ADMET-AI

This git repo contains the code for ADMET-AI, an ADMET prediction platform that uses Chemprop models trained on ADMET datasets from the Therapeutics Data Commons (TDC). ADMET-AI can be used to make ADMET predictions on new molecules via the command line, via the Python API, or via a web server. A live web server hosting ADMET-AI is at admet.ai.greenstonebio.com

Please see the following paper and this blog post for more details, and please cite us if ADMET-AI is useful in your work.

ADMET-AI: A machine learning ADMET platform for evaluation of large-scale chemical libraries

Versions

This repo contains ADMET-AI v2, whereas our paper and the ADMET-AI live web server are based on ADMET-AI v1. The main difference is that ADMET-AI v2 uses Chemprop v2 (without RDKit fingerprints), which is faster and is compatible with more packages (e.g., recent versions of PyTorch). Note that since the models were retrained from scratch, the predictions of ADMET-AI v2 will not exactly match those of ADMET-AI v1. If you still need to use ADMET-AI v1 or want to view instructions for reproducing our paper, you can see ADMET-AI v1.4.0 at this commit.

Table of Contents

• Installation
• Predicting ADMET properties
  • Command line tool
  • Python module
  • Web server

Installation

ADMET-AI can be installed in a few minutes on any operating system using pip (optionally within a conda environment). If a GPU is available, it will be used by default, but the code can also run on CPUs only.

Optionally, create a conda environment.

conda create -y -n admet_ai python=3.14
conda activate admet_ai

Install ADMET-AI via pip.

pip install admet-ai

Alternatively, clone the repo and install ADMET-AI locally.

git clone https://github.com/swansonk14/admet_ai.git
cd admet_ai
pip install -e .

By default, the pip installation only includes dependencies required for making ADMET predictions, either via the command line or via the Python API. To install dependencies required for hosting the ADMET-AI web server, run pip install admet-ai[web].

If there are version issues with the required packages, create a conda environment with specific working versions of the packages as follows.

pip install -r requirements_frozen.txt
pip install -e .

Note: If you get the issue ImportError: libXrender.so.1: cannot open shared object file: No such file or directory, run conda install -c conda-forge xorg-libxrender.

Predicting ADMET properties

ADMET-AI can be used to make ADMET predictions in three ways: (1) as a command line tool, (2) as a Python module, or (3) as a web server.

Command line tool

ADMET predictions can be made on the command line with the admet_predict command, as illustrated below.

admet_predict \
    --data_path data.csv \
    --save_path preds.csv \
    --smiles_column smiles

This command assumes that there exists a file called data.csv with SMILES strings in the column smiles. The predictions will be saved to a file called preds.csv.

Python module

ADMET predictions can be made using the predict function in the admet_ai Python module, as illustrated below.

from admet_ai import ADMETModel

model = ADMETModel()
preds = model.predict(smiles="O(c1ccc(cc1)CCOC)CC(O)CNC(C)C")

If a SMILES string is provided, then preds is a dictionary mapping property names to values. If a list of SMILES strings is provided, then preds is a Pandas DataFrame where the index is the SMILES and the columns are the properties.

Web server

ADMET predictions can be made using the ADMET-AI web server, as illustrated below. Note: Running the following command requires additional web dependencies (i.e., pip install admet-ai[web]).

admet_web

Then navigate to http://127.0.0.1:5000 to view the website.

Analysis plots

The DrugBank reference plot and radial plots displayed on the ADMET-AI website can be generated locally using the scripts/plot_drugbank_reference.py and scripts/plot_radial_summaries.py scripts, respectively. Both scripts take as input a CSV file with ADMET-AI predictions along with other parameters.