Advanced optimization of drug solubility through statistical analysis using machine learning

Predictive modeling of scientific systems is often difficult due to limited data, impaired bias due to shifts in the data, and commonly applied techniques used to model data for generalized solutions. A proof of concept for the aforementioned problem is provided by utilizing a new approach to data analysis and modeling development that results in more accurate solubility predictions from a physics-constrained, hybrid machine learning (ML) and statistical modeling method with added value of embedded constraints into a well-defined, flexible model. A quantitative evaluation of the methodology discussed will be conducted via interpolation, extrapolation, and noise scenarios using realistic datasets to evaluate the performance of the framework. Traditional modeling methods have provided accurate predictions but exhibited unstable and non-physical behaviors during the extrapolation and interpolation process. In contrast, the proposed approach is able to predict solubility accurately, maintain a physical relationship between pressure and solubility, and produce more accurate and robust predictions while prioritizing predictive utility instead of statistical accuracy. A predefined λ weighting factor will adjust the extent to which the methodology incorporates data-driven methodology versus physics-based guided solutions into the models.