When Do Quantum-Chemical Descriptors Improve Aqueous Solubility Prediction? A Residual-Rescue Benchmark under Chemical Distribution Shift
- 1Stanford Online High School, 415 Broadway, Redwood City, California, 94063
Abstract
Quantum-chemical descriptors are increasingly appended to molecular machine-learning representations, yet their computational cost is difficult to justify when the average accuracy gain is small. We asked a stricter question for aqueous solubility: do low-cost quantum descriptors contain information that is genuinely nonredundant with conventional two-dimensional chemistry, and under which chemical regimes is that information useful? A frozen 2,996-molecule internal cohort derived from AqSolDB was evaluated with random, Bemis-Murcko scaffold-disjoint, and Butina-cluster-disjoint splits, and an independent 820-molecule OChemUnseen cohort was reserved for external validation. GFN2-xTB/ALPB-water descriptors produced small but reproducible direct improvements when added to RDKit descriptors (mean internal RMSE reductions 0.009-0.018 logS across model/split combinations; all 10 seeds improved). More importantly, a strict out-of-fold residual-rescue architecture, in which a quantum-only model learned errors left by the 2D model, improved RMSE by 0.028, 0.047, and 0.028 logS under random, scaffold, and cluster splits, respectively (10/10 seeds; paired Wilcoxon p = 0.001953 for each). On the external set, quantum augmentation reduced the 95th-percentile absolute error by 0.055 logS. Frontier-orbital descriptors accounted for most of the external gain, and significant benefits after false-discovery-rate correction were concentrated in nitro, amine, amide, and alcohol/phenol subsets. A learned low-cost gate did not reliably predict which molecules would benefit, providing a useful negative result. These results show that inexpensive quantum descriptors add modest average accuracy but measurable orthogonal information, with the clearest value under chemical distribution shift and in difficult or electronically distinctive cases.
Related
- How Molecular Weight Shapes the Relative Importance of Chemical and Structural Descriptors for Aqueous Solubility
- A Machine Learning Approach to logBB Prediction Combining Molecular Docking and Quantum-Chemical Descriptors
- Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
- MatUQ: a benchmark for uncertainty-aware out-of-distribution materials property prediction with graph neural networks
- PeptSol: A Translational Benchmark for Protected-Peptide Solubility in Drug Manufacturing
BibTeX
@misc{tsai2026when,
title = {{When Do Quantum-Chemical Descriptors Improve Aqueous Solubility Prediction? A Residual-Rescue Benchmark under Chemical Distribution Shift}},
author = {Aris Tsai},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009706/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009706/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - When Do Quantum-Chemical Descriptors Improve Aqueous Solubility Prediction? A Residual-Rescue Benchmark under Chemical Distribution Shift
AU - Aris Tsai
PY - 2026
DA - 2026/09/30
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009706/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009706/v1
AB - Quantum-chemical descriptors are increasingly appended to molecular machine-learning representations, yet their computational cost is difficult to justify when the average accuracy gain is small. We asked a stricter question for aqueous solubility: do low-cost quantum descriptors contain information that is genuinely nonredundant with conventional two-dimensional chemistry, and under which chemical regimes is that information useful? A frozen 2,996-molecule internal cohort derived from AqSolDB was evaluated with random, Bemis-Murcko scaffold-disjoint, and Butina-cluster-disjoint splits, and an independent 820-molecule OChemUnseen cohort was reserved for external validation. GFN2-xTB/ALPB-water descriptors produced small but reproducible direct improvements when added to RDKit descriptors (mean internal RMSE reductions 0.009-0.018 logS across model/split combinations; all 10 seeds improved). More importantly, a strict out-of-fold residual-rescue architecture, in which a quantum-only model learned errors left by the 2D model, improved RMSE by 0.028, 0.047, and 0.028 logS under random, scaffold, and cluster splits, respectively (10/10 seeds; paired Wilcoxon p = 0.001953 for each). On the external set, quantum augmentation reduced the 95th-percentile absolute error by 0.055 logS. Frontier-orbital descriptors accounted for most of the external gain, and significant benefits after false-discovery-rate correction were concentrated in nitro, amine, amide, and alcohol/phenol subsets. A learned low-cost gate did not reliably predict which molecules would benefit, providing a useful negative result. These results show that inexpensive quantum descriptors add modest average accuracy but measurable orthogonal information, with the clearest value under chemical distribution shift and in difficult or electronically distinctive cases.
ER -