Classical Structural Priors Improve Neural Molecular Representations in Molecule Language Models
Abstract Classical molecular similarity is central to chemical information systems, yet molecule-language models trained on paired structure–text data do not necessarily organize molecular embeddings according to graded structural relationships. Here, we introduce Structure-Aware Guided Alignment (SAGA), a training-only regularizer that uses Centered Kernel Alignment to align neural molecular representations with fingerprint-based Tanimoto similarity and requires no additional parameters or fingerprints at inference. In MolCA and Atomas, SAGA improves molecule–text retrieval, retrieves closer structural analogues, and raises average performance in scaffold-split property prediction, without clear degradation of molecular description or text-to-molecule generation. These results, together with control experiments that attribute the gains to the chemical content of the prior, show that established chemoinformatics priors can make molecule-language representations more chemically faithful while retaining their multimodal utility.
Authors
- Sungroh Yoon (ORCID: https://orcid.org/0000-0002-2367-197X)
- Munyoung Chang
- Donghyeok Lee (ORCID: https://orcid.org/0000-0002-6671-8117)
- Sangkwon Park
Institutions
- Seoul National University (KR)
- Seoul National University Bundang Hospital (KR)
Publication Details
- Journal
- Journal of Chemical Information and Modeling
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1021/acs.jcim.6c02225
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Ministry of Health and Welfare
- Ministry of Science and ICT, South Korea