Skip to content
Discovery Stage 09 · Clinical Outcomes

Only 1 in 10 compounds that enter Phase I clears it

NovoExpert-3 returns a calibrated Phase I clearance probability with SHAP explanations, on demand, the kind of model that used to live only inside large pharma. You see which risk factors to address before committing to the clinic.

“Predict Phase I clinical clearance probability for this candidate and explain the top risk factors.”

0.756
AUROC
63
Input features
11,952
Training compounds
SHAP
Explained
The funnel

How it works

01

Submit a profiled candidate

Provide a SMILES with ADMET results from upstream stages. NovoExpert-3 uses 63 features, physicochemical, CYP, toxicity, Tox21, and composite scores, all already computed by the pipeline.

02

Calibrated prediction with explanations

Isotonic calibration makes the probability reliable, a true likelihood rather than a raw model score. SHAP feature attribution identifies the top risk factors driving the prediction.

03

Competence-checked

Out-of-domain predictions are flagged automatically. If the compound falls outside the model's validated domain, you get an explicit warning in place of false confidence.

Proof

NovoExpert-3: CatBoost classifier trained on TDC Trial Outcome Prediction dataset. 11,952 compounds. AUROC 0.756 in validated domains.

63 features: physicochemical (8), CYP inhibitor/substrate (8), toxicity (12), Tox21 NR (7), Tox21 SR (5), composite scores (3), all computed by upstream ADMET pipeline.

Isotonic calibration. SHAP feature attribution. Competence check flags out-of-domain predictions.

Research preview

Know before the clinic

Calibrated clearance probability. SHAP explained. Sign up and predict your first candidate.