Abstract
Background
Reliable prediction of pharmaceutical solubility in supercritical carbon dioxide (SC-CO
2
) can support particle engineering and formulation design. However, record-wise validation may overestimate generalization when measurements for the same compound occur in both the training and test sets.
Methods
A leakage-aware machine-learning workflow was evaluated using molecular weight, melting point, temperature, and pressure as predictors of drug solubility in SC-CO
2
. Experimental records compiled from published measurements were screened using Isolation Forest and scaled by min-max normalization. Kernel ridge regression (KRR), histogram-based gradient boosting (HGB), and neural additive models (NAMs), together with linear regression and multilayer perceptron baselines, were evaluated using a 90/10 record-wise split and leave-one-compound-out (LOCO) validation. Hyperparameters were tuned using Hyperband.
Results
HGB provided the strongest overall performance. Under the random record-wise test split, it achieved R² = 0.99372, RMSE = 0.247192 g/L, and MAE = 0.113038 g/L. Under the more stringent LOCO evaluation, its performance decreased to R² = 0.93128, RMSE = 0.498217 g/L, and MAE = 0.271894 g/L. The difference between the two validation schemes demonstrates that random splitting yields an optimistic estimate when compound-specific observations are shared across subsets.
Conclusion
HGB accurately correlated solubility within the investigated data distribution and retained useful predictive ability for compounds excluded from model training. Nevertheless, external validation and additional molecular descriptors are required before extrapolating the model to chemically distinct pharmaceuticals.