ReMoD-Mol: Reliability-Aware Multimodal Learning with Descriptor-Guided Knowledge Distillation for Molecular Property Prediction
Bioinformatics · 16.11.2026
DOI: 10.1093/bioinformatics/btag838 ↗Kurz gefasst
Fachartikel in der Zeitschrift Bioinformatics, online erschienen am 16.11.2026. Verfasst von Yilin Yang und 3 weiteren Autor:innen. Beteiligte Einrichtung: School of Informatics, Hunan University of Chinese Medicine, Changsha 410208, China. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Motivation Accurate molecular property prediction (MPP) remains a central challenge in drug discovery. Given the limited availability of labeled molecular data, multimodal molecular representation learning offers a promising means of integrating complementary structural information for MPP. Existing multimodal models still face two limitations: (i) common fusion strategies, particularly direct concatenation, can introduce redundant information or signals that are weakly relevant to the target property; and (ii) the effective incorporation of complementary physicochemical priors into multimodal representations remains insufficiently explored. Results Here, we proposed ReMoD-Mol, a reliability-aware multimodal framework that integrated molecular graphs, SMILES sequences and two-dimensional molecular images. Task-adaptive modality calibration learned one set of weights per benchmark from downstream supervision to regulate the relative contributions of the three molecular views. Descriptor- and fingerprint-based teachers further provided task-conditioned chemical priors through train-only, five-fold out-of-fold knowledge distillation. Prototype-guided embedding shaping together with graph-view and SMILES-view consistency objectives regularized the fused representation space. Across nine MoleculeNet benchmark datasets spanning regression and classification, ReMoD-Mol achieved the best mean performance on seven datasets and the second-best on the other two in a scaffold-split comparison with 14 baselines. Representation analyses revealed smoother property-related manifolds and more discriminative class structures, while learned modality coefficients exhibited distinct task-dependent fusion patterns. Teacher-prior and ablation analyses further supported the complementary roles of chemical-prior transfer and multi-view representation regularization. Availability and implementation Source code, processed datasets and implementation details are publicly available at https://github.com/27167199/ReMoD-Mol.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- School of Informatics, Hunan University of Chinese Medicine, Changsha 410208, China.
- School of Computer Science and Engineering, Central South University, Changsha, China.
- School of Biomedical Sciences, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong SAR 999077, China.
Zitieren
APA
Yang, Y., Chen, S., Zhao, Q., & Yan, C. (2026). ReMoD-Mol: Reliability-Aware Multimodal Learning with Descriptor-Guided Knowledge Distillation for Molecular Property Prediction. Bioinformatics. https://doi.org/10.1093/bioinformatics/btag838
Vancouver
Yang Y, Chen S, Zhao Q, Yan C. ReMoD-Mol: Reliability-Aware Multimodal Learning with Descriptor-Guided Knowledge Distillation for Molecular Property Prediction. Bioinformatics. 2026. doi:10.1093/bioinformatics/btag838
BibTeX
@article{Yang2026,
author = {Yang, Yilin and Chen, Shuqi and Zhao, Qichang and Yan, Cheng},
title = {ReMoD-Mol: Reliability-Aware Multimodal Learning with Descriptor-Guided Knowledge Distillation for Molecular Property Prediction},
journal = {Bioinformatics},
year = {2026},
doi = {10.1093/bioinformatics/btag838}
}
EUVIMED