Machine-Learning-Guided Cocrystal Prediction Based on Large Data Base

Abstract

A machine-learning model trained on the whole Cambridge Structural Database was developed to assist high-throughput cocrystal screening. With only 2D structures taken as inputs, the probability of cocrystal formation is returned for two given molecules. All of the cocrystal records in the CSD were used as positive samples, while negative samples were constructed by randomly combining different molecules into chemical pairs. Our model showed a prediction ability comparable with that of a widely used ab initio method in a head-to-head comparison test. Both experimental and virtual cocrystal screening against captopril were conducted at the same time to further validate the model. Two cocrystals of captopril with L-proline and sarcosine were obtained and characterized by PXRD, DSC, and FT-IR. These two coformers were also successfully predicted by our model. These results suggest that the tool we developed can be used to effectively guide coformer selection in the discovery of new cocrystals.

SMINBR: An Integrated Network and Chemoinformatics Tool Specialized for Prediction of Two-Component Crystal Formation

“…It is difficult to retrieve negative crystallization samples in the databases and the literature. Experimentally constructed negative samples will lead to the narrow applicability domain. , The generated artificial invalid molecular pairs lack real experimental evidence and inevitably contain positive samples that have not been confirmed, , which might affect the model accuracy. Although one classification model reported recently can build a predictive model without negative data, this model is still trained on a type of cocrystal system with a limited applicability domain .…”

“…11−13 To enlarge the applicability domain of such classification models, artificially invalid molecular combinations and cocrystals retrieved from Cambridge Structural Database are merged to train the machine learning or deep learning models. 14,15 The lattice energy landscapes of molecular combination and two pure components have been explored to determine cocrystal formation. 16−18 Hydrogen bond energies (HBE) between two molecules calculated via molecular electrostatic potential surfaces have also been utilized for cocrystal virtual screening.…”

“…In silico models that can predict a priori which coformers (interaction partners) are likely to form molecular interactions with a target molecule will provide a potential to reduce the laboratory time and costs. , To date, various types of in silico models have been developed to guide experimental cocrystal screening. Machine learning and related models have been conducted to discriminate cocrystal formation or not for a certain type of molecule. − To enlarge the applicability domain of such classification models, artificially invalid molecular combinations and cocrystals retrieved from Cambridge Structural Database are merged to train the machine learning or deep learning models. , The lattice energy landscapes of molecular combination and two pure components have been explored to determine cocrystal formation. − Hydrogen bond energies (HBE) between two molecules calculated via molecular electrostatic potential surfaces have also been utilized for cocrystal virtual screening. − In addition, hydrogen bond propensity (HBP), hydrogen bond coordination (HBC), molecular dynamics, and COSMO-RS , have been proposed to guide cocrystal screening.…”

Novel Cocrystals of Vonoprazan: Machine Learning-Assisted Discovery

“…Various ML models have been applied for cocrystal prediction, including support vector machines (SVMs) [30], multivariate adaptive regression splines [31], random forest (RF) [32], and network-based link-prediction [33]. In our previous work, we developed a virtual screening model based on an artificial neural network (ANN) algorithm for cocrystal prediction [34].…”

Cocrystal Prediction Using Machine Learning Models and Descriptors

“…Devogelaer and co-workers introduced a comprehensive approach to study cocrystallization using network science and linkage prediction algorithms and constructed a data-driven co-crystal prediction tool with co-crystal data extracted from the CSD [42]. Wang et al also used a data set with co-crystal data available in the CSD and ultimately developed a machine learning model using different model types and molecular fingerprints that can be used to select appropriate coformers for a target molecule [43]. The above existing studies have shown successful results, but they have a common limitation that they only compared model performance (e.g., accuracy) without investigating features (i.e., descriptors) importance.…”