Machine learning models can accurately predict pharmaceutical solubility in supercritical carbon dioxide to support industrial process design, but their effectiveness depends heavily on rigorous controls against target leakage and validation bias. A new evaluation of small, heterogeneous chemical datasets demonstrates that while algorithmic models achieve strong predictive accuracy when interpolating within previously represented drug compounds, their predictive reliability breaks down entirely when applied to unseen molecules.
The investigation examined an experimental dataset comprising 196 discrete observations covering multiple distinct pharmaceutical drugs. These measurements spanned an operating temperature range of 308 to 343 Kelvin, system pressures between 9 and 30 megapascals, and carbon dioxide solvent densities ranging from 326.09 to 924.56 kilograms per cubic meter.
To eliminate data contamination across model training stages, the researchers constructed a strictly leakage-free input feature space. This input space incorporated the physical and chemical properties of the drugs and solvent, including molecular weight, melting temperature, operating temperature, system pressure, and carbon dioxide density. In addition to these fundamental properties, the feature space contained five pre-specified engineered mathematical transformations: density raised to the power of 1.5, pressure raised to the power of 2.8, the product of temperature and pressure, the ratio of temperature to pressure, and density multiplied by the natural logarithm of density.
Model optimization and training were governed by strict data separation protocols to prevent optimistic evaluation bias. The researchers split the 196 observations into a development partition and an independent test partition using a 90-to-10 ratio under a fixed random seed of 5. All hyperparameter selection was conducted exclusively within the development data, ensuring that the untouched 20-point test set remained isolated from tuning. Furthermore, feature scaling procedures for support vector regression and multilayer perceptron artificial neural networks were fitted exclusively on training folds.
Across the investigation, eight distinct machine learning approaches were evaluated, including a limited particle-swarm-optimized Random Forest model alongside standard tree-based and neural algorithms. When evaluated on the untouched 20-point test partition, the Decision Tree model recorded the highest coefficient of determination, reaching an R-squared of 0.955 and a root mean square error of 0.217 in the natural logarithm of solubility. However, the Decision Tree model exhibited extreme overfitting, exactly fitting the development data. In comparison, the particle-swarm-optimized Random Forest, AdaBoost, and standard Random Forest models yielded test R-squared values of 0.920, 0.917, and 0.911, respectively.
To establish more dependable performance estimates, the researchers implemented repeated nested validation across the development pipeline. Under this rigorous resampling protocol, AdaBoost achieved a mean outer-fold R-squared value of 0.899 with a standard deviation of 0.112. The standard Random Forest architecture demonstrated comparable stability, producing a mean outer-fold R-squared value of 0.908 with a standard deviation of 0.062 across the repeated validation folds.
In sharp contrast to the strong metrics observed in random test splits, leave-one-compound-out validation produced negative pooled R-squared values across the evaluated algorithms. This stark failure confirmed that high accuracy achieved during standard data splitting merely reflects interpolation among familiar compounds. Consequently, the authors established that interpolation success within represented drug families cannot be interpreted as evidence of reliable solubility prediction for unseen chemical compounds.
To determine the physical drivers behind model predictions, SHAP analysis was conducted for the leakage-free Random Forest using solely the independent test set. Molecular weight and melting temperature emerged as the two most influential input features governing solubility predictions, followed closely by the interaction terms of temperature multiplied by pressure and temperature divided by pressure. These empirical findings demonstrate that current machine learning architectures serve effectively as interpolation tools within the specific chemical and thermodynamic domains studied, while emphasizing that larger and chemically diverse external datasets are required before models can achieve genuine compound-level generalization.
Subject of Research: Machine learning modeling of pharmaceutical solubility in supercritical carbon dioxide
Article Title: ​Modeling drug solubility in supercritical CO₂via machine learning
Article References: Sayed, U., Alhasso, B., Priya, G. P., Sahoo, S., Srivastav, S. K., Rani, A., Kumar, N., Nwr, H. K., & Abolhekmat, M. (2026). ​Modeling drug solubility in supercritical CO₂via machine learning. Discover Materials. https://doi.org/10.1007/s43939-026-01067-7
Image Credits: AI Generated
DOI: 10.1007/s43939-026-01067-7
Keywords: supercritical carbon dioxide, drug solubility, machine learning, particle swarm optimization, random forest, decision tree, adaboost, support vector regression, multilayer perceptron, target leakage, nested validation, leave-one-compound-out validation, SHAP analysis, molecular weight, melting temperature, carbon dioxide density, process design, feature engineering
News Source: Teresa Odom. (November 12, 2026). Machine Learning Interpolates Known Drug Solubility in Supercritical Carbon Dioxide. Scienmag.



