	This paper
212	RGCL ext data GNN baseline actual #212 vs paper #219 7 places worse	0.7530 ±0.0050 best reported	actual #212 vs paper #219 7 places worse paper 0.7520 ±0.0034 best today: Rethinking Tokenizer and Decoder in Masked Graph Modeling for Molecules (NeurIPS 2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7520 ±0.0034 Δ today 0.0010
240	AD-GCL ext data GNN baseline actual #240 vs paper #272 32 places worse	0.7470 ±0.0110 best reported	actual #240 vs paper #272 32 places worse paper 0.7356 ±0.0072 best today: Decoupling Weighing and Selecting for Integrating Multiple Graph Pre-training Tasks (ICLR 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7356 ±0.0072 Δ today 0.0114
259	JOAOv2 ext data GNN baseline	0.7427 ±0.0062	This paper
264	No Pre-train GNN baseline	0.7400 ±0.0080	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 ↗
43	GIN code ↗ GNN global	0.8646 ±0.0081	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
98	GraphLoG ext data GNN baseline actual #98 vs paper #115 17 places worse	0.8350 ±0.0120 best reported	actual #98 vs paper #115 17 places worse paper 0.8260 ±0.0125 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: UGMAE: A Unified Framework for Graph Masked Autoencoders (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8260 ±0.0125 Δ today 0.0090
100	JOAO ext data GNN baseline actual #100 vs paper #226 126 places worse	0.8340 ±0.0330 best reported	actual #100 vs paper #226 126 places worse paper 0.7734 ±0.0048 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7734 ±0.0048 Δ at pub 0.0606
176	EdgePred ext data GNN baseline actual #176 vs paper #176	0.7990 ±0.0090	actual #176 vs paper #176 paper 0.7990 ±0.0090 prev best: 16 Pre-training Molecular Graph Representation with 3D Geometry (ICLR 2021) paper 0.7990 ±0.0090 Δ at pub 0.1530
186	ContextPred ext data GNN baseline actual #186 vs paper #186	0.7960 ±0.0120	actual #186 vs paper #186 paper 0.7960 ±0.0120 prev best: 16 Pre-training Molecular Graph Representation with 3D Geometry (ICLR 2021) Improvement within reported uncertainty; not highlighted. paper 0.7960 ±0.0120 Δ at pub 0.0100
191	AttrMasking ext data GNN baseline	0.7930 ±0.0160	This paper
202	AD-GCL ext data GNN baseline actual #202 vs paper #250 48 places worse	0.7900 ±0.0080 best reported	actual #202 vs paper #250 48 places worse paper 0.7502 ±0.0188 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Community-Invariant Graph Contrastive Learning (ICML 2024) paper 0.7502 ±0.0188 Δ today 0.0398
206	GraphSC ext data GNN ★ proposed	0.7882 ±0.0138	This paper
213	GraphCL ext data GNN baseline actual #213 vs paper #248 35 places worse	0.7820 ±0.0120 best reported	actual #213 vs paper #248 35 places worse paper 0.7538 ±0.0144 best at pub: Omni-Granular Ego-Semantic Propagation for Self-Supervised Graph Representation Learning (ICML 2022) paper 0.7538 ±0.0144 Δ at pub 0.0282
224	DGI ext data GNN baseline actual #224 vs paper #243 19 places worse	0.7780 ±0.0050 best reported	actual #224 vs paper #243 19 places worse paper 0.7590 ±0.0160 best today: ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7590 ±0.0160 Δ today 0.0190
242	RGCL ext data GNN baseline	0.7603 ±0.0077	This paper
248	JOAOv2 ext data GNN baseline	0.7549 ±0.0127	This paper
281	No Pre-train GNN baseline	0.7010 ±0.0540	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 …
34	GIN code ↗ GNN global	0.8875 ±0.0049	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
36	GraphCL ext data GNN baseline actual #36 vs paper #160 124 places worse	0.8870 ±0.0190 best reported	actual #36 vs paper #160 124 places worse paper 0.6968 ±0.0067 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6968 ±0.0067 Δ at pub 0.1902
39	JOAO ext data GNN baseline actual #39 vs paper #139 100 places worse	0.8790 ±0.0200 best reported	actual #39 vs paper #139 100 places worse paper 0.7022 ±0.0098 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7022 ±0.0098 Δ at pub 0.1768
40	ContextPred ext data GNN baseline actual #40 vs paper #201 161 places worse	0.8770 ±0.0260 best reported	actual #40 vs paper #201 161 places worse paper 0.6800 ±0.0200 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6800 ±0.0200 Δ at pub 0.1970
44	EdgePred ext data GNN baseline actual #44 vs paper #225 181 places worse	0.8730 ±0.0160 best reported	actual #44 vs paper #225 181 places worse paper 0.6730 ±0.0024 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6730 ±0.0024 Δ at pub 0.2000
54	GraphLoG ext data GNN baseline actual #54 vs paper #120 66 places worse	0.8460 ±0.0080 best reported	actual #54 vs paper #120 66 places worse paper 0.7104 ±0.0186 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.7104 ±0.0186 Δ at pub 0.1356
97	GraphSC ext data GNN ★ proposed	0.7216 ±0.0142	This paper
104	JOAOv2 ext data GNN baseline actual #104 vs paper #117 13 places worse	0.7198 ±0.0018 best reported	actual #104 vs paper #117 13 places worse paper 0.7139 ±0.0092 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 0.7139 ±0.0092 Δ today 0.0059
117	RGCL ext data GNN baseline actual #117 vs paper #117	0.7142 ±0.0066	actual #117 vs paper #117 paper 0.7142 ±0.0066 prev best: Subgraph Networks Based Contrastive Learning (2023) Improvement within reported uncertainty; not highlighted. paper 0.7142 ±0.0066 Δ at pub 0.0002
118	SimGRACE ext data GNN baseline	0.7125 ±0.0086	This paper
144	AD-GCL ext data GNN baseline actual #144 vs paper #196 52 places worse	0.7010 ±0.0120 best reported	actual #144 vs paper #196 52 places worse paper 0.6826 ±0.0103 best today: Decoupling Weighing and Selecting for Integrating Multiple Graph Pre-training Tasks (ICLR 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.6826 ±0.0103 Δ today 0.0184
184	DGI ext data GNN baseline	0.6880 ±0.0080	This paper
230	AttrMasking ext data GNN baseline actual #230 vs paper #279 49 places worse	0.6712 ±0.0045 best reported	actual #230 vs paper #279 49 places worse paper 0.6430 ±0.0280 best at pub: GraphMAE: Self-Supervised Masked Graph Autoencoders (KDD 2022) 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 0.6430 ±0.0280 Δ today 0.0282
257	No Pre-train GNN baseline	0.6580 ±0.0450	This paper

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 ↗
19	GCN GNN global	0.9130 ±0.0170	OOD-GNN: Out-of-Distribution Generalized Graph Neural Network
57	RGCL ext data GNN baseline actual #57 vs paper #63 6 places worse	0.8500 ±0.0080 best reported	actual #57 vs paper #63 6 places worse paper 0.8338 ±0.0091 best at pub: Subgraph Networks Based Contrastive Learning (2023) best today: Rethinking Tokenizer and Decoder in Masked Graph Modeling for Molecules (NeurIPS 2023) paper 0.8338 ±0.0091 Δ at pub 0.0002 Δ today 0.0162
74	GraphSC ext data GNN ★ proposed	0.8229 ±0.0121	This paper
79	JOAO ext data GNN baseline actual #79 vs paper #79	0.8132 ±0.0249	actual #79 vs paper #79 paper 0.8132 ±0.0249 prev best: GraphMAE: Self-Supervised Masked Graph Autoencoders (KDD 2022) Improvement within reported uncertainty; not highlighted. paper 0.8132 ±0.0249 Δ at pub 0.0002
80	JOAOv2 ext data GNN baseline	0.8097 ±0.0164	This paper
82	AD-GCL ext data GNN baseline actual #82 vs paper #107 25 places worse	0.8077 ±0.0392 best reported	actual #82 vs paper #107 25 places worse paper 0.7763 ±0.0421 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Community-Invariant Graph Contrastive Learning (ICML 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7763 ±0.0421 Δ today 0.0314
97	GraphCL ext data GNN baseline actual #97 vs paper #121 24 places worse	0.7890 ±0.0420 best reported	actual #97 vs paper #121 24 places worse paper 0.7599 ±0.0265 best at pub: 16 Pre-training Molecular Graph Representation with 3D Geometry (ICLR 2021) best today: Improving Molecular Pretraining with Complementary Featurizations (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7599 ±0.0265 Δ at pub 0.0291 Δ today 0.0291
103	GraphLoG ext data GNN baseline	0.7872 ±0.0258	This paper
115	ContextPred ext data GNN baseline actual #115 vs paper #197 82 places worse	0.7720 ±0.0088 best reported	actual #115 vs paper #197 82 places worse paper 0.6590 ±0.0380 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 0.6590 ±0.0380 Δ at pub 0.1130 Δ today 0.1130
162	AttrMasking ext data GNN baseline	0.7180 ±0.0410	This paper
175	DGI ext data GNN baseline	0.6990 ±0.0300	This paper
192	EdgePred ext data GNN baseline actual #192 vs paper #207 15 places worse	0.6730 ±0.0200 best reported	actual #192 vs paper #207 15 places worse paper 0.6410 ±0.0370 best at pub: Subgraph Networks Based Contrastive Learning (2023) best today: Decoupling Weighing and Selecting for Integrating Multiple Graph Pre-training Tasks (ICLR 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.6410 ±0.0370 Δ today 0.0320
232	No Pre-train GNN baseline	0.5800 ±0.0440	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 ↗
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 #183 115 places worse	0.6400 ±0.0100 best reported	actual #68 vs paper #183 115 places worse paper 0.5997 ±0.0079 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.5997 ±0.0079 Δ at pub 0.0403
103	ContextPred GNN baseline actual #103 vs paper #144 41 places worse	0.6220 ±0.0059 best reported	actual #103 vs paper #144 41 places worse paper 0.6090 ±0.0060 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 0.6090 ±0.0060 Δ at pub 0.0130 Δ today 0.0130
114	GraphCL GNN baseline actual #114 vs paper #162 48 places worse	0.6183 ±0.0060 best reported	actual #114 vs paper #162 48 places worse paper 0.6053 ±0.0088 best at pub: Omni-Granular Ego-Semantic Propagation for Self-Supervised Graph Representation Learning (ICML 2022) 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 0.6053 ±0.0088 Δ at pub 0.0027 Δ today 0.0130
123	GraphLoG GNN baseline actual #123 vs paper #215 92 places worse	0.6150 ±0.0130 best reported	actual #123 vs paper #215 92 places worse paper 0.5786 ±0.0144 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.5786 ±0.0144 Δ at pub 0.0364
128	RGCL GNN baseline actual #128 vs paper #130 2 places worse	0.6140 ±0.0060 best reported	actual #128 vs paper #130 2 places worse paper 0.6138 ±0.0061 best at pub: Subgraph Networks Based Contrastive Learning (2023) best today: Dual-perspective Cross Contrastive Learning in Graph Transformers (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.6138 ±0.0061 Δ at pub 0.0002 Δ today 0.0002
133	AttrMasking GNN baseline actual #133 vs paper #141 8 places worse	0.6121 ±0.0065 best reported	actual #133 vs paper #141 8 places worse paper 0.6100 ±0.0070 best at pub: GraphMAE: Self-Supervised Masked Graph Autoencoders (KDD 2022) 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 0.6100 ±0.0070 Δ today 0.0021
146	AD-GCL GNN baseline actual #146 vs paper #198 52 places worse	0.6080 ±0.0090 best reported	actual #146 vs paper #198 52 places worse paper 0.5924 ±0.0086 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Community-Invariant Graph Contrastive Learning (ICML 2024) paper 0.5924 ±0.0086 Δ today 0.0156
156	SimGRACE GNN baseline actual #156 vs paper #161 5 places worse	0.6060 ±0.0100 best reported	actual #156 vs paper #161 5 places worse paper 0.6059 ±0.0096 best at pub: Graph Self-supervised Learning with Accurate Discrepancy Learning (NeurIPS 2022) best today: Dual-perspective Cross Contrastive Learning in Graph Transformers (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.6059 ±0.0096 Δ today 0.0001
168	GraphSC ext data GNN ★ proposed	0.6049 ±0.0041	This paper
167	JOAOv2 GNN baseline	0.6049 ±0.0074	This paper
170	EdgePred GNN baseline	0.6040 ±0.0070	This paper
180	DGI GNN baseline actual #180 vs paper #206 26 places worse	0.6009 ±0.0582 best reported	actual #180 vs paper #206 26 places worse paper 0.5840 ±0.0080 best today: MAGPrompt: Message-Adaptive Graph Prompt Tuning for Graph Neural Networks (2026) Difference within reported uncertainty; not counted as underperformance. paper 0.5840 ±0.0080 Δ today 0.0169
222	No Pre-train GNN baseline	0.5730 ±0.0160	This paper

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 ext data GNN baseline actual #28 vs paper #131 103 places worse	0.7140 ±0.0180 best reported	actual #28 vs paper #131 103 places worse paper 0.6390 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6390 ±0.0060 Δ at pub 0.0750
30	EdgePred ext data GNN baseline actual #30 vs paper #127 97 places worse	0.7120 ±0.0110 best reported	actual #30 vs paper #127 97 places worse paper 0.6410 ±0.0060 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6410 ±0.0060 Δ at pub 0.0710
31	JOAO ext data GNN baseline actual #31 vs paper #165 134 places worse	0.7110 ±0.0140 best reported	actual #31 vs paper #165 134 places worse paper 0.6294 ±0.0048 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6294 ±0.0048 Δ at pub 0.0816
42	GraphCL ext data GNN baseline actual #42 vs paper #178 136 places worse	0.6960 ±0.0230 best reported	actual #42 vs paper #178 136 places worse paper 0.6240 ±0.0057 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6240 ±0.0057 Δ at pub 0.0720
64	GraphLoG ext data GNN baseline actual #64 vs paper #180 116 places worse	0.6770 ±0.0190 best reported	actual #64 vs paper #180 116 places worse paper 0.6232 ±0.0051 best at pub: KPGT: Knowledge-Guided Pre-training of Graph Transformer for Molecular Property Prediction (KDD 2022) paper 0.6232 ±0.0051 Δ at pub 0.0538
122	GraphSC ext data GNN ★ proposed	0.6427 ±0.0023	This paper
126	AttrMasking ext data GNN baseline	0.6420 ±0.0050	This paper
138	AD-GCL ext data GNN baseline actual #138 vs paper #159 21 places worse	0.6350 ±0.0070 best reported	actual #138 vs paper #159 21 places worse paper 0.6310 ±0.0066 best at pub: Let Invariant Rationale Discovery Inspire Graph Contrastive Learning (ICML 2022) best today: Community-Invariant Graph Contrastive Learning (ICML 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.6310 ±0.0066 Δ today 0.0040
142	No Pre-train GNN baseline	0.6340 ±0.0060	This paper
145	SimGRACE ext data GNN baseline	0.6336 ±0.0052	This paper
148	RGCL ext data GNN baseline	0.6333 ±0.0017	This paper
158	JOAOv2 ext data GNN baseline	0.6316 ±0.0045	This paper
169	DGI ext data GNN baseline	0.6270 ±0.0040	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 at pub — reported value improved over the best known when published Best today — OK at publication; today's leaderboard best is elsewhere