ADMET Prediction: What Every Drug Developer Needs to Know

Why ADMET Is the Make-or-Break of Drug Development

You can design the most potent molecule ever seen in a binding assay, and it will still fail as a drug if the body cannot absorb it, if the liver destroys it in minutes, or if it poisons the heart. This is the ADMET problem – and it kills more drug candidates than any other single factor.

ADMET stands for Absorption, Distribution, Metabolism, Excretion, and Toxicity. Together, these five properties describe what the body does to a drug (pharmacokinetics) and what the drug does to the body (safety). Getting ADMET wrong is extraordinarily expensive: roughly 40% of drugs that enter clinical trials fail because of pharmacokinetic or toxicity issues, each failure costing hundreds of millions of dollars.

The good news? AI can now predict ADMET properties from nothing more than a molecular structure – a SMILES string. Instead of synthesizing a compound and running weeks of wet-lab assays, you can screen thousands of candidates computationally in minutes. This guide explains what each ADMET property means, why it matters, and how to predict them using SciRouter's API.

The Five Properties, Explained

A – Absorption

Absorption answers a deceptively simple question: does the drug get into the bloodstream? For oral drugs – the vast majority of medicines – this means crossing the intestinal wall. A compound with poor absorption will never reach its target, no matter how potent it is in a test tube.

Key metrics include Caco-2 permeability (how well a compound crosses a model of the intestinal epithelium), oral bioavailability (the fraction that reaches systemic circulation), and aqueous solubility (a compound that won't dissolve in the gut can't be absorbed). A concrete example: the antifungal drug itraconazole has excellent potency but notoriously poor solubility, requiring complex formulation strategies that add cost and limit patient compliance.

D – Distribution

Once absorbed, the drug must travel through the body to reach its target tissue. Distribution is governed by factors like plasma protein binding (drugs bound to albumin are inactive), volume of distribution (how widely the drug spreads beyond the bloodstream), and blood-brain barrier penetration(critical for CNS drugs, dangerous for drugs that should stay out of the brain).

Consider the difference between a cancer drug targeting a solid tumor versus one treating a brain tumor. The former needs to concentrate in peripheral tissue; the latter must cross the blood-brain barrier – one of the most selective membranes in the body. Getting distribution wrong means your drug accumulates in the wrong place.

M – Metabolism

The liver is the body's chemical processing plant, and cytochrome P450 (CYP) enzymes are its primary workforce. Metabolism determines how quickly a drug is broken down. Too fast, and the drug is cleared before it can work – requiring higher doses or more frequent dosing. Too slow, and the drug accumulates to toxic levels.

The most critical concern is CYP inhibition. If your drug inhibits a major CYP enzyme like CYP3A4 (which metabolizes roughly 50% of all marketed drugs), it will alter the metabolism of every other medication the patient takes. This is how grapefruit juice became a pharmacology footnote: it inhibits CYP3A4, causing dangerous interactions with statins, immunosuppressants, and dozens of other drugs.

E – Excretion

What goes in must come out. Excretion describes how the body eliminates the drug, primarily through the kidneys (renal clearance) and liver (biliary excretion). The key metric is half-life – how long it takes for the drug's concentration to drop by half. A drug with a 2-hour half-life needs dosing four times daily; one with a 24-hour half-life can be taken once a day.

Half-life also determines how quickly a drug washes out if problems arise. Drugs with very long half-lives (think of the anticoagulant warfarin) require careful monitoring because adverse effects persist long after the last dose.

T – Toxicity

Toxicity is the ultimate gatekeeper. A drug can have perfect pharmacokinetics and still fail if it damages organs. The key screens include hERG channel inhibition(blocking the hERG potassium channel causes cardiac arrhythmias – this ended the careers of multiple marketed drugs), hepatotoxicity (liver damage is the most common reason for post-market drug withdrawal), and Ames mutagenicity(whether the compound damages DNA, raising cancer risk).

The antihistamine terfenadine (Seldane) was pulled from the market after it was found to cause fatal cardiac arrhythmias via hERG inhibition – a property that modern AI models can now flag computationally before a single pill is manufactured.

How AI Predicts ADMET Properties

Traditional ADMET assessment requires synthesizing each compound and running it through a battery of in vitro assays: Caco-2 cells for permeability, microsomal stability for metabolism, patch-clamp electrophysiology for hERG. Each assay takes days and costs hundreds to thousands of dollars per compound.

AI models learn the relationship between molecular structure and ADMET properties from hundreds of thousands of experimental measurements. The approach works in three layers:

• Molecular representation – The SMILES string is converted into a numerical fingerprint or graph representation that captures the molecule's structural features
• Feature learning – Graph neural networks or transformer models learn which structural patterns correlate with each ADMET property (e.g., certain ring systems predict CYP inhibition)
• Ensemble prediction – Multiple models vote on each property, providing both a prediction and a confidence score

The training data comes from public databases like ChEMBL (containing millions of bioactivity measurements) and proprietary datasets from pharmaceutical companies. Modern models achieve 80–92% accuracy on classification tasks and strong correlation with experimental values for continuous properties.

Predicting ADMET Properties with the SciRouter API

SciRouter's /v1/pharma/adme endpoint accepts a SMILES string and returns a comprehensive ADMET profile. Let's walk through a complete example using aspirin and a more complex drug candidate.

Single Compound ADMET Prediction

import requests

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"

# Aspirin (acetylsalicylic acid)
response = requests.post(
    f"{BASE}/pharma/adme",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={"smiles": "CC(=O)Oc1ccccc1C(=O)O"}
)

result = response.json()
print(f"Compound: Aspirin")
print(f"Caco-2 Permeability: {result['absorption']['caco2_class']}")
print(f"Oral Bioavailability: {result['absorption']['bioavailability']}")
print(f"Plasma Protein Binding: {result['distribution']['ppb_percent']}%")
print(f"BBB Penetration: {result['distribution']['bbb_penetration']}")
print(f"CYP3A4 Inhibitor: {result['metabolism']['cyp3a4_inhibitor']}")
print(f"CYP2D6 Inhibitor: {result['metabolism']['cyp2d6_inhibitor']}")
print(f"Half-life Class: {result['excretion']['half_life_class']}")
print(f"hERG Inhibitor: {result['toxicity']['herg_inhibitor']}")
print(f"Ames Mutagenicity: {result['toxicity']['ames_mutagenicity']}")
print(f"Hepatotoxicity Risk: {result['toxicity']['hepatotoxicity']}")

The response is organized by ADMET category, making it straightforward to assess each property independently. Here is what the output looks like:

{
  "smiles": "CC(=O)Oc1ccccc1C(=O)O",
  "absorption": {
    "caco2_class": "high",
    "caco2_papp": 32.5e-6,
    "bioavailability": 0.68,
    "solubility_class": "soluble",
    "solubility_log_mol_L": -2.1
  },
  "distribution": {
    "ppb_percent": 87.3,
    "vd_L_kg": 0.15,
    "bbb_penetration": "yes"
  },
  "metabolism": {
    "cyp3a4_inhibitor": false,
    "cyp2d6_inhibitor": false,
    "cyp2c9_inhibitor": false,
    "microsomal_stability": "stable"
  },
  "excretion": {
    "half_life_class": "short",
    "clearance_mL_min_kg": 8.2,
    "renal_clearance": "moderate"
  },
  "toxicity": {
    "herg_inhibitor": false,
    "ames_mutagenicity": false,
    "hepatotoxicity": "low_risk",
    "ld50_mg_kg": 200
  },
  "confidence": 0.89
}

Batch ADMET Screening

In practice, you rarely evaluate one compound at a time. Here is how to screen a batch of candidates and filter for the best ADMET profiles:

import requests

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"

candidates = [
    {"name": "Aspirin", "smiles": "CC(=O)Oc1ccccc1C(=O)O"},
    {"name": "Celecoxib", "smiles": "Cc1ccc(-c2cc(C(F)(F)F)nn2-c2ccc(S(N)(=O)=O)cc2)cc1"},
    {"name": "Ibuprofen", "smiles": "CC(C)Cc1ccc(C(C)C(=O)O)cc1"},
    {"name": "Caffeine", "smiles": "Cn1c(=O)c2c(ncn2C)n(C)c1=O"},
]

results = []
for compound in candidates:
    resp = requests.post(
        f"{BASE}/pharma/adme",
        headers={"Authorization": f"Bearer {API_KEY}"},
        json={"smiles": compound["smiles"]}
    )
    data = resp.json()
    results.append({
        "name": compound["name"],
        "admet": data
    })

# Filter: keep only compounds with no hERG risk and high absorption
safe_candidates = [
    r for r in results
    if not r["admet"]["toxicity"]["herg_inhibitor"]
    and r["admet"]["absorption"]["caco2_class"] == "high"
]

print(f"Passed ADMET filter: {len(safe_candidates)}/{len(candidates)}")
for c in safe_candidates:
    print(f"  {c['name']}: bioavailability={c['admet']['absorption']['bioavailability']:.2f}, "
          f"hERG_safe=True, hepatotox={c['admet']['toxicity']['hepatotoxicity']}")

Interpreting ADMET Results: What Good Looks Like

Raw numbers are only useful if you know what they mean. Here is a practical guide to interpreting the key ADMET predictions:

Absorption Benchmarks

• Caco-2 Papp > 8 × 10¹²³ cm/s – High permeability. The compound crosses the intestinal wall readily.
• Bioavailability > 0.5 – More than half the oral dose reaches systemic circulation. This is the target for most oral drugs.
• Solubility > 60 µg/mL – The compound dissolves adequately in gastrointestinal fluid. Below this, formulation challenges increase significantly.

Distribution Benchmarks

• Plasma protein binding < 90% – Acceptable. Above 95%, very little free drug is available to act on the target.
• BBB penetration = "yes" – Essential for CNS drugs. For non-CNS drugs, BBB penetration may increase the risk of neurological side effects.

Metabolism Benchmarks

• CYP inhibition = false for all major isoforms – The compound is unlikely to cause drug-drug interactions. CYP3A4 inhibition is the most critical flag.
• Microsomal stability = "stable" – The compound is not rapidly degraded by liver enzymes. Unstable compounds may need prodrug strategies or non-oral formulation.

Toxicity Red Flags

• hERG inhibitor = true – Serious cardiac safety concern. This alone can kill a drug program. Medicinal chemistry optimization to remove hERG liability is standard practice.
• Ames positive – The compound may cause DNA mutations. Requires careful evaluation; some Ames-positive compounds proceed if the benefit-risk ratio is high (e.g., cancer drugs).
• Hepatotoxicity = "high_risk" – Liver damage potential. The most common cause of post-approval drug withdrawals.

Integrating ADMET into Your Drug Discovery Pipeline

ADMET prediction is most powerful when integrated early in the discovery process. Here is a recommended workflow:

• Hit identification – Use molecular properties to check Lipinski's Rule of Five and basic drug-likeness
• Hit-to-lead – Screen hits with ADMET prediction to eliminate compounds with fatal ADMET liabilities before investing in synthesis
• Lead optimization – Iteratively modify leads, re-predicting ADMET after each structural change to track how modifications affect the pharmacokinetic profile
• Candidate selection – Compare final candidates across all ADMET dimensions to select the compound with the best overall profile for preclinical development

For a deeper understanding of the molecular representations used as input, see our complete guide to SMILES notation.

Common ADMET Pitfalls

Even with good predictions, there are traps that catch experienced drug developers:

• Chasing potency at the expense of ADMET – A 1 nM IC50 means nothing if the compound has zero oral bioavailability. Balance potency optimization with ADMET monitoring from day one.
• Ignoring metabolite toxicity – The parent compound may be safe, but its metabolites can be toxic. Acetaminophen is a classic example: the parent drug is safe at therapeutic doses, but a CYP-generated metabolite (NAPQI) causes liver failure in overdose.
• Over-reliance on single metrics – A compound that passes every individual ADMET filter can still fail if the overall profile is mediocre across the board. Look at the holistic picture.
• Applicability domain – AI models are most accurate for compounds structurally similar to their training data. Novel scaffolds or unusual chemotypes may produce unreliable predictions. Always check the confidence score.

Next Steps

ADMET prediction transforms drug discovery from a serial, expensive process into a computational screening funnel. By filtering compounds early, you focus synthesis and wet-lab resources on candidates with the best chance of clinical success.

To get started, calculate molecular properties for your compounds, then run them through the ADMET endpoint to build a complete pharmacokinetic profile. If you are working with molecular representations for the first time, our SMILES notation guide will help you encode your structures correctly.

Ready to profile your compounds? Sign up for a free SciRouter API key and start screening today.