Explainable Machine Learning for Efficient Cocrystal Prediction: Leveraging Morgan Fingerprints to Decode Local...

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Cocrystallization is crucial for tuning molecular material properties but faces experimental bottlenecks.
• Existing machine learning (ML) models for virtual screening lack interpretability and rely on global descriptors, hindering rational design.
• Developing efficient and interpretable ML tools is essential for advancing cocrystal engineering.

Purpose of the Study:

• To develop a high-performance, interpretable machine learning framework for rapid cocrystal screening.
• To utilize Morgan fingerprints for capturing local chemical environment information in molecular structure encoding.
• To establish a direct link between ML predictions and underlying chemical mechanisms for rational cocrystal design.

Main Methods:

• Constructed a large-scale cocrystal dataset of 7700 samples.
• Employed Morgan fingerprints to encode molecular structures, preserving critical functional group and substructure information.
• Developed and optimized five ML models (MF-KNN, MF-RF, MF-XGBoost, MF-SVM, MF-ANN), with MF-ANN achieving state-of-the-art performance.
• Applied SHapley Additive exPlanations (SHAP) for model interpretability at both global and local levels.

Main Results:

• The MF-ANN model demonstrated high accuracy (97.16%) and F1-score (98.35%) on an independent test set.
• Robustness was confirmed through extensive cross-validation (10-fold, 5-fold, 3-fold).
• SHAP analysis identified key substructures driving cocrystal formation and correlated predictions with specific molecular interactions like hydrogen bonding and π-π stacking.

Conclusions:

• The developed ML framework offers a high-performance and interpretable solution for accelerating cocrystal screening.
• This approach provides actionable insights into chemical mechanisms, facilitating rational design of functional cocrystals.
• The interpretable ML models are applicable to diverse fields including pharmaceuticals, energetic materials, and optoelectronics.