ADMET-AI

What is ADMET-AI?

ADMET-AI is a machine learning model for predicting the drug-like properties of small molecules. Developed by Stanford University and Greenstone Biosciences and published in Bioinformatics in 2024, it provides fast and accurate predictions for Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET).

In essence, this is another ADMET model that helps researchers filter large chemical libraries and prioritize compounds that have a higher probability of success in clinical trials.

How does ADMET-AI work?

ADMET-AI uses a deep learning model called Chemprop-RDKit, which combines two powerful approaches:

• Graph Neural Networks (Chemprop): The model represents molecules as graphs, where atoms are nodes and bonds are edges. This allows it to learn complex relationships between a molecule's structure and its properties.
• Physicochemical Descriptors (RDKit): This graph-based representation is enhanced with 200 molecular features calculated by the RDKit toolkit, adding established chemical knowledge to the model.

The models were trained on 41 curated datasets from the Therapeutics Data Commons (TDC), the most comprehensive collection of standardized ADMET data available for machine learning.

Predicted properties

ADMET-AI consists of 41 predictive models, each trained on its own curated datasets, whcih allows for a detailed assessment of a compound's potential pharmacokinetic and toxicological profile.

Absorption

Good absorption is crucial for orally administered drugs. ADMET-AI assesses this through the following models:

• Oral bioavailability: Predicts the percentage of a drug that reaches systemic circulation after oral administration, a critical factor for determining effective dosage.
• Intestinal absorption (Caco-2): Models permeability across Caco-2 cells, a standard in vitro assay that mimics the human intestinal barrier.
• Aqueous solubility: Estimates a compound's solubility in water, which directly impacts its dissolution and subsequent absorption in the gut.

Distribution

Once absorbed, a drug must be distributed to its target tissues. ADMET-AI evaluates distribution with models for:

• Blood-Brain Barrier (BBB) Penetration: Predicts whether a compound is likely to cross the highly selective barrier protecting the central nervous system, which is essential for CNS-acting drugs but undesirable for others.
• Plasma Protein Binding: Estimates the degree to which a molecule will bind to proteins in the blood. High binding can limit the amount of free drug available to exert its therapeutic effect.
• P-glycoprotein (P-gp) Interaction: Classifies whether a compound is a substrate of P-gp, a key efflux transporter that can pump drugs out of cells, affecting both distribution and efficacy.

Metabolism

Metabolism determines how a drug is processed and cleared from the body. ADMET-AI covers this by predicting:

• Cytochrome P450 (CYP) Interactions: Includes models for inhibition of major CYP isoforms (1A2, 2C9, 2C19, 2D6, 3A4) and whether the compound is a substrate for CYP2D6 and 3A4. These interactions are a primary source of drug-drug interactions.
• Metabolic Stability: Predicts the half-life of a compound in human liver microsomes, indicating how quickly it is broken down by metabolic enzymes.

Excretion

This refers to how a drug is eliminated from the body. The platform provides models for:

• Half-Life (t½): Predicts the time it takes for the concentration of a drug in the plasma to be reduced by half.
• Renal Clearance: Estimates the rate at which a drug is cleared by the kidneys, a major route of elimination.

Toxicity

Early identification of potential toxicity is one of the most critical steps in drug discovery. ADMET-AI includes several widely recognized toxicity endpoints:

• hERG Inhibition: Predicts blockage of the hERG potassium channel, a key indicator of potential cardiotoxicity and risk of arrhythmias.
• Hepatotoxicity (Liver Toxicity): Assesses the potential for a compound to cause drug-induced liver injury.
• AMES Mutagenicity: Predicts the outcome of the Ames test, which screens for a compound's potential to cause DNA mutations and be carcinogenic.
• Acute Oral Toxicity (LD₅₀): Estimates the lethal dose (LD₅₀) of a substance, providing a measure of its short-term toxic potential.

Limitations

• Chemical space: The models are trained primarily on small molecules. Performance may be less reliable for macrocycles, natural products, or other molecules that are underrepresented in the training data.
• Need for validation: Predictions from any model, including ADMET-AI, are not a substitute for experimental validation. This tool should be used for screening and prioritization.

Cost

Using ADMET-AI through ProteinIQ costs 2 credits per molecule. This includes a complete analysis of all 41 ADMET properties.

---

Based on: Swanson, K., Walther, P., Leitz, J., Mukherjee, S., Wu, J.C., Shivnaraine, R.V., & Zou, J. (2024). ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics, 40(7), btae416. DOI: 10.1093/bioinformatics/btae416