Compounds that look promising on potency fail on absorption, metabolism, or toxicity after significant investment.
A major driver of discovery attrition is candidates that are potent but fail on pharmacokinetics or toxicity — often after considerable time and spend. Catching these liabilities earlier saves both. AI approaches predict ADMET and toxicity properties computationally so teams can deprioritize risky chemistry before committing to synthesis and assays.
How AI helps predict ADMET and toxicity earlier
ADMET-prediction tools estimate absorption, distribution, metabolism, excretion, and toxicity properties from a molecule's structure before it is ever synthesized. Trained on public and proprietary assay data, they score candidate chemistry for liabilities — poor solubility, metabolic instability, likely off-target toxicity, cardiac or hepatic risk — so teams can deprioritize the riskiest structures before committing synthesis and assay resources. The practical value is triage rather than certainty. A model that flags a probable metabolic soft spot or a structural alert lets chemists redesign around it early, and lets project teams weigh potency against developability when ranking a series. The prediction is a hypothesis for the bench to confirm — qualified scientists still run the assays and own the go/no-go decision.
What to evaluate before buying ADMET prediction AI
The deciding factor is whether the model's training data resembles your chemistry. A tool built largely on drug-like small molecules may extrapolate poorly to novel scaffolds, macrocycles, or your proprietary series, so ask how the model behaves on chemistry outside its training distribution and whether it reports confidence or applicability-domain estimates alongside each prediction. A number with no uncertainty attached is hard to act on. Also probe how the tool handles the endpoints you actually care about — different toxicity endpoints are predicted with very different reliability — and whether you can retrain or calibrate on your own assay results. Explainability matters too: a structural alert a chemist can inspect is more useful than an opaque score.
How teams typically get started
A low-risk entry point is retrospective benchmarking: run compounds with known assay outcomes — ideally including some of your own discontinued chemistry — through the tool and compare its predictions against what the lab measured. That tests the model's relevance to your chemical space before it influences any live project, and builds internal confidence in which endpoints to trust and which to treat with caution.
AI Use Cases That Address This Problem
ADMET & Toxicity Prediction
Hit Identification & Lead Optimization
Frequently asked questions
What is ADMET and why do candidates fail on it?
ADMET covers absorption, distribution, metabolism, excretion, and toxicity — the properties that determine whether a potent molecule can actually become a usable medicine. Candidates that look strong on target potency can still fail because they are poorly absorbed, cleared too quickly, or carry toxicity liabilities, and these problems are expensive when they surface late.
Can AI predict toxicity reliably?
It depends heavily on the endpoint and the chemistry. Some properties are predicted well because there is abundant, consistent training data; others remain difficult, and model outputs should be treated as risk flags rather than verdicts. Reliable use means checking predictions against your own assay results and understanding where the model is extrapolating beyond what it has seen.
Does in silico ADMET replace laboratory assays?
No. These tools are designed to help teams decide which compounds are worth taking into the lab, not to replace measurement. The assays remain the source of truth; the value of prediction is reducing how many low-value molecules consume synthesis and screening capacity before that measurement happens.
What should we ask an ADMET prediction vendor?
Ask which endpoints the model covers and how it was validated for each, how it performs on chemistry unlike its training set, whether predictions come with confidence or applicability-domain estimates, and whether you can calibrate the model on your own data. Request a retrospective benchmark on compounds with known outcomes before committing.
Eli Lilly collaboration worth up to $670 million to discover novel therapies
Genesis Therapeutics (now Genesis Molecular AI) entered a strategic collaboration with Eli Lilly to discover novel therapies for up to five targets across a range of therapeutic areas using its Genesis Molecular AI drug-discovery platform, receiving $20 million upfront with a potential total value of up to $670 million.
Nimbus TYK2 program (TAK-279) acquired by Takeda for $4 billion upfront
A selective TYK2 inhibitor discovered using Schrödinger's physics-based computational platform (developed by Nimbus Therapeutics) was acquired by Takeda for $4 billion upfront, demonstrating commercial validation of a computationally designed clinical asset.
Eli Lilly AI and robotics drug-discovery collaboration
XtalPi and Eli Lilly announced a collaboration to apply XtalPi's AI and robotics-driven drug discovery platform to uncover first-in-class small-molecule therapeutics.
Cube Biotech collaboration to design novel Amylin Receptor agonists
Iktos and Cube Biotech announced a strategic collaboration to discover novel small-molecule agonists of the Amylin Receptor, pairing Iktos's AI-driven generative design platform with Cube Biotech's membrane protein production and purification technologies.
Pfizer collaboration expanded for next-generation molecular modeling
XtalPi announced an expansion of its research collaboration with Pfizer to build a next-generation molecular modeling platform for drug discovery, applying XtalPi's AI and robotics capabilities to improve the accuracy of computational predictions.
Pierre Fabre Laboratories oncology drug-discovery collaboration
Pierre Fabre Laboratories and Iktos announced an integrated drug-discovery collaboration to identify and develop novel small-molecule candidates in oncology, with Iktos applying its AI-driven generative design platform alongside Pierre Fabre's medicinal chemistry, biology, and preclinical development.