Conformal Prediction and Multi-Ensemble Gradient Boosting for Subtype-Selective Binding Affinity Estimation Across Human Adenosine Receptors
- 1Department of Medicinal Chemistry and Computer-Aided Drug Design
Abstract
Subtype selectivity across human adenosine GPCRs remains an intractable medicinal chemistry challenge. The four orthosteric binding pockets share over 70% sequence homology across transmembrane helices III, V, VI, and VII. Consequently, standard QSAR models fail prospectively. They suffer from systemic scaffold leakage, overestimating held-out affinity while producing point estimates that lack calibrated error bars. We built an open-source, leak-free computational platform to solve both failure modes. The architecture combines XGBoost gradient boosting with MAPIE Jackknife+ cross-conformal prediction, Random Forest, and LightGBM, trained on 9,589 curated ChEMBL v34 and GPCRdb bioactivity records. Under strict Bemis-Murcko scaffold partitioning (N_train = 6,332; N_test = 1,583), the ensemble achieved an overall R2 of 0.693 and MAE of 0.390 pChEMBL units. On active compounds alone (N_test = 3,771, structural decoys removed), accuracy reached an overall R2 of 0.865 and MAE of 0.314. Per-subtype R2 values reached 0.753 for A1, 0.884 for A2A, 0.912 for A2B, and 0.886 for A3. Conformal intervals delivered 85.80% empirical coverage at a 90% nominal confidence level. Uncertainty quartiles scaled monotonically with absolute prediction error. A GINE graph neural network trained on identical scaffold splits managed only R2 = 0.248 overall, demonstrating that curated physicochemical descriptors decisively outperform deep graph convolutions in low-to-medium data regimes. Twenty-fold Y-randomization confirmed genuine structure-activity relationships, with all permuted R2 values falling below zero (p < 0.001). External blind validation on 15 novel GPCRdb ligands yielded a 75% selectivity recall accuracy. TreeSHAP features attributions verified that model decisions follow interpretable electrostatic and steric properties. All source code, curated data splits, model weights, and interactive deployment are publicly available.
Related
- Molecular Property Prediction with Well-Calibrated Uncertainties via Graph Neural Networks and Conformal Quantile Regression
- A Study on Lithium-Ion Battery Health Estimation and Remaining Life Prediction Based on Real-World Operational Data
- Novel Triazole-Adenosine Conjugates as Mammalian Seryl-tRNA Synthetase Inhibitors to protect cells from Ischemia-Reperfusion Injury
- Fluorescence Sensing of Anions Like Adenosine 5ʹ‐Triphosphate With pH‐Gated Porous Dye‐Loaded Cerasomes
- The Design and Application of Diureido Functionalized Spiroligomers as Diphosphate Receptors
BibTeX
@misc{patel2026conformal,
title = {{Conformal Prediction and Multi-Ensemble Gradient Boosting for Subtype-Selective Binding Affinity Estimation Across Human Adenosine Receptors}},
author = {Utkarsh Patel},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008546/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008546/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Conformal Prediction and Multi-Ensemble Gradient Boosting for Subtype-Selective Binding Affinity Estimation Across Human Adenosine Receptors
AU - Utkarsh Patel
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008546/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008546/v1
AB - Subtype selectivity across human adenosine GPCRs remains an intractable medicinal chemistry challenge. The four orthosteric binding pockets share over 70% sequence homology across transmembrane helices III, V, VI, and VII. Consequently, standard QSAR models fail prospectively. They suffer from systemic scaffold leakage, overestimating held-out affinity while producing point estimates that lack calibrated error bars. We built an open-source, leak-free computational platform to solve both failure modes. The architecture combines XGBoost gradient boosting with MAPIE Jackknife+ cross-conformal prediction, Random Forest, and LightGBM, trained on 9,589 curated ChEMBL v34 and GPCRdb bioactivity records. Under strict Bemis-Murcko scaffold partitioning (N_train = 6,332; N_test = 1,583), the ensemble achieved an overall R2 of 0.693 and MAE of 0.390 pChEMBL units. On active compounds alone (N_test = 3,771, structural decoys removed), accuracy reached an overall R2 of 0.865 and MAE of 0.314. Per-subtype R2 values reached 0.753 for A1, 0.884 for A2A, 0.912 for A2B, and 0.886 for A3. Conformal intervals delivered 85.80% empirical coverage at a 90% nominal confidence level. Uncertainty quartiles scaled monotonically with absolute prediction error. A GINE graph neural network trained on identical scaffold splits managed only R2 = 0.248 overall, demonstrating that curated physicochemical descriptors decisively outperform deep graph convolutions in low-to-medium data regimes. Twenty-fold Y-randomization confirmed genuine structure-activity relationships, with all permuted R2 values falling below zero (p < 0.001). External blind validation on 15 novel GPCRdb ligands yielded a 75% selectivity recall accuracy. TreeSHAP features attributions verified that model decisions follow interpretable electrostatic and steric properties. All source code, curated data splits, model weights, and interactive deployment are publicly available.
ER -