Hierarchical Molecular Representation Learning via Fragment-Based Self-Supervised Embedding Prediction
#Graph Self-Supervised Learning #Molecular Representation #GraSPNet #Hierarchical Learning #Fragment-Based Prediction #Molecular Properties #Transfer Learning
📌 Key Takeaways
- Researchers developed GraSPNet framework for hierarchical molecular representation learning
- Method decomposes molecular graphs into chemically meaningful fragments without predefined vocabularies
- Multi-level message passing captures both atomic and fragment-level semantics
- Outperformed existing methods in transfer learning for molecular property prediction
📖 Full Retelling
Researchers Jiele Wu, Haozhe Ma, Zhihan Guo, Thanh Vinh Vo, and Tze Yun Leong introduced the Graph Semantic Predictive Network (GraSPNet) framework in a paper submitted to arXiv on February 23, 2026, addressing limitations in current Graph Self-Supervised Learning methods that overlook chemically relevant molecular substructures affecting molecular properties. The paper presents a novel hierarchical self-supervised approach that explicitly models both atomic-level and fragment-level semantics in molecular graphs, representing a significant advancement in computational chemistry and machine learning. GraSPNet innovatively decomposes molecular structures into chemically meaningful fragments without relying on predefined vocabularies, enabling more accurate representation of molecular properties that depend on specific substructures. The framework employs multi-level message passing to simultaneously learn representations at both node and fragment levels, incorporating masked semantic prediction at each hierarchical level to capture structural information at multiple resolutions. This comprehensive approach allows the model to capture relationships between atoms and larger molecular fragments that traditional methods often miss, leading to more expressive and transferable molecular representations. The researchers validated their method through extensive experiments on multiple molecular property prediction benchmarks, demonstrating consistent performance improvements over existing state-of-the-art GSSL approaches in transfer learning scenarios.
🏷️ Themes
Machine Learning, Molecular Representation, Self-Supervised Learning
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
--> Computer Science > Machine Learning arXiv:2602.20344 [Submitted on 23 Feb 2026] Title: Hierarchical Molecular Representation Learning via Fragment-Based Self-Supervised Embedding Prediction Authors: Jiele Wu , Haozhe Ma , Zhihan Guo , Thanh Vinh Vo , Tze Yun Leong View a PDF of the paper titled Hierarchical Molecular Representation Learning via Fragment-Based Self-Supervised Embedding Prediction, by Jiele Wu and 4 other authors View PDF HTML Abstract: Graph self-supervised learning has demonstrated strong potential for generating expressive graph embeddings without the need for human annotations, making it particularly valuable in domains with high labeling costs such as molecular graph analysis. However, existing GSSL methods mostly focus on node- or edge-level information, often ignoring chemically relevant substructures which strongly influence molecular properties. In this work, we propose Graph Semantic Predictive Network , a hierarchical self-supervised framework that explicitly models both atomic-level and fragment-level semantics. GraSPNet decomposes molecular graphs into chemically meaningful fragments without predefined vocabularies and learns node- and fragment-level representations through multi-level message passing with masked semantic prediction at both levels. This hierarchical semantic supervision enables GraSPNet to learn multi-resolution structural information that is both expressive and transferable. Extensive experiments on multiple molecular property prediction benchmarks demonstrate that GraSPNet learns chemically meaningful representations and consistently outperforms state-of-the-art GSSL methods in transfer learning settings. Comments: 15 pages (8 pages main text),8 figures Subjects: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI); Quantitative Methods (q-bio.QM) Cite as: arXiv:2602.20344 [cs.LG] (or arXiv:2602.20344v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2602.20344 Focus to learn more arXiv-issued DOI via...
Read full article at source