Exploring celecoxib polymorph landscape using AIMNet2 machine learning interatomic potential.

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Materials Science

Background:

• Drug crystal form (polymorphism) significantly impacts dissolution, efficacy, and stability.
• Predicting polymorphs of flexible molecules is a major challenge due to subtle energy differences (<2 kJ mol⁻¹).
• Accurate prediction requires high computational cost, often limiting quantum chemistry approaches.

Purpose of the Study:

• To develop and apply a machine-learned interatomic potential (AIMNet2) for efficient and accurate crystal structure prediction.
• To map the polymorphic landscape of celecoxib, a flexible COX-2 inhibitor.
• To identify novel low-energy polymorphs and assess the limitations of static-lattice models for flexible crystals.

Main Methods:

• Utilized AIMNet2, a machine-learned potential refined by active learning on cluster reference data.
• Employed a GPU-accelerated workflow to generate and rank hundreds of thousands of candidate structures.
• Performed hybrid-DFT calculations and finite-temperature analyses for validation and deeper insights.

Main Results:

• Achieved near-quantum accuracy in predicting the polymorphic landscape of celecoxib.
• Recovered the experimental ordering of celecoxib forms I, II, and III with high geometric fidelity.
• Identified two novel low-energy candidate structures within 4 kJ mol⁻¹ of the most stable polymorph.
• Highlighted the limitations of static-lattice models for ultra-soft crystals like celecoxib form I.

Conclusions:

• AIMNet2 provides a computationally efficient and accurate framework for crystal structure prediction of flexible drug molecules.
• The study offers a transferable strategy for polymorph screening and discovery.
• Findings advance the understanding of polymorphism in pharmaceutical science and materials design.