Admet Prediction

<!--
# COPYRIGHT NOTICE
# This file is part of the "Universal Biomedical Skills" project.
# Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu>
# All Rights Reserved.
#
# This code is proprietary and confidential.
# Unauthorized copying of this file, via any medium is strictly prohibited.
#
# Provenance: Authenticated by MD BABU MIA

-->

---
name: bio-admet-prediction
description: Predicts ADMET properties using ADMETlab 3.0 API or DeepChem models. Estimates bioavailability, CYP inhibition, hERG liability, and 119 toxicity endpoints with uncertainty quantification. Filters for PAINS and other structural alerts. Use when filtering compounds for drug-likeness or prioritizing leads by predicted safety.
tool_type: python
primary_tool: ADMETlab
measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes.
allowed-tools:
  - read_file
  - run_shell_command
---

# ADMET Prediction

Predict absorption, distribution, metabolism, excretion, and toxicity properties.

## ADMETlab 3.0 API

ADMETlab 3.0 provides 119 endpoints with uncertainty estimates.

```python
import requests
import pandas as pd

def predict_admet_batch(smiles_list, api_url='https://admetlab3.scbdd.com/api/predict'):
    '''
    Predict ADMET properties using ADMETlab 3.0 API.

    Note: SwissADME has NO API - it is web-only.
    '''
    payload = {
        'smiles': smiles_list
    }

    response = requests.post(api_url, json=payload)
    response.raise_for_status()

    return pd.DataFrame(response.json())

# Example usage
# smiles = ['CCO', 'c1ccccc1O', 'CC(=O)Oc1ccccc1C(=O)O']
# results = predict_admet_batch(smiles)
```

## Key ADMET Endpoints

| Category | Endpoints | Thresholds |
|----------|-----------|------------|
| Absorption | Caco-2, HIA, Pgp substrate | HIA > 30% |
| Distribution | BBB penetration, PPB, VDss | BBB+: penetrates |
| Metabolism | CYP inhibition (1A2, 2C9, 2C19, 2D6, 3A4) | Inhibitor threshold |
| Excretion | Clearance, Half-life | - |
| Toxicity | hERG, AMES, hepatotoxicity, carcinogenicity | hERG IC50 > 10 μM |

## DeepChem Models

DeepChem supports both PyTorch and TensorFlow backends.

```python
import deepchem as dc

# Load pre-trained toxicity model
tox21_tasks, tox21_datasets, transformers = dc.molnet.load_tox21()
train_dataset, valid_dataset, test_dataset = tox21_datasets

# Featurize new molecules
featurizer = dc.feat.CircularFingerprint(size=1024)
smiles = ['CCO', 'c1ccccc1']
features = featurizer.featurize(smiles)

# Load trained model
model = dc.models.GraphConvModel(
    n_tasks=12,
    mode='classification',
    model_dir='tox21_model'
)

# Predict (after training/loading)
# predictions = model.predict_on_batch(features)
```

## PAINS Filter

```python
from rdkit.Chem.FilterCatalog import FilterCatalog, FilterCatalogParams

def filter_pains(molecules):
    '''
    Filter out PAINS (pan-assay interference compounds).
    These are promiscuous compounds that give false positives in assays.
    '''
    params = FilterCatalogParams()
    params.AddCatalog(FilterCatalogParams.FilterCatalogs.PAINS)
    catalog = FilterCatalog(params)

    clean = []
    flagged = []

    for mol in molecules:
        if mol is None:
            continue
        entry = catalog.GetFirstMatch(mol)
        if entry is None:
            clean.append(mol)
        else:
            flagged.append((mol, entry.GetDescription()))

    print(f'Clean: {len(clean)}, PAINS flagged: {len(flagged)}')
    return clean, flagged

# Other filter catalogs available:
# FilterCatalogs.BRENK - Brenk structural alerts
# FilterCatalogs.NIH - NIH structural alerts
# FilterCatalogs.ZINC - ZINC clean leads
```

## Lipinski and Beyond

```python
from rdkit import Chem
from rdkit.Chem import Descriptors, Lipinski, QED

def calculate_druglikeness(mol):
    '''
    Calculate multiple drug-likeness criteria.
    '''
    if mol is None:
        return None

    props = {
        # Lipinski Rule of 5
        'MW': Descriptors.MolWt(mol),
        'LogP': Descriptors.MolLogP(mol),
        'HBD': Lipinski.NumHDonors(mol),
        'HBA': Lipinski.NumHAcceptors(mol),

        # Additional properties
        'TPSA': Descriptors.TPSA(mol),
        'RotatableBonds': Lipinski.NumRotatableBonds(mol),
        'AromaticRings': Lipinski.NumAromaticRings(mol),

        # QED (quantitative estimate of drug-likeness)
        # 0-1 scale, > 0.5 generally drug-like
        'QED': QED.qed(mol)
    }

    # Lipinski violations
    violations = 0
    if props['MW'] > 500: violations += 1
    if props['LogP'] > 5: violations += 1
    if props['HBD'] > 5: violations += 1
    if props['HBA'] > 10: violations += 1
    props['LipinskiViolations'] = violations

    # Veber criteria (oral bioavailability)
    # RotatableBonds <= 10, TPSA <= 140
    props['VeberCompliant'] = (props['RotatableBonds'] <= 10 and props['TPSA'] <= 140)

    return props
```

## Prioritization Pipeline

```python
def prioritize_compounds(molecules):
    '''
    Multi-stage ADMET filtering pipeline.
    '''
    results = []

    for mol in molecules:
        if mol is None:
            continue

        props = calculate_druglikeness(mol)
        if props is None:
            continue

        # Stage 1: Lipinski filter
        if props['LipinskiViolations'] > 1:
            continue

        # Stage 2: Additional filters
        if not props['VeberCompliant']:
            continue

        # Stage 3: QED cutoff
        if props['QED'] < 0.5:
            continue

        results.append((mol, props))

    return results
```

## Related Skills

- molecular-descriptors - Calculate descriptors for ML
- substructure-search - Filter reactive groups
- virtual-screening - Screen after ADMET filtering


<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->