Crystal structure prediction for flexible organic molecules: complete landscapes with sub-kJ/mol free-energy ranking
Difficulty 56: generation and ranking each have visible paths and recent progress, but sub-kJ/mol free-energy ranking of flexible molecules is not demonstrated blind. Application value 72: polymorph risk is a major pharmaceutical cost, comparable to the best-valued items in the hardware packs. Leverage 46.5: it reshapes how solid-state free energies and ML potentials are validated.
The problem
Predicting every way a floppy drug molecule can pack into a crystal, and which is stable, from its drawing alone.
- 1Floppy molecule, many packings
- 2Generate candidate crystals
- 3Energies differ by few kJ/mol
- 4Temperature changes ranking
- 5Blind tests measure progress
Task. From a 2D molecular diagram, predict the observable crystal forms of a flexible molecule. This involves (1) generating candidate structures (random or evolutionary searches over cell parameters, orientations and conformers), (2) ranking them by free energy, and (3) judging kinetics and disorder.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Machine-learned interatomic potentials with calibrated, guaranteed error bounds outside the training distributionComputational Chemistry · D60
- Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised trainingComputational Chemistry · D54
- Step 2
- Step 3
- A transferable exchange–correlation functional free of delocalization and self-interaction errorComputational Chemistry · D72
- Step 4
- Transferable ML potentials with non-local charge transfer, variable charge/spin states and long-range electrostaticsComputational Chemistry · D52
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Step 2
- Step 3
- Free-energy methods and enhanced samplingAlgorithm · host
- Then
- Crystal structure prediction for flexible organic molecules: complete landscapes with sub-kJ/mol free-energy rankingThe problem
State of the art
- Blind tests. The seventh CCDC blind test (Acta Cryst. B, 2024; two papers on structure generation and structure ranking; 22 groups for ranking) found that many methods reproduced the structure of a small flexible agrochemical, few succeeded on the more complex targets (copper complex, cocrystal, flexible drug candidate, salt), periodic DFT-D methods broadly matched experiment within expected error, an AIMNet-based machine-learned approach matched in many cases, and two groups predicted disorder blindly for the first time. The released database holds 171,679 structures from 207 landscapes.
- Free energies. Kapil and Engel (PNAS 2022) show that finite-temperature and anharmonic effects can reverse polymorph stability orders computed from static lattice energies. Cruz-Cabeza, Feeder and Davey (Commun. Chem. 2020) catalogue open questions about polymorph occurrence.
- Beyond DFT energies (preprint). Shi, Herman, Della Pia, Kapil, Zen and Nagy (arXiv:2603.02180, March 2026, preprint) report a correlated wavefunction plus many-body-expansion workflow with lattice energies within experimental uncertainty for 23 organic molecules and 13 ice polymorphs, and sub-chemical accuracy for relative energies of drug-like polymorphs including axitinib and ROY at cost close to hybrid DFT; this has not been through blind testing.
- Open. A blind, prospective demonstration across flexible drug-like molecules with free-energy ranking and disorder, and agreement between independent workflows, is still missing; the search space for large, flexible molecules is still incompletely sampled.
Realistic applications
Late-appearing polymorphs have caused pharmaceutical withdrawals (ritonavir) and cost the industry heavily in patents and bioavailability; computational polymorph screening would reduce experimental screening and enable the design of organic semiconductors, explosives and agrochemicals with the desired packing.
What counts as solved
A blind-test result (CCDC or equivalent) in which at least one method predicts all experimentally observed forms within the top 5 ranked structures, including free-energy ordering, for at least five drug-like targets with 6+ rotatable bonds, with independent reproduction on a following round.
How it gets verified
ObservationSettled by new measurements or observations.
Settled by a blind test (CCDC or equivalent) against newly determined experimental crystal forms: at least one method must place all observed forms in its top 5, with correct free-energy ordering, for at least five drug-like targets with six or more rotatable bonds, and reproduce this in a following round.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Compute-heavy generation and ranking; the new part is accurate, affordable energy models and sampling.
Bottlenecks
ranked by votes- #1TheorySub-kJ/mol finite-temperature free energies
Anharmonic, nuclear quantum and entropic contributions need ML or wavefunction-quality methods at scale.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Mar 2, 2026Sub-chemical-accuracy lattice energies (preprint)source
Shi et al. report correlated wavefunction-quality lattice energies for drug-like crystals.
- Oct 1, 2024Seventh blind test publishedsource
Generation and ranking results published in Acta Cryst. B (date approximate).
- Mar 1, 2022Finite-temperature stabilitysource
Kapil and Engel show that anharmonic and quantum effects can reorder polymorph stabilities (date approximate).
- Oct 1, 2020Open questions in polymorphismsource
Cruz-Cabeza, Feeder and Davey set out what is unknown about organic polymorphism (date approximate).