A Zero-Shot Single-Point Chemical Language Model Mimics Medicinal Chemist Reasoning for Molecular Optimization and STING Inhibitor Development in Alzheimer’s Disease
8 affiliations
- 1Bioinnovation center, Department of biomedical informatics and genomics, Tulane school of medicine
- 2Tulane University
- 3Department of Chemical and Systems Biology, School of Medicine
- 4Stanford University
- 5Genion Quantum Technologies Inc.
- 6Zermatt Biotech Inc.
- 7The University of Texas Health Science Center
- 8Johns Hopkins University
Abstract
Drug development relies heavily on the ability of medicinal chemists to identify productive, localized structural modifications that improve potency, selectivity, and developability while preserving favorable features of a lead scaffold. Here, we introduce the Single-Point Chemical Language Model (SpCLM), a zero-shot molecular optimization framework designed to computationally emulate this iterative medicinal chemistry paradigm. Built on a Transformer architecture, SpCLM performs controlled single-point molecular modifications encompassing common medicinal chemistry operations and generates focused, chemically interpretable analog spaces rather than large, unconstrained molecular libraries. This strategy enables efficient exploration of structure–activity relationships while maintaining close structural relationships to experimentally tractable lead compounds. Across multiple protein targets and molecular optimization tasks, SpCLM generated compact libraries typically comprising only a few hundred molecules, yet recovered 60%–80% of experimentally validated active compounds from held-out test sets that were not included during model training. Generated molecules showed substantial agreement with experimentally observed structure–activity relationships, including changes in binding affinity and functional activity, demonstrating that the model can reproduce productive chemical transformations without target-specific retraining. These results establish single-point molecular editing as an efficient strategy for translating learned medicinal chemistry knowledge into experimentally relevant molecular optimization. We further applied SpCLM to the development of stimulator of interferon genes (STING) inhibitors as potential therapeutics for Alzheimer’s disease (AD). Because aberrant activation of the cGAS–STING innate immune pathway contributes to neuroinflammatory processes associated with AD and related neurodegenerative disorders, pharmacological inhibition of STING represents an emerging therapeutic strategy. Starting from experimentally characterized STING inhibitor chemotypes, SpCLM generated focused analog series through medicinal-chemistry-like single-point modifications. Integration of model-guided generation with structure-based prioritization and experimental evaluation enabled efficient exploration and optimization of STING inhibitor chemical space, identifying analogs with improved activity and providing experimentally supported structure–activity relationships for further lead development. Together, these results demonstrate that SpCLM bridges generative molecular modeling and practical medicinal chemistry by converting broad chemical knowledge into focused, experimentally actionable structural modifications. By recovering a substantial fraction of experimentally active chemical space from only hundreds of generated candidates and enabling the optimization of therapeutically relevant STING inhibitor chemotypes, SpCLM provides a generalizable framework for reducing the experimental search space and accelerating iterative drug discovery.
Related
- Development of 3D-Printed Photoreactors for Medicinal Chemistry Applications
- Drug Molecular and Route Similarity-Aware Retrieval-Augmented LLM Generation with Knowledge Graph for Semantic Reasoning
- Optimizing Affinity and Pharmacokinetics of a Cysteine-sulfoxide Derivative as Potential Inhibitor of BACE-1 using Fragment Growing and Molecular Dynamics
- Corosolic Acid in Hepatocellular Carcinoma: Molecular Properties, Evidence Boundaries, and Development Priorities
- Minimum Knowledge Collaboration enables joint materials development through protected optimization
BibTeX
@misc{gao2026zeroshot,
title = {{A Zero-Shot Single-Point Chemical Language Model Mimics Medicinal Chemist Reasoning for Molecular Optimization and STING Inhibitor Development in Alzheimer’s Disease}},
author = {Peng Gao and Ying Qin and Zhilian Dai and Yichao Liu and Hao Wen and Jie Zhang and Songyou Zhong and Tenghuan Ge and Chen Wang and Jiawei Fu and Dan Zhang and Ning Zhi and Zhen Zhao},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv-2025-m82r5/v2},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-m82r5/v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - A Zero-Shot Single-Point Chemical Language Model Mimics Medicinal Chemist Reasoning for Molecular Optimization and STING Inhibitor Development in Alzheimer’s Disease
AU - Peng Gao
AU - Ying Qin
AU - Zhilian Dai
AU - Yichao Liu
AU - Hao Wen
AU - Jie Zhang
AU - Songyou Zhong
AU - Tenghuan Ge
AU - Chen Wang
AU - Jiawei Fu
AU - Dan Zhang
AU - Ning Zhi
AU - Zhen Zhao
PY - 2026
DA - 2026/09/08
PB - ChemRxiv
DO - 10.26434/chemrxiv-2025-m82r5/v2
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-m82r5/v2
AB - Drug development relies heavily on the ability of medicinal chemists to identify productive, localized structural modifications that improve potency, selectivity, and developability while preserving favorable features of a lead scaffold. Here, we introduce the Single-Point Chemical Language Model (SpCLM), a zero-shot molecular optimization framework designed to computationally emulate this iterative medicinal chemistry paradigm. Built on a Transformer architecture, SpCLM performs controlled single-point molecular modifications encompassing common medicinal chemistry operations and generates focused, chemically interpretable analog spaces rather than large, unconstrained molecular libraries. This strategy enables efficient exploration of structure–activity relationships while maintaining close structural relationships to experimentally tractable lead compounds. Across multiple protein targets and molecular optimization tasks, SpCLM generated compact libraries typically comprising only a few hundred molecules, yet recovered 60%–80% of experimentally validated active compounds from held-out test sets that were not included during model training. Generated molecules showed substantial agreement with experimentally observed structure–activity relationships, including changes in binding affinity and functional activity, demonstrating that the model can reproduce productive chemical transformations without target-specific retraining. These results establish single-point molecular editing as an efficient strategy for translating learned medicinal chemistry knowledge into experimentally relevant molecular optimization. We further applied SpCLM to the development of stimulator of interferon genes (STING) inhibitors as potential therapeutics for Alzheimer’s disease (AD). Because aberrant activation of the cGAS–STING innate immune pathway contributes to neuroinflammatory processes associated with AD and related neurodegenerative disorders, pharmacological inhibition of STING represents an emerging therapeutic strategy. Starting from experimentally characterized STING inhibitor chemotypes, SpCLM generated focused analog series through medicinal-chemistry-like single-point modifications. Integration of model-guided generation with structure-based prioritization and experimental evaluation enabled efficient exploration and optimization of STING inhibitor chemical space, identifying analogs with improved activity and providing experimentally supported structure–activity relationships for further lead development. Together, these results demonstrate that SpCLM bridges generative molecular modeling and practical medicinal chemistry by converting broad chemical knowledge into focused, experimentally actionable structural modifications. By recovering a substantial fraction of experimentally active chemical space from only hundreds of generated candidates and enabling the optimization of therapeutically relevant STING inhibitor chemotypes, SpCLM provides a generalizable framework for reducing the experimental search space and accelerating iterative drug discovery.
ER -