ADMETlab 3.0 outputs & interpretation

You get back a per-compound table spanning a broad ADMET endpoint panel, organized across absorption, distribution, metabolism, excretion, and toxicity. These scores drive your next call: an organ-specific toxicity or CYP-interaction alert flags a liability to weigh against potency before you advance a compound, calibrating each decision to that endpoint's recommended cutoff.

Artifacts you receive

ADMETlab 3.0 returns predictions for up to 119 ADMET endpoints organized across absorption, distribution, metabolism, excretion, and toxicity categories. The model uses a multi-task Directed Message Passing Neural Network (DMPNN) architecture trained on over 400,000 data entries.

• CSV: Per-structure scores, confidences, and flags for all requested endpoints. Input columns are preserved.
• JSONL (optional): Machine-friendly rows with explicit error objects for programmatic processing.
• Job metadata: Model version, endpoint bundle, runtime logs, and parameter set.

Endpoint categories (119 total)

Absorption endpoints: Caco-2 permeability, HIA (human intestinal absorption), Pgp substrate/inhibitor, bioavailability, and solubility predictions.

Distribution endpoints: Plasma protein binding, volume of distribution, BBB penetration, and tissue distribution.

Metabolism endpoints: CYP450 substrate/inhibitor predictions (CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4), metabolic stability, and clearance.

Excretion endpoints: Renal clearance, half-life predictions, and biliary excretion.

Toxicity endpoints (31 new in v3.0): Comprehensive organ-specific toxicity including:

• Nephrotoxicity: Kidney damage potential
• Neurotoxicity: CNS toxicity risks
• Ototoxicity: Hearing damage potential
• Hematotoxicity: Blood cell toxicity
• Hepatotoxicity: Liver toxicity (DILI)
• Cardiotoxicity: hERG inhibition and cardiac liability
• Carcinogenicity, mutagenicity, and developmental toxicity

Reading results

• Binary classification endpoints: Probabilities between 0 and 1 with recommended cutoffs per endpoint. Higher values indicate greater likelihood of the predicted property.
• Regression endpoints: Numeric outputs (e.g., log values for solubility, clearance rates) plus confidence or applicability domain tags when available.
• Multi-class endpoints: Category predictions with associated probabilities for each class.
• Flags: rejected, invalid_smiles, and out_of_domain clarify why scores may be missing. Out-of-domain warnings indicate compounds outside the training set chemical space.

Exports and joins

• Download artifacts from the job page or programmatically via the artifact API.
• Join by your identifier column or original_index to merge back into screening tables.
• Keep the job URL as provenance in ELN or tickets.

Interpretation tips

• Calibrate decisions to each endpoint’s recommended cutoff; avoid reusing thresholds across tasks.
• Large swings across runs often tie to preprocessing differences—compare canonical SMILES.
• Use confidence bands to prioritize confirmatory assays for low-confidence predictions.

Canvas Visualization

When an ADMETLab3 job completes, a result card appears on the canvas.

• Card size: 350×300px
• Border color: Orange (#FF9800)
• Visualization: Property radar chart + detailed property table

Tier 2 Metrics

• Druglikeness score
• Lipinski violations count
• Bioavailability percentage
• Absorption / Distribution / Metabolism / Toxicity category flags

Interpreting Results

• Radar chart: Shows multi-dimensional property profile across ADMET categories
• Severity indicators: Green (favorable), Yellow (borderline), Red (unfavorable)
• 119 prediction endpoints: Grouped by Absorption, Distribution, Metabolism, Excretion, Toxicity
• Alerts: Highlighted for problematic properties (hERG inhibition, CYP interactions, mutagenicity)

Read next