MMPK is an end-to-end multimodal deep learning model for human oral pharmacokinetic parameters prediction. It leverages multiple molecular representations learned from molecular graphs, substructure graphs, and SMILES sequences to comprehensively capture multi-scale molecular information. Multi-task learning and data imputation strategies are also employed to improve data efficiency and model robustness. Additionally, the substructure-aware cross-attention mechanism enhances the interpretability of MMPK by identifying chemically meaningful substructures that contribute to the predictions.
Due to the scheduled network maintenance of East China University of Science and Technology (华东理工大学), MMPK Web Server is previously restricted to access only during 8:00 AM to 10:00 PM Beijing Time (UTC+8). The issue has now been resolved, and the webserver is currently accessible 24 hours a day.
To prevent similar issues in the future, we provide an alternative link hosted on Streamlit Community Cloud, which offers identical functionality and content, except for a different URL and slightly slower prediction speed. According to Streamlit Community Cloud policy, if the app has no traffic for 12 hours, it will go to sleep to conserve resources. If you visit a sleeping app, simply click the "Yes, get this app back up!" button and wait a moment — the app will be available shortly.
We appreciate your understanding and are happy to provide further assistance if needed.
MMPK can predict the following eight human oral pharmacokinetic parameters:
- Area under the concentration–time curve (AUC [ng*h/mL])
- Maximum plasma concentration (Cmax [ng/mL])
- Time to reach Cmax (Tmax [h])
- Elimination half-life (t1/2 [h])
- Apparent clearance (CL/F [L/h])
- Apparent volume of distribution (Vz/F [L])
- Mean residence time (MRT [h])
- Oral absolute bioavailability (F [%])
Download the repository and create the environment:
git clone https://github.com/xli7654321/MMPK.git
cd ./MMPK
conda env create -f environment.yml
conda activate mmpkInstall PyTorch (with CUDA 11.8 support) after activating the environment:
pip install torch --index-url https://download.pytorch.org/whl/cu118👉 We strongly recommend using our MMPK Web Server for new prediction.
You can download our trained models from Zenodo. The trained 10-fold MMPK models should be stored in the checkpoints/ directory, under a folder name specified by the --checkpoints_folder argument (e.g., checkpoints/mmpk/).
Use the following command to start prediction:
python predict.py --checkpoints_folder mmpk --output_file_pred [PREDICTION_FILENAME] --output_file_att [ATTENTION_FILENAME]--output_file_pred: Output filename (*.csv) for predicted PK parameters (saved inprediction/)--output_file_att: Output filename (*.csv) for substructure attention weights (saved inattention/)
After running the command, you will be prompted to select an input mode:
Please select input mode [1: Manually, 2: CSV file] (default: 1):
> 1
Please enter SMILES (separated by commas):
> Clc1cccc(Cl)c1NC1=NCCN1, NCC(CC(=O)O)c1ccc(Cl)cc1
Please enter administered dose (separated by commas):
> 0.1, 10
Please select dose unit [mg or mg/kg]:
> mg
Do you want to standardize SMILES? [y/n] (default: y):
> yYour CSV file should contain at least two columns: smiles and dose, and should be placed in the examples/ directory.
Please select input mode [1: Manually, 2: CSV file] (default: 1):
> 2
Please enter the path to your CSV file:
> example.csv
Please select dose unit [mg or mg/kg]:
> mg
Do you want to standardize SMILES? [y/n] (default: y):
> yAfter input is confirmed, the script will summarize the compounds and doses, then begin prediction.
Note: Here, the substructure attention weights are independent of dose.
We provide the raw datasets (raw_dataset) and preprocessed datasets (model_dataset) at Zenodo. To reproduce the results, the data for approved drugs should be placed in the data/approved directory. Then, you can execute the split.py script to split the dataset into training, validation, and test sets.
python utils/split.pyThen, use the following command to run training and evaluation.
bash run.sh 0 "mmpk" "auc cmax tmax hl cl vz mrt f"The best model checkpoints and prediction results will be saved under the checkpoints/mmpk/ and results/mmpk/ directories, respectively.
The script run.sh accepts three arguments:
DEVICE: the CUDA device ID to use (default:0)FOLDER: the name of the subfolder undercheckpoints/andresults/to store model checkpoints and prediction results (default:"mmpk")PK_PARAMS: the list of PK parameters to be predicted (default:"auc cmax tmax hl cl vz mrt f")
For external evaluation, you should ensure that the external datasets are placed under the following directories:
data/investigational/data/approved_2024/
Similarly, the trained 10-fold MMPK models should be stored in the checkpoints/ directory, under a folder name specified by the --checkpoints_folder argument (e.g., checkpoints/mmpk/).
python test.py --checkpoints_folder mmpkThis script will load all 10 trained MMPK models and generate averaged predictions for the compounds. The predicted values of each PK parameter for each compound and the overall metrics will be saved under the results/investigational/ and results/2024/ directories, respectively.
@article{li2025mmpk,
title={MMPK: A Multimodal Deep Learning Framework to Predict Human Oral Pharmacokinetic Parameters},
author={Li, Xiang and Zhan, Meiling and Fang, Jiaojiao and Liu, Guixia and Tang, Yun and Li, Weihua},
journal={Journal of Medicinal Chemistry},
year={2025},
volume={68},
number={15},
pages={16678--16690},
publisher={American Chemical Society},
doi={10.1021/acs.jmedchem.5c01522}
}This project is licensed under the MIT License.
This project is developed by the Laboratory of Molecular Modeling and Design (LMMD) at the School of Pharmacy, East China University of Science and Technology (ECUST).
To learn more about about our research projects, publications, and other tools, please visit our website.