Enhancing Molecular Property Predictions by Learning from Bond Modelling and Interactions

Yunqing Liu; Yi Zhou; Wenqi Fan

cs.LGFeb 28, 2026

Enhancing Molecular Property Predictions by Learning from Bond Modelling and Interactions

Authors: Yunqing Liu, Yi Zhou, Wenqi Fan

Organizations: Department of Computing (COMP), The Hong Kong Polytechnic University · Department of Management and Marketing (MM), The Hong Kong Polytechnic University

Abstract

Molecule representation learning is crucial for understanding and predicting molecular properties. However, conventional atom-centric models, which treat chemical bonds merely as pairwise interactions, often overlook complex bond-level phenomena like resonance and stereoselectivity. This oversight limits their predictive accuracy for nuanced chemical behaviors. To address this limitation, we introduce DeMol, a dual-graph framework whose architecture is motivated by a rigorous information-theoretic analysis demonstrating the information gain from a bond-centric perspective. DeMol explicitly models molecules through parallel atom-centric and bond-centric channels. These are synergistically fused by multi-scale Double-Helix Blocks designed to learn intricate atom-atom, atom-bond, and bond-bond interactions. The framework's geometric consistency is further enhanced by a regularization term based on covalent radii to enforce chemically plausible structures. Comprehensive evaluations on diverse benchmarks, including PCQM4Mv2, OC20 IS2RE, QM9, and MoleculeNet, show that DeMol establishes a new state-of-the-art, outperforming existing methods. These results confirm the superiority of explicitly modelling bond information and interactions, paving the way for more robust and accurate molecular machine learning.

Figures & tables

Figure 1: An example of the different bonds within benzene (C 6 H 6 ). Formed by $p$ -orbital overlap, the delocalized $\pi$ system creates electron density above/below the ring plane, providing stability and unique reactivity.

Figure 2: An example of bond-bond interactions in molecules that affect properties. Left is the anticancer drug cisplatin, where two ammonia ligands and two chloride ions bind to a central platinum atom in a cis configuration. Right is transplatin, which possesses the same atomic composition but features ligands in a trans configuration, rendering it pharmacologically ineffective.

Figure 3: DeMol integrates atom-centric and bond-centric channels via dual-graph representations. Cross-level (atom-bond) interactions are enforced through double-helix blocks, ensuring geometric consistency.

Model	# param.	MAE ( $\downarrow$ )
MLP-Fingerprint	16.1M	0.1735
GCN	2.0M	0.1379
GIN	3.8M	0.1195
GINE - ${}_{\text{VN}}$	13.2M	0.1167
GCN - ${}_{\text{VN}}$	4.9M	0.1153
GIN - ${}_{\text{VN}}$	6.7M	0.1083

Table 1: Performance on PCQM4MV2 validation set ( Hu et al., 2021 ) . Bold values indicate the best performance.

	Energy MAE (eV) $\downarrow$					EwT (%) $\uparrow$
Model	ID	OOD Ads.	OOD Cat.	OOD Both	AVG.	ID	OOD Ads.	OOD Cat.	OOD Both	AVG.
SchNet	0.6465	0.7074	0.6475	0.6626	0.6660	2.96	2.22	3.03	2.38	2.65
DimeNet++	0.5636	0.7127	0.5612	0.6492	0.6217	4.25	2.48	4.40	2.56	3.42
GemNet-T	0.5561	0.7342	0.5659	0.6964	0.6382	4.51	2.24	4.37	2.38	3.38
SphereNet	0.5632	0.6682	0.5590	0.6190	0.6024	4.56	2.70	4.59	2.70	3.64
Graphormer-3D	0.4329	0.5850	0.4441	0.5299	0.4980	-	-	-	-	-

Table 2: Performance on OC20 IS2RE validation set. NN refers to "Noisy Nodes" ( Godwin et al., 2021 ) . Bold values indicate the best performance.

Model	$\mu$	$\alpha$	$\epsilon_{HOMO}$	$\epsilon_{LUMO}$	$\Delta\epsilon$	$R^{2}$	ZPVE	$U_{0}$	U	H	G	$C_{v}$	Avg. Rank $\#$
EdgePred	0.039	0.086	37.4	31.9	58.2	0.112	1.81	14.7	14.2	14.8	14.5	0.038	18.92
AttrMask	0.020	0.072	31.3	37.8	50.0	0.423	1.90	10.7	10.8	11.4	11.2	0.062	15.25
InfoGraph	0.041	0.099	48.1	38.1	72.2	0.114	1.69	16.4	14.9	14.5	16.5	0.030	19.08
GraphCL	0.027	0.066	26.8	22.9	45.5	0.095	1.42	9.6	9.7	9.6	10.2	0.028	10.0
GPT-GNN	0.039	0.103	35.7	28.8	54.1	0.158	1.75	12.0	24.8	14.8	12.2	0.032	18.17
GraphMVP	0.031	0.070	28.5	26.3	46.9	0.082	1.63	10.2	10.3	10.4	11.2	0.033	12.00

Table 3: Results on QM9. The evaluation metric is the Mean Absolute Error (MAE). We report the official results of baselines from Appendix D.4 . Bold values indicate the best performance.

Figure 4: Results on molecular property classification tasks. The table version is Appendix Table 6 .

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Inference Time
Transformer-M	$\sim$ 28 ms/molecule
GPS++ (Ensemble)	$\sim$ 100 ms/molecule
DeMol	$\sim$ 34 ms/molecule

Appendix

Table 5: Inference Time Comparison on PCQM4Mv2 dataset.

Methods	BBBP $\uparrow$	Tox21 $\uparrow$	ToxCast $\uparrow$	SIDER $\uparrow$	ClinTox $\uparrow$	MUV $\uparrow$	HIV $\uparrow$	BACE $\uparrow$	Avg $\uparrow$
AttrMask	$65.0\pm 2.3$	$74.8\pm 0.2$	$62.9\pm 0.1$	$61.2\pm 0.1$	$87.7\pm 1.1$	$73.4\pm 2.0$	$76.8\pm 0.5$	$79.7\pm 0.3$	72.68
ContextPred	$65.7\pm 0.6$	$74.2\pm 0.0$	$62.5\pm 0.3$	$62.2\pm 0.5$	$77.2\pm 0.8$	$75.3\pm 1.5$	$77.1\pm 0.8$	$76.0\pm 2.0$	71.28
GraphCL	$69.7\pm 0.6$	$73.9\pm 0.6$	$62.4\pm 0.5$	$60.5\pm 0.8$	$76.0\pm 2.6$	$69.8\pm 2.6$	$78.5\pm 1.2$	$75.4\pm 1.4$	70.78
InfoGraph	$67.5\pm 0.1$	$73.2\pm 0.4$	$63.7\pm 0.5$	$59.9\pm 0.3$	$76.5\pm 1.0$	$74.1\pm 0.7$	$75.1\pm 0.9$	$77.8\pm 0.8$	70.98
GROVER	$70.0\pm 0.1$	$74.3\pm 0.1$	$65.4\pm 0.4$	$64.8\pm 0.6$	$81.2\pm 3.0$	$67.3\pm 1.8$	$62.5\pm 0.9$	$82.6\pm 0.7$	71.01
MolCLR	$66.6\pm 1.8$	$73.0\pm 0.1$	$62.9\pm 0.3$	$57.5\pm 1.7$	$86.1\pm 0.9$	$72.5\pm 2.3$	$76.2\pm 1.5$	$71.5\pm 3.1$	70.79

Appendix

Table 6: Results on molecular property classification tasks. We report the (mean

\pm

standard deviation) ROC-AUC score (higher is better) of 10 random seeds under scaffold splitting. The best results are highlighted in bold.

Hyperparameter	Small	Large	HIV
Learning rate	[5e-5, 8e-5, 1e-4, 4e-4, 5e-4]	[2e-5, 1e-4]	[2e-5, 5e-5]
Batch size	[32, 64, 128, 256]	[128, 256]	[128, 256]
Epochs	[40 ,60, 80, 100]	[20, 40]	[2, 5, 10]
Pooler dropout	[0.0, 0.1, 0.2, 0.5]	[0.0, 0.1]	[0.0, 0.2]
Warmup ratio	[0.0, 0.06, 0.1]	[0.0, 0.06]	[0.0, 0.1]

Appendix

Table 7: Search space for small datasets: BBBP, BACE, ClinTox, Tox21, Toxcast, SIDER, for large datasets: MUV and HIV.

Figure 5: Visualisation on self-attention map of multi-heads independently.

Figure 6: Visualisation on qualitative analysis of bond-level interactions.

Methods	$\epsilon_{HOMO}$	$\epsilon_{LUMO}$	$\Delta\epsilon$	$G$	$C_{v}$
No Pre-training	21.2	19.3	30.4	10.97	0.029
Pre-training on PCQM4Mv2	16.4	15.7	26.8	6.29	0.021

Appendix

Table 8: The impact of pre-training on PCQM4Mv2.

Figure 7: An example of the two graph representations of the methane molecule. Atom-Centric Graph : Graph representation with atoms as nodes, which can explicitly encode structural information such as interatomic distances. Bond-Centric Graph : Graph representation with bonds as nodes, which can explicitly encode structural information such as inter-bond angles.

Molecular Property Prediction

Molecular property prediction estimates chemical, physical, and biological properties from molecular structure and context.

See all 201 papers

Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction
May 15, 2026Xiayu Liu, Zhengyi Lu, Hou-biao LiSelf-Supervised Pre-Training Hierarchical Representation Learning
School of Mathematical Sciences, University of Electronic Science and Technology of China, No.2006, Xiyuan Avenue, West Hi-tech Zone,2026 611731, Chengdu, China · Department of Computer Science and Engineering, Oakland University, 201 Meadow Brook RD, Rochester Hills, 48309, MI, USA
DMMRL: Disentangled Multi-Modal Representation Learning via Variational Autoencoders for Molecular Property Prediction
Mar 22, 2026Long Xu, Junping Guo, Jianbo Zhao +2Molecular Representation Learning Multimodal Disentangled Representation Learning
Guangxi Key Lab of Human-machine Interaction and Intelligent Decision, Nanning Normal University, Nanning, China · College of Big Data and Software Engineering, Zhejiang Wanli University, Ningbo, China
Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints
Sep 14, 2026Gwang-Hyeon Yun, Jong-Hoon Park, Bing Hu +3Graph SSL Hierarchical Representation Learning
Yonsei University · Wonju-si, Republic of Korea · University of Waterloo +1

May 15, 2026J/K move · Enter open · S save

Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction

Xiayu Liu, Zhengyi Lu, Hou-biao Li

School of Mathematical Sciences, University of Electronic Science and Technology of China, No.2006, Xiyuan Avenue, West Hi-tech Zone,2026 611731, Chengdu, China · Department of Computer Science and Engineering, Oakland University, 201 Meadow Brook RD, Rochester Hills, 48309, MI, USA

Self-Supervised Pre-Training Hierarchical Representation Learning Molecular Property Prediction Molecular Representation Learning

Open paper AskPDF arXiv

Self-supervised pretraining on molecular graphs has emerged as a promising approach for molecular property prediction, yet most existing methods operate at a single structural granularity and treat bond information as auxiliary edge attributes rather than as an independent semantic layer. In this work, we propose MolCHG, a multi-level self-supervised pretraining framework built upon a novel Compositional Hierarchical Graph that organizes molecular structure into four types of nodes across three semantic levels. By introducing a bond graph that operates in parallel with the atom graph, our architecture elevates bond-level information to independently evolving node representations, enabling fragment nodes to aggregate atom-level and bond-level semantics on an equal footing. We design three level-specific pretraining objectives: an atom-bond cross-view contrastive task that aligns the atom-view and bond-view representations within each fragment, a fragment-level functional group prediction task to inject domain-relevant chemical knowledge, and graph-level structure prediction tasks to encode global molecular topology. Experiments on nine MoleculeNet benchmarks demonstrate that MolCHG achieves the best performance on seven datasets across both classification and regression tasks, remaining competitive with the strongest baselines on the rest. Ablation studies further confirm that the multi-level supervision signals are complementary and that each component contributes to the overall performance.

Abstract

Molecular Property Prediction

Explore similar work

Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction

DMMRL: Disentangled Multi-Modal Representation Learning via Variational Autoencoders for Molecular Property Prediction

Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints