(W1230-03-20) AI in Action: Accelerating Co-Crystal Design Using Machine Learning

Purpose: Artificial intelligence (AI) is transforming pharmaceutical development by enhancing data organization, integration, and application, ultimately leading to faster decision-making and process optimization. This study explores the implementation of machine learning techniques to improve co-crystal design, a critical aspect of solid form screening that influences the physicochemical properties of active pharmaceutical ingredients (APIs). The objective is to evaluate effectiveness of artificial intelligence, specifically machine learning models, in predicting co-crystal formation and optimizing co-crystal design to improve efficiency of experimental workflows during the early-phase solid form screening of pharmaceutical compounds drug development.

Methods: Three predictive models were developed to predict co-crystal formation using eXtreme Gradient Boosting (XGBoost) and message-passing neural networks (MPNN), trained on existing known positive and negative co-crystal pairs. While positive co-crystallization hits are readily accessible due to their frequent documentation in the literature, failures are rarely reported. As such, negative hits are underrepresented, leading to potential bias in model training. To mitigate this, we implemented an algorithmic strategy to synthesize negative samples, ensuring a more balanced and robust dataset. Input features included relevant molecular descriptors and fingerprints. The MPNN operates directly on molecular graph representations, allowing it to learn interatomic and topological features critical to molecular interactions in an unsupervised manner. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating curve (AUC-ROC). For validation, the models were evaluated using previously unseen API molecular structures that were excluded from the training dataset, to assess the generalization capability of the predictive framework. Experimental validation was designed with 10 APIs and 10 co-formers using slurry and liquid assisted grinding methods in 5 different solvents, and so far, experimental trials have been completed for 6 APIs, totaling 300 experiments. The selection of APIs and co-formers was guided by molecular size and functional group characteristics. Initial sample characterization was conducted using powder X-ray diffraction (PXRD), with subsequent confirmation by thermogravimetric analysis (TGA), differential scanning calorimetry (DSC), and NMR spectroscopy.

Results: For model training, 20000 datapoints were used including known positive hits and the simulated negative hits (8000). Both the XGBoost and MPNN models exhibited strong classification performance, with the MPNN model demonstrating superior predictive capability. As illustrated in Figure 1, the XGBoost model has high recall (0.92) and AUC (0.84) performance indicating strong sensitivity capability for predicting co-crystallization. Evaluation metrics such as ROC (AUC=0.83), precision-recall curves (0.84) and high true positives from confusion matrix confirmed robust classifier behavior for the MPNN. Also, the probability distribution plot revealed clear class separation at a decision threshold of 0.5 as shown in Figure 2. While some false positives were observed, likely influenced in part by mild class imbalance, the model maintains good performance. Eight novel solid forms were identified through experimental screening (300 combinations) and confirmed via PXRD, TGA, DSC, and NMR. PXRD was crucial not only for confirming the formation of new crystalline phases but also for verifying failed co-crystallizations, as diffraction patterns in those cases matched the API, co-former, or their physical mixtures. Solids with novel PXRD patterns were further characterized using TGA, DSC, and NMR. TGA confirmed the absence of solvate formation, while DSC provided insights into phase purity and revealed distinct melting transitions relative to the starting material. NMR was essential in confirming chemical identity and determining stoichiometry of the new solid forms. Figure 3 presents the experimental workflow, model performance metrics derived from training evaluations, and results from initial experimental validation. When applied to in-lab validation, an ensemble of all models was used for classification of co-crystallization probabilities which achieved an experimental accuracy of 82% and good model performance, correctly predicting 49 of 60 co-crystal outcomes. These results validate the model's capability to guide efficient co-former selection and reduce trial-and-error screening.

Conclusion: In summary, this work shows that machine learning significantly enhances co-crystal screening efficiency by enabling data-driven decision-making and efficient experimental planning. The results demonstrate strong classification capabilities for both XGBoost and MPNN models. Preliminary experimental validation further supports these findings, with the models achieving an accuracy of 82%, indicating reliable predictive performance. The integration of predictive models enables targeted selection of co-formers, significantly reducing experimental burden and resource consumption. The MPNN model, in particular, offers robust generalization across structural diversity. This data-driven approach promotes exhaustive exploration of the co-crystal landscape, facilitating identification of previously unrecognized solid forms. Such capability not only enhances intellectual property strategy by exhausting the co-crystal landscape, but also de-risks formulation and scale-up challenges in later stages of drug development. Ultimately, the use of machine learning supports more informed decision-making, efficient experimental planning, and facilitates efficient progression toward commercialization.

References: J. Chem. Inf. Model. 2019, 59, 8, 3370–3388

Acknowledgements: No conflicts of interest are declared.

Figure 1. XGBoost model performance metrics showing high recall (0.92), AUC (0.84), accuracy (0.76), precision (0.71), and F-1 score (0.80) indicating strong sensitivity capability for predicting co-crystallization

Figure 2. MPNN Model: Confusion matrix, probability distribution, ROC (AUC=0.83), and precision-recall curves (0.84) illustrating model performance, class separation and predictive reliability across the dataset

Figure 3. Screening of 300 API-coformer-solvent combinations; solid forms characterized by PXRD, TGA, DSC, and NMR. Tables summarize the performance metrics of the XGBoost and MPNN models and predictive model accuracy across six-APIs