ADMET Prediction for Drug Candidates

Version Compatibility

Reference examples tested with: RDKit 2024.09+, requests 2.31+, DeepChem 2.8+, chemprop 2.0+ (note major API change from 1.x), admet-ai 1.3+, pandas 2.2+.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

ADMET Prediction

Predict absorption, distribution, metabolism, excretion, and toxicity properties of drug candidates. ADMET prediction underpins lead selection and de-risking; calibrated, applicability-domain-aware predictions distinguish a working filter from a costly false-confidence rejection. Modern best practice combines online services (ADMETlab 3.0 with uncertainty estimates), open-source models (chemprop D-MPNN), and rule-based filters (Lipinski / Veber / BBB heuristics) -- each with known failure modes.

For PAINS / Brenk / structural alerts, see chemoinformatics/substructure-search. For QSAR model building from in-house data, see chemoinformatics/qsar-modeling.

ADMET Model Taxonomy

Decision: For batch screening with no in-house data, ADMETlab 3.0 provides 119 reported platform features and uncertainty for modeled endpoints through its web service and hosted API; verify the live API documentation before automating access. For a sufficiently large, relevant in-house endpoint dataset, benchmark a chemprop D-MPNN, descriptors, and simpler baselines under a deployment-relevant split rather than assuming a universal sample-size threshold. Shan et al. (2022) reported an AUC of 0.956 for a D-MPNN combined with 206 MOE descriptors on their random-split hERG benchmark.

Decision Tree by Scenario

OECD QSAR Principles (5 Pillars)

For regulatory-grade ADMET QSAR (REACH, ECHA, FDA submissions), models must satisfy:

1. Defined endpoint -- specific bioassay, units, conditions
2. Unambiguous algorithm -- reproducible model + code
3. Defined applicability domain (AD) -- where the model is valid
4. Appropriate statistical validation -- external test set, cross-validation
5. Mechanistic interpretation -- biological / chemical rationale

For non-regulatory work, AD assessment is still critical. The OECD's applicability domain is the workhorse: predictions outside the AD are unreliable, but operational AD measures (leverage, kNN, conformal prediction) often disagree.

Applicability Domain Methods

For deep-learning ADMET, conformal prediction can provide calibrated prediction sets or intervals when its exchangeability and calibration assumptions are appropriate (McShane et al. 2024).

ADMETlab 3.0 API

ADMETlab 3.0 reports 119 platform features: 77 prediction models, 34 computed physicochemical properties, and 8 medicinal-chemistry rules. The modeled endpoints include prediction uncertainty; do not imply that computed properties and rules have model uncertainty.

Goal: Obtain the platform's 119 features for a batch of SMILES, including uncertainty for the 77 modeled endpoints, using the hosted service.

Approach: Follow the live ADMETlab 3.0 API tutorial to wash molecules, submit batch predictions, and retrieve the returned results. The 2024 paper documents API/batch support and modeled-endpoint uncertainty. Obtain current rate limits, routes, payloads, task identifiers, and output contracts from the live official documentation rather than attributing them to the paper or hard-coding an unofficial example.

import pandas as pd

# After submitting with the current official API example, load its CSV output.
results = pd.read_csv('admetlab3_results.csv')
# Preserve the uncertainty columns and task identifier in downstream reports.

ADMETlab 3.0 endpoints (sample):

• Absorption: Caco-2 permeability (logPapp), HIA (%), Pgp inhibitor/substrate, MDCK
• Distribution: BBB+, PPB (%), VDss (L/kg), Fu (fraction unbound)
• Metabolism: CYP1A2/2C9/2C19/2D6/3A4 inhibitor / substrate
• Excretion: CL (mL/min/kg), T1/2 (h)
• Toxicity: hERG, AMES, hepatotoxicity (DILI), carcinogenicity, immunotoxicity, mutagenicity, respiratory, skin, eye, cardiotoxicity, mitochondrial, NR-AR, NR-ER, SR-MMP
• Drug-likeness: Lipinski, Veber, Ghose, Egan, Muegge, QED, SAscore

chemprop D-MPNN for Custom Endpoints

When in-house data is available, train a target-specific model. chemprop provides a widely used open-source D-MPNN architecture with atom/bond features and optional molecular descriptors; benchmark it against appropriate baselines on the project's data.

Goal: Train a target-specific ADMET classifier or regressor on in-house bioassay data.

Approach: Use the installed Chemprop 2.x CLI with a scaffold split, a release-supported descriptor featurizer, replicated models, and an explicit prediction-time uncertainty/calibration workflow.

# Chemprop 2.2 CLI; verify flags against the installed release.
# chemprop train --data-path data.csv --task-type classification \
#                --save-dir model_dir --split-type scaffold_balanced \
#                --molecule-featurizers v1_rdkit_2d_normalized \
#                --num-replicates 5 --ensemble-size 5
# chemprop predict --test-path test.csv --model-paths model_dir \
#                  --uncertainty-method ensemble --preds-path predictions.csv

# Or chemprop 2.x programmatic API (full programmatic API documented at chemprop.readthedocs.io)
# See chemoinformatics/qsar-modeling for the full chemprop 2.x training pipeline.

Key: Replicates and an ensemble estimator produce an uncertainty estimate, not automatic calibration. Fit and evaluate a documented calibrator on a separate calibration set when calibrated probabilities or intervals are required. Descriptor benefit must be demonstrated on the intended endpoint.

hERG Cardiotoxicity Endpoint

hERG (KCNH2) blockade can contribute to QT prolongation and Torsades de Pointes and is an important non-clinical cardiac-safety endpoint. Follow the current ICH S7B/E14 and regulator-specific guidance applicable to the program rather than treating one model output as a regulatory conclusion.

Interpretation: A single-model probability > 0.5 is NOT a kill signal. Triangulate multiple hERG-specific models and a literature search. ProTox-3.0 calls the endpoint cardiotoxicity, but its model specifically predicts small-molecule hERG blockers; it should not be treated as an independent non-hERG mechanism. Consider exposure relative to measured hERG potency and confirm important decisions experimentally rather than applying a universal safe/unsafe IC50 cutoff.

CYP Inhibition (DDI Risk)

5 CYP isoforms cover most clinically relevant DDIs: