Twenty-two case studies planned and run by PlayMolecule AI

PlayMolecule^® AI planned and ran 22 case studies. Each one started from a single prompt stating a discovery question. From that prompt the agent retrieved the public structures and activity data, decided which calculations to run, ran them, and reported what they showed; a scientist reviewed the retained outputs and logs before anything was published. Twenty-one address small-molecule discovery questions; one is a retrospective enzyme-engineering benchmark. The full reports for two of them are now published, alongside the prompt that started every study, on the PlayMolecule AI case studies page.

The published reports state the conditions of each calculation, the quality-control decisions taken during review, and the results that were excluded from the conclusions.

What the studies covered

The questions span oncology, neuroscience, immunology and immuno-oncology, metabolic and rare disease, fibrosis, and biocatalysis. The work behind them covers target assessment and go or no-go decisions, cryptic and allosteric pocket discovery, binding-mode prediction including covalent and macrocyclic cases, selectivity between close relatives, free-energy ranking under blinded conditions, membrane simulations that separate specific binding from lipid partitioning, cross-species pose transfer, and active-site engineering.

NAMPT: which computational steps held up

The NAMPT study tested four steps against public structures and disclosed inhibitors. Free redocking of the disclosed 4LTS ligand reached the crystal geometry, at 1.90 Å, and the score ranked that pose twentieth of 100. Ligand-free pocket prediction covered at least half the heavy atoms of all 180 retrospective ligand instances, while its discrete top-three centres recovered 59% of them.

Sampling and ranking separate on this target: the workflow reached the reference geometry and its score did not select it, so a pose picked on docking score alone stands as a hypothesis until an independent geometric check is applied. A comparison of two starting-hydration conditions, three replicas each, did not resolve a difference in mean ligand RMSD, which leaves starting-water choice an open setup variable at that sampling depth. For the extended NAMPT binding region, volume coverage is the pocket-prediction endpoint this benchmark supports.

ATA-117: recovering the hotspots of an evolved enzyme

The ATA-117 benchmark worked from the parent transaminase and public structures, blind to the evolved answer, and designed six active-site substitutions to accept a bulky non-native ketone. All six identities are present in ATA-117-Rd11, the experimentally evolved enzyme used in sitagliptin manufacture. The six-site combination was the only construct whose selected pose scored favourably under all three scoring families, which is consistent with epistasis among the substitutions.

Molecular dynamics bounded what that recovery means. Across three independent 50 ns replicas per construct the fold stayed stable, and no construct held a meaningful population of strict near-attack geometry. Run as a positive control, the same docking setup did not recover productive geometry for the active Rd11 enzyme. The evidence therefore supports locating engineering hotspots, not predicting catalytic rate.

The other twenty

Write-ups and data packages for the remaining studies are available on request to interested researchers. The index page names each target, the question it addressed and the prompt it started from. If one of them is close to something you are working on, or you would like the same workflow run on a program of your own, ask us for a walkthrough. With access to the underlying data through PlayMolecule AI, you can ask your own questions, follow the evidence and explore the aspects most relevant to your program.