Predicts in vivo pharmacokinetics from structure alone: plasma clearance, steady-state volume, and residence time in four preclinical species.
No providers recorded yet. Browse all providers
A medicinal chemist asking how long a molecule will survive in an animal usually does not measure it. They incubate it with liver microsomes or hepatocytes, watch it disappear, and push that intrinsic clearance through the well-stirred scaling equation — in vitro to in vivo extrapolation, or IVIVE. It is mechanistic and interpretable, and it captures hepatic metabolism and nothing else: not renal excretion, not transporter effects, not the rest of what an intact organism does to a drug. On mouse and rat plasma clearance it tracks the animal outcome at a Pearson r of roughly 0.5. Two endpoints that matter as much — tissue distribution and how long a molecule stays intact — have no in vitro proxy at all, being properties of a whole body.
This model predicts those endpoints directly from structure. insitro built it on the proprietary preclinical corpus of Eli Lilly and Company — decades of standardized ADME assays and IV pharmacokinetic studies on hundreds of thousands of unique molecules — and it reached Lilly TuneLab member companies on 6 October 2026. It sits beside ChemLab, Lilly's own in vitro ADMET model, in the TuneLab catalog, but it is a different artifact from a different builder: ChemLab reports assay-level readouts, while this model reports the three outputs of an intravenous PK study in four species.
What makes the in vivo heads trainable is transfer inside a single network. Full IV PK studies number in the thousands; in vitro ADME measurements on the same chemistry run far higher. A shared encoder trained against both learns the physicochemistry driving all of them, and the in vivo heads inherit a representation a few thousand PK studies could not support alone.
The architecture is a Chemprop multitask directed message-passing neural network over raw molecular graphs, with one multi-layer perceptron head per endpoint across the in vivo PK targets and the in vitro ADME and physicochemical endpoints. Splits came from Morgan fingerprints clustered by sphere exclusion with RDKit's LeaderPicker, whole clusters allocated to train, validation, and test at a target 75/15/10 ratio across all endpoints at once; every measurement on a molecule travels with it into the same split. The validation set drove early stopping and best-checkpoint selection, the test set only final metrics. On that held-out set, all twelve species-endpoint combinations reach Pearson correlations of roughly 0.6–0.8, with up to 75% of predictions within twofold of the measured value and 69–91% within threefold; steady-state volume of distribution in dog and monkey reaches 90% or better within threefold. On mouse and rat plasma clearance the model attains r of approximately 0.54–0.62, against 0.39–0.45 for raw in vitro intrinsic clearance and about 0.5 for IVIVE. These are insitro's internal figures. Parameter count and per-assay dataset sizes are undisclosed, and per-endpoint model cards are visible only inside TuneLab.
The deployment target is compound prioritization during hit-to-lead and lead optimization at small and mid-size biotechs — companies with preclinical programs but not the decades-long assay archive such a model requires. A team ranks a virtual library or a proposed analogue series on predicted clearance, distribution volume, and residence time, synthesizes the survivors, and sends fewer compounds into IV PK studies. Needing no assay input, the predictions arrive earlier in the design cycle than any in vitro result can.
Pharmacokinetic failure is one of the main reasons small-molecule candidates die, and it usually surfaces after substantial time and capital are committed, so moving an in vivo signal before synthesis changes where a program spends its effort. It also makes the TuneLab premise concrete: data no outside group could reconstruct, made usable without either side's molecules changing hands. Nothing has been externally validated or peer reviewed, no code or checkpoint is public, access is contingent on TuneLab membership, and the reported generalization is to held-out structural clusters inside one company's chemistry. As insitro puts it, the true test of generalization will be performance on new chemistry.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.