ADMET: The evaluation of pharmacokinetics and toxicity is crucial for the design of new therapeutic candidates. In silico virtual screens and generative AI output a vast number of molecules that must be filtered to a tractable number for synthesis and experimental validation. An effective primary filter is to evaluate candidate compounds based on their Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties.
ADMET-AI: ADMET-AI is a simple, fast, and accurate web interface for predicting the ADMET properties of molecules using machine learning models.
Models: ADMET-AI predicts ADMET properties using a graph neural network architecture called Chemprop-RDKit (see the Chemprop package for details). ADMET-AI's Chemprop-RDKit models were trained on 41 ADMET datasets from the Therapeutics Data Commons (TDC). ADMET-AI’s Chemprop-RDKit models have the highest average rank on the TDC ADMET Benchmark Group leaderboard. ADMET-AI is also currently the fastest web-based ADMET predictor.
References: The ADMET-AI code can be found at github.com/swansonk14/admet_ai, and ADMET-AI is described in detail in this paper: ADMET-AI: A machine learning ADMET platform for evaluation of large-scale chemical libraries. Please cite us if ADMET-AI is useful in your work.
Molecules: ADMET-AI can make predictions for up to 1,000 molecules at a time by (1) providing SMILES (one per line) in the text box, (2) uploading a CSV file with SMILES, or (3) drawing a molecule using the interactive tool and then converting it to a SMILES.
DrugBank Reference: To provide relevant context, ADMET predictions on input molecules are compared to predictions on 2,579 approved drugs from the DrugBank. This reference set can be filtered to a specific category of drugs for a more relevant comparison by selecting an Anatomical Therapeutic Chemical (ATC) code.
Predict: After selecting input molecules and a DrugBank reference set, click the “Predict” button to make ADMET predictions.
Summary Plot: The summary plot shows the distribution of ADMET predictions for all input molecules compared to the DrugBank reference set. The x- and y-axes can be changed to show any two ADMET properties.
Radial Plot: For each input molecule, a radial plot is shown summarizing five key ADMET properties in terms of their DrugBank percentile:
ADMET Predictions: Clicking on each displayed molecule will show the molecule’s ADMET predictions in tabular form. For each molecule, ADMET-AI computes 8 physicochemical properties using RDKit and predicts 41 ADMET properties using its Chemprop-RDKit graph neural networks. For regression properties, the property is directly predicted with the units shown. For classification properties, the predicted value is the probability that the molecule has the property (e.g., probability of blood-brain barrier penetration). Additionally, the percentile of the molecule’s property compared to the DrugBank reference is shown. Note: For regression properties, the displayed value is clipped to a valid range (e.g., ≥0 for half life) while the downloaded results contain the original value.
Download Predictions: Predictions for the first 25 molecules are shown on the website. Predictions for all input molecules can be downloaded as a CSV file by clicking the “Download Results” button.
Storing Molecules: ADMET-AI does not store any molecules uploaded to the website. All molecules are deleted after predictions are made.
ADMET-AI can be run locally as a command line tool for large-scale batch prediction or as a Python module for use within other Python-based drug discovery tools. Please see github.com/swansonk14/admet_ai for more details.
Input up to 1,000 molecules to make ADMET predictions.
Create a DrugBank reference set to provide context for ADMET predictions.
Choose an Anatomical Therapeutic Chemical (ATC) code to select a subset of DrugBank approved drugs or choose "all" for all approved drugs.
Selected ATC code "all" includes 2,579 molecules.