MTAN-ADMET: A Multi-Task Adaptive Neural Network for Efficient and Accurate Prediction of ADMET Properties

📖 Overview

MTAN-ADMET is a deep learning framework for efficient and accurate prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties.
It integrates multi-task adaptive neural networks with learned molecular embeddings to achieve state-of-the-art performance.

This repository contains:

• data/ → datasets and generated embeddings
• envs/ → environment setup files
• notebooks/ → training, evaluation, and analysis notebooks
• plot_styles/ → plotting configurations
• scripts/ → utility scripts

This software corresponds to the research article:

Shahid S, Maity D, Chakrabarty S. MTAN-ADMET: A Multi-Task Adaptive Neural Network for Efficient and Accurate Prediction of ADMET Properties. ChemRxiv. 2025; doi: 10.26434/chemrxiv-2025-zhrsk.
(This content is a preprint and has not been peer-reviewed.)

Please refer to the paper for model details and methodology.

---

⚙️ Installation

1. Clone the repository

git clone https://github.com/TeamSuman/MTAN-ADMET.git
cd MTAN-ADMET

2. Create environments

Two environments are required:

• CDDD environment (for generating embeddings)
• MTAN-ADMET environment (for training and evaluation)

# Create cddd environment
cd envs
conda env create -f cddd.yml
conda activate cddd

# Create mtan-admet environment
conda env create -f mtan_admet.yml
conda activate mtan_admet

---

🚀 Usage

Step 1: Download the default model

The embeddings are generated using the pretrained sequence-to-sequence autoencoder by Winter et al.

Winter, R.; Montanari, F.; Noé, F.; Clevert, D.-A. Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations. Chemical Science 2019, 10, 1692–1701.

The pretrained model is available on Google Drive. Download and unzip by execuiting the bash script "download_default_model.sh":

cd scripts/cddd
./download_default_model.sh

The default_model.zip file can also be downloaded manualy under https://drive.google.com/open?id=1oyknOulq_j0w9kzOKKIHdTLo5HphT99h

Step 2: Generate molecular embeddings

Activate the cddd environment and run:

conda activate cddd
jupyter notebook notebooks/test/embeddings_generator.ipynb

This produces:

• data/admet_embeddings_clean.csv → contains SMILES, groups, embeddings, and property values
• data/embeddings_features.dat → contains embeddings + labels (used for training)

---

Step 3: Train the MTAN-ADMET model

Activate the mtan-admet environment and run:

conda activate mtan_admet
jupyter notebook notebooks/test/mtan_admet_run.ipynb

• Trains the adaptive neural network.
• Results can be evaluated within the notebook.
• Multiple runs can be performed by changing the random seed to collect performance statistics.

---

Step 4: Baseline comparison

To compare MTAN-ADMET against other models:

jupyter notebook notebooks/test/baseline_analysis.ipynb

---

Step 5: Dataset analysis

To analyze the ADMET dataset (collected from Du et al.):

Du, B.-X.; Xu, Y.; Yiu, S.-M.; Yu, H.; Shi, J.-Y. ADMET property prediction via multi-task graph learning under adaptive auxiliary task selection. iScience 2023, 26, 108285.

jupyter notebook notebooks/analysis/Data_analysis.ipynb

---

📊 ADMET Properties Included

The MTAN-ADMET framework predicts the following properties:

• HIA
• OB
• P-gp inhibitor
• P-gp substrates
• BBB
• Respiratory toxicity
• Hepatotoxicity
• Half-life
• CL (Clearance)
• Cardiotoxicity (1 μM)
• Cardiotoxicity (5 μM)
• Cardiotoxicity (10 μM)
• Cardiotoxicity (30 μM)
• CYP1A2
• CYP2C19
• CYP2C9
• CYP2D6
• CYP3A4
• Acute oral toxicity (LD50)
• IGC50
• ESOL (solubility)
• logD
• Caco-2 permeability
• PPB (Plasma Protein Binding)

---

📑 Citation

If you use this software, please cite:

Shahid S, Maity D, Chakrabarty S. 
MTAN-ADMET: A Multi-Task Adaptive Neural Network for Efficient and Accurate Prediction of ADMET Properties. 
ChemRxiv. 2025. doi:10.26434/chemrxiv-2025-zhrsk