Evaluating Molecular Complexity with Open-Source Machine Learning Approaches to Predict Process Mass Intensity.

Area of Science:

• Pharmaceutical Manufacturing
• Green Chemistry
• Computational Chemistry

Background:

• Process Mass Intensity (PMI) is a key metric for resource efficiency in pharmaceutical manufacturing.
• Accurate PMI target setting is challenging and crucial for sustainable practices.
• Existing machine learning tools show potential but require further development for broader application.

Purpose of the Study:

• To refine and expand the SMART-PMI tool into an open-source model and application.
• To enhance the prediction of molecular complexity for setting PMI targets.
• To promote explainability and informed decision-making in pharmaceutical process development.

Main Methods:

• Developed an open-source machine learning model based on four molecular descriptors: heteroatom count, stereocenter count, unique topological torsion, and connectivity index chi4n.
• Correlated molecular features to molecular complexity to predict PMI targets.
• Created a user-friendly application accepting SDF files for rapid complexity quantification and PMI target generation.

Main Results:

• The refined model achieved 82.6% predictive accuracy and a 0.349 RMSE for molecular complexity.
• The developed application provides accessible PMI targets to guide process development.
• The model emphasizes explainability and parsimony for better understanding and decision-making.

Conclusions:

• The open-source SMART-PMI tool offers a flexible and explainable approach to optimizing pharmaceutical manufacturing processes.
• This advancement supports the adoption of green chemistry principles and sustainable practices.
• The integrated tools facilitate data-driven process development for improved resource efficiency.