Computational modeling and optimized machine learning for multidrug solubility prediction in supercritical carbon dioxide.
TL;DR
Sparse, leakage-controlled calibration is established as a more defensible use case than unconditional prediction for unseen drugs, while identifying the need for richer molecular descriptors and external prospective validation.
Abstract
Reliable prediction of pharmaceutical solubility in supercritical carbon dioxide (scCO2) is complicated by strong compound-specific behavior and response values. This study investigated whether sparse measurements from a newly introduced drug can calibrate transferable machine-learning predictions for computational estimation of solubility in wide range of parameters. A curated workbook containing 1,618 observations for 60 drugs was modeled using molecular weight, melting point, temperature, and pressure. Solubility was transformed as log₁₀(S), and XGBoost, CatBoost, and multilayer perceptron regression were each tuned using tree-structured Parzen estimation, differential evolution, and particle swarm optimization. Evaluation used five-fold nested group cross-validation, with complete drugs held out from training and tuning. For each test drug, predictions were assessed before calibration and after 1, 3, or 5 randomly selected measurements, repeated ten times. Within the present evaluation framework, the lowest-log-MAE zero-shot configuration was XGBoost optimized by tree-structured Parzen estimation (log-MAE = 0.584; R² = -0.087; RMSE = 1.713 g/L), showing that the four descriptors were insufficient for reliable extrapolation to unseen compounds. Sparse calibration substantially improved transfer: the lowest-log-MAE one-shot result was obtained by CatBoost-particle swarm optimization (log-MAE = 0.418; R² = 0.166), whereas XGBoost-tree-structured Parzen estimation had the lowest log-MAE at three shots (log-MAE = 0.298; R² = 0.317) and at five shots (log-MAE = 0.251; R² = 0.452; RMSE = 1.223 g/L). Median absolute relative deviation decreased from 82.4% to 42.9% between zero and five shots. These results establish sparse, leakage-controlled calibration as a more defensible use case than unconditional prediction for unseen drugs, while identifying the need for richer molecular descriptors and external prospective validation. In practical terms, this few-shot strategy can reduce the experimental burden for a newly introduced drug by using a small screening set to calibrate a global model before mapping the remaining pressure-temperature domain.