A fast and efficient machine learning assisted prediction of urea and its derivatives to screen crystal propensity with experimental validation

Predicting crystal propensity is crucial yet challenging in various industries where it significantly influences product stability, performance, and efficacy. Predicting a crystal propensity identifies their optimal chemical structures for desired properties including solubility, bioavailability, shelf-life stability etc. Herein, A machine learning (ML) assisted analysis is performed to predict their crystal propensity by collecting a dataset of 6000 non-crystalline and over 200 crystalline urea and its derivatives. The data is trained by employing a Support Vector Machine (SVM) with its Radial Basis Function (RBF) and linear kernels along with Random Forest regression analysis. The trained data is compared with four other ML models, including Linear Regression, Gradient Boosting, Random Forest and Decision Tree Regressions to predict their crystal propensity. It yields an accuracy of 79% for identifying their non-crystalline compounds and 59 % in predicting crystallization failure. Their dimensionality reduction via t-SNE reveals their distinct clustering patterns to underscore their complex interplay between molecular structure and crystal propensity. Their experimental validation also corroborates the current findings to demonstrate their efficacy to streamline their crystal engineering for pharmaceutical formulation-based workflows. Notably, the number of rotatable bonds and molecular connectivity index (chi(v)(o)) emerges as pivotal descriptors for enabling their accurate classification with minimal input features. This study elucidates its quantitative structure-crystallinity relationship to provide a valuable tool for crystal design and optimization.