From molecular uncertainty
to clinical confidence.

The Predictor begins by running the candidate SMILES through six independent ADMET prediction tools — ADMETlab 3, Deep-PK, pkCSM, ADMET-AI, PHD Predictor, and SwissADME. Each tool returns its own estimate of absorption, distribution, metabolism, excretion, and toxicity; each one has its own biases, its own training data, its own blind spots. Taken alone, any one of them is a noisy guess.

A Bayesian model then treats each tool’s output as one observation in a joint probability distribution and marginalizes across methods to recover a posterior that is independent of any single tool’s uncertainty.

A downstream machine-learning layer, trained on clinical data from FDA-approved drugs, reads that distribution and returns the expected ADMET properties with calibrated uncertainty. The result is a single optimal assessment — the collective intelligence of the industry’s best tools, anchored in clinical reality.

Inputs

Candidate SMILES, ingested by six independent ADMET prediction tools.

Method

Bayesian marginalization over the six tool outputs, refined by an ML layer trained on FDA-approved drugs.

Output

Expected ADMET properties with calibrated mean and uncertainty intervals — seven ADMET and BBB endpoints in the demo.

Most industry methods translate ADMET predictions into human context using generalized physiological assumptions — assumptions that break down for novel chemical matter where the relationship between structure and in vivo PK is not yet understood.

The Scaler replaces that guesswork with empirical calibration. Using the Predictor’s posterior, the engine screens the US FDA drug library and identifies clinical cousins — existing therapeutics that share a high degree of structural (SMILES) and ADME similarity to your candidate, within a thirty-percent variance window. Published clinical data for those cousins tunes the scaling factors for the most sensitive parameters, producing a human-calibrated model of how the body will handle your compound.

Mechanism

Matches the candidate against the US FDA drug library to find structurally and pharmacokinetically similar approved drugs — within 30% variance.

Replaces

Theoretical guesswork — the generalized physiological assumptions industry tools apply to novel chemistry.

Result

Scaling factors tuned against the real clinical data of the matched cousin drugs.

In the final stage, the engine constructs a high-fidelity PBPK/PD model and runs it inside a virtual clinical trial environment. The pharmacokinetic model is projected in time, organ by organ; the pharmacodynamic layer attaches a custom response curve specific to the therapeutic endpoint — tumor shrinkage, receptor occupancy, biomarker flux, or a bespoke mechanism.

Because the scaling factors were tuned in Module II, the simulation reflects how a real human — not a generalized average — will respond. Dosage forms, administration routes, and physiological variability are all configurable per program.

PK

Traces how the drug moves through the body after an IV, oral, or injected dose.

PD

Predicts what the drug actually does — shrink a tumor, hit a receptor, shift a biomarker.

Compartments

Focuses on specific organs — brain, liver, kidney — for local effects.