Computational design and prioritization of database-absent small-molecule GLP-1 receptor agonist candidates from public bioactivity data
Abstract
Public bioactivity data were used to construct a reproducible workflow for the computational prioritization of small-molecule glucagon-like peptide-1 receptor (GLP-1R) agonist candidates. Exact IC50, EC50, and Ki measurements were standardized, converted to pActivity, and deduplicated after structural exclusion of peptide-like records. The resulting dataset comprised 337 compounds and 155 Murcko scaffolds. Morgan fingerprints and physicochemical descriptors were evaluated using ridge regression, random forest, and XGBoost with scaffold-based validation. Random forest provided the lowest mean five-fold cross-validation root-mean-square error (RMSE; 0.497 pActivity units). Its scaffold-disjoint holdout performance was R2 = 0.236, RMSE = 0.760, and mean absolute error = 0.499. Local modifications of orforglipron generated 1,701 valid, unique analogs. Composite prioritization incorporated predicted activity, quantitative estimate of druglikeness, synthetic accessibility, and database absence. All 146 queried structures were absent from ChEMBL and PubChem under the implemented identity criteria. Of these, 46 passed an exploratory physicochemical filter, and 20 were shortlisted. No generated candidate passed the stricter drug-likeness filter. The leading candidate, GLP1R-01162, had predicted pActivity 7.677 and composite score 0.5400. These results defined an auditable set of computational hypotheses; they did not establish agonism, experimental potency, synthetic feasibility, or global novelty. Functional testing and independent novelty assessment remained necessary.
Related
- A Computational Workflow Integrating Explainable AI and Database Mining for the Prioritization of Antimicrobial Natural Products
- Structural, supramolecular and computational insights into the ALS inhibitor pyroxsulam: crystal structure, reactivity and bioactivity
- Old Drugs, New Targets: Multi-parameter computational prioritization of FDA‑approved, brain‑penetrant drugs as isoform‑selective GSK‑3β inhibitors for Alzheimer’s disease
- From Computational Chemistry to Generative Models: A Survey of AI-Driven Small-Molecule Drug Discovery
- A century of structures: historical fidelity and computational fitness in the Inorganic Crystal Structure Database (ICSD)
BibTeX
@misc{mohan2026computational,
title = {{Computational design and prioritization of database-absent small-molecule GLP-1 receptor agonist candidates from public bioactivity data}},
author = {JOGU MOHAN},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010310/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010310/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Computational design and prioritization of database-absent small-molecule GLP-1 receptor agonist candidates from public bioactivity data
AU - JOGU MOHAN
PY - 2026
DA - 2026/10/12
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010310/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010310/v1
AB - Public bioactivity data were used to construct a reproducible workflow for the computational prioritization of small-molecule glucagon-like peptide-1 receptor (GLP-1R) agonist candidates. Exact IC50, EC50, and Ki measurements were standardized, converted to pActivity, and deduplicated after structural exclusion of peptide-like records. The resulting dataset comprised 337 compounds and 155 Murcko scaffolds. Morgan fingerprints and physicochemical descriptors were evaluated using ridge regression, random forest, and XGBoost with scaffold-based validation. Random forest provided the lowest mean five-fold cross-validation root-mean-square error (RMSE; 0.497 pActivity units). Its scaffold-disjoint holdout performance was R2 = 0.236, RMSE = 0.760, and mean absolute error = 0.499. Local modifications of orforglipron generated 1,701 valid, unique analogs. Composite prioritization incorporated predicted activity, quantitative estimate of druglikeness, synthetic accessibility, and database absence. All 146 queried structures were absent from ChEMBL and PubChem under the implemented identity criteria. Of these, 46 passed an exploratory physicochemical filter, and 20 were shortlisted. No generated candidate passed the stricter drug-likeness filter. The leading candidate, GLP1R-01162, had predicted pActivity 7.677 and composite score 0.5400. These results defined an auditable set of computational hypotheses; they did not establish agonism, experimental potency, synthetic feasibility, or global novelty. Functional testing and independent novelty assessment remained necessary.
ER -