The Drug Crystallizes However It Wants

Many drug molecules can crystallize into multiple distinct solid forms (polymorphs) with different physical properties — solubility, dissolution rate, stability, and bioavailability. During manufacturing scale-up, a new, unexpected polymorph can appear spontaneously, converting the intended crystal form into one with wrong dissolution properties, potentially rendering entire batches non-bioequivalent. The phenomenon is unpredictable: polymorphic transitions can occur years into commercial production, triggered by subtle changes in impurity profiles, mixing conditions, or even the surfaces of manufacturing equipment. Current computational crystal structure prediction (CSP) methods can identify thermodynamically plausible polymorphs but cannot reliably predict which form will crystallize under specific process conditions.

The ritonavir case (1998) is the most famous example: Abbott Laboratories' HIV protease inhibitor spontaneously converted to a previously unknown, less-soluble polymorph two years into commercial production, forcing withdrawal of the oral capsule formulation and costing an estimated $250 million. Similar (though usually less dramatic) polymorph surprises occur across the pharmaceutical industry — an estimated 50–80% of drug molecules exhibit polymorphism, and 30% of development programs encounter an unexpected form during scale-up or commercial production. Regulatory agencies (FDA, EMA) require demonstration that a drug's crystal form is controlled and consistent, making polymorph control a critical quality attribute.

Polymorph screening campaigns (high-throughput crystallization from many solvents, temperatures, and conditions) attempt to identify all possible forms before scale-up, but completeness is never guaranteed — new forms can appear after thousands of experiments. Computational CSP (using DFT-D and global lattice energy minimization) has improved dramatically but still produces lists of 10–100 plausible structures ranked by energy, without predicting which will actually nucleate under process conditions. Process analytical technology (PAT) — in-situ Raman, FTIR, and X-ray diffraction in crystallizers — can detect polymorph conversion during manufacturing but only after it has begun, often too late to intervene. Seeding with the desired polymorph improves control but doesn't prevent conversion if the undesired form is thermodynamically more stable.

Reliable prediction of nucleation kinetics — not just which polymorphs are possible (thermodynamics) but which will form under specific conditions (kinetics). This requires understanding crystal nucleation at the molecular level, which remains one of the most fundamental unsolved problems in physical chemistry. Machine learning models trained on large crystallization datasets could potentially learn empirical relationships between process parameters and polymorph outcomes. Continuous crystallization (rather than batch) offers tighter control of supersaturation and temperature profiles, potentially reducing the window for polymorphic conversion.

A team could conduct a systematic crystallization study of a model polymorphic compound (e.g., glycine, carbamazepine, sulfathiazole — well-characterized polymorphic systems) across a matrix of solvents, temperatures, and cooling rates, correlating process conditions with polymorph outcomes. A computational team could benchmark CSP methods (CrystalPredictor, AIRSS) against known polymorphic systems and identify systematic failure modes. Both approaches use standard equipment and published reference data.