Data-Driven Prediction of Pharmaceutical Solid Forms Across Polymorphism, Crystallization, Molecular Packing, Stability, and Developability Decisions

Download PDF

Pharmaceutical solid-form development increasingly relies on data-driven and physics-informed models to anticipate crystal structures, polymorph landscapes, multicomponent forms, crystallization behavior, stability, and downstream developability, yet these tasks are often discussed together despite representing different prediction targets, evidence levels, and decision consequences. This systematic review evaluates how current methods connect molecular or process data to solid-state predictions and how far those predictions can support pharmaceutical decisions. A structured search and screening strategy was applied, covering molecular-crystal machine learning, crystal structure prediction, polymorphism, solvates, salts, cocrystals, nucleation, stability, phase transformation, crystal habit, and developability. Study information was coded by material class, target, representation, algorithm, validation design, experimental confirmation, uncertainty source, and transferability, while risk-of-bias appraisal adapted general prediction-model principles with crystal-specific concerns including structural leakage, generated versus experimental labels, split strategy, condition dependence, and external validation. Heterogeneous endpoints were synthesized without indiscriminate quantitative pooling. The literature resolves into distinct layers: solid-form propensity classification, crystal packing and polymorph ranking, process and nucleation modeling, finite-temperature or environment-dependent stability, dynamic transformation modeling, and developability-oriented decision support. The strongest evidence connects computational prediction to blind, external, or experimental validation, whereas major weaknesses arise when model confidence, benchmark accuracy, computed stability, or packing plausibility is treated as equivalent to experimentally realized form behavior. Data-driven solid-form prediction is therefore most defensible as a staged evidence problem rather than a single performance problem, and a proposed task-to-decision hierarchy is used to preserve boundaries between prediction, physical validation, experimental realization, and pharmaceutical consequence.