MMCLKin, an attention consistency-guided multimodal and multiscale contrastive learning framework for efficient and accurate prediction of kinase-inhibitor activity and selectivity
Python
10
203 commits
updated Oct 8, 2025
This repository has been tested on Linux operating systems.
Create a virtual environment to run the code of MMCLKin.
It is recommended to use conda to manage the virtual environment.The installation method for conda can be found here.
Make sure to install torch with the cuda version that fits your device.
This process usually takes few munites to complete.
git clone https://github.com/Yanara-Tian/MMCLKin.git
cd MMCLKin
export PYTHONPATH=$PWD:$PYTHONPATH
This package is tested with Python 3.8.15 and CUDA 11.0 on Ubuntu 20.04. Run the following to create a conda environment and install the required Python packages (modify torch+cu116 and dgl-cu116 according to your CUDA version).
conda create -n mmclk python=3.8.15
conda activate mmclk
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install --pre dgl-cu116 -f https://data.dgl.ai/wheels-test/repo.html
pip install matplotlib rdkit scipy Bio transformers sympy scikit-learn jupyterlab lifelines notebook
pip install pyg_lib torch_geometric torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.12.0+cu116.html
pip install networkx==2.7.1
Running the above lines of pip install should be sufficient to install all MMCLKin's required packages (and their dependencies). Specific versions of the packages we tested were listed in requirements.txt.
[1] Download checkpoints(~275MB)(~2mins) and dataset(~1.2GB)(~6mins) using the links, and transfer them to main directory.
https://1drv.ms/u/s!ApQsEObuotQkbxMGi7ezUTOd3CI?e=UYMaM3
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./pkls
https://1drv.ms/u/s!ApQsEObuotQkbsFbZlJ_l8RXMGg?e=fhUyvv
tar -xvf demo_3dkd_affinity.tar.xz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on a small samples from 3DKDavis dataset under the kinase cold-start splitting strategy, run the following script(~3 mins):
python test_3dkdavis_affinity_demo.py
If you would like to replicate the results published in this article, please run the scripts following the steps below.
[1] Download checkpoints(~275MB) and dataset(~15GB) using the links, and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkdavis_new_kinase_affinity.tar.gz
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkdavis_new_kinase_affinity.tar.gz
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on 3DKDavis dataset. We offer dataset splits based on drug cold-start, kinase cold-start, or kinase-drug cold-start, ensuring that the model is tested on unseen kinases, unseen drugs, or both. To enable this option, set the --label argument to new_kinase, new_drug, or both_new. For example, to test on unseen kinases, run the following script(~3 mins):
python test_3dkdavis_affinity.py
MAE CI MSE PCC
0.295 0.853 0.284 0.741
[1] Download checkpoints(~300MB) and dataset(~6.8GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkkiba_new_kinase_affinity.tar.gz
tar zxvf 3dkkiba_new_kinase_affinity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkkiba_new_kinase_affinity.tar.gz
tar zxvf 3dkkiba_new_kinase_affinity.tar.gz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on low sequence similarity dataset of 3DKDavis. For example, to evaluate on unseen kinases, run the following script(~3 mins):
python test_3dkkiba_affinity.py
MAE CI MSE PCC
0.374 0.733 0.300 0.658
[1] Download checkpoints(~300MB) and dataset(~15GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkdavis_selectivity.tar.gz
tar zxvf 3dkdavis_selectivity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkdavis_new_drug_selectivity.tar.gz
tar zxvf 3dkdavis_new_drug_selectivity.tar.gz -C ./test_datasets
[2] To ensure comprehensive learning of human kinases, the predictive performance for kinase inhibitor selectivity of MMCLKin was evalueated under the drug cold-start splitting strategy (~1.5 h).
python test_3dkdavis_selectivity.py
[1] Download checkpoints(~300MB) and dataset(~6.1GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkkiba_selectivity.tar.gz
tar zxvf 3dkkiba_selectivity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkkiba_new_drug_selectivity.tar.gz
tar zxvf 3dkkiba_new_drug_selectivity.tar.gz -C ./test_datasets
[2] Test the predictive performance for kinase inhibitor selectivity of MMCLKin on the low sequence similarity dataset of 3DKKIBA(~25 mins).
python test_3dkkiba_selectivity.py
To provide a more intuitive demonstration, we have also presented the virtual screening process and results of MMCLKin on the LRRK2 kinase target, which is based on a known experimental structure, using a Jupyter Notebook file.
examples/virtual_screening_lrrk2.ipynb
Additionally, we have also presented the virtual screening process and results of MMCLKin on the CRK12 kinase target, for which the crystal structure remains unknown, using a Jupyter Notebook file.
examples/virtual_screening_crk12.ipynb
[1] Download the new constructed 3DKDavis dataset(~1GB)to main directory and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkdavis/3dkdavis.tar.gz
tar zxvf 3dkdavis.tar.gz
[2] Feature extraction, encompassing the biochemical and conformational characteristics of kinase inhibitors, evolutionary information and the intricate spatial structural features of binding pockets and kinase domains.To expedite feature generation for each complex, we recommend executing the following command, which directly generates complex-level features using our pre-generated feature files for all kinases, binding pockets, and kinase inhibitors:
python process_3dkdavis.py
[3] Three splitting strategies (kinase cold-start, drug cold-start and kinase-drug cold-start) remain available for training and evaluating the predictive performance of MMCLKin on kinase-inhibitor binding affinity. For example, to train and test on unseen kinases, run the following script:
python train_3dkdavis_affinity.py
For the selectivity of kinase inhibitors, only the drug cold-start splitting method is provided，execute the following command:
python train_3dkdavis_selectivity.py
[1] Download the new constructed low sequence similarity dataset(~35GB)of 3DKDavis and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkkiba/30sm_3dkkiba_gra_seq.tar.gz
tar zxvf 30sm_3dkkiba_gra_seq.tar.gz
If you wish to obtain features for the entire 3DKKIBA dataset (approximately 500GB), download the 3dkkiba.tar.gz and then execute the following command:
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkkiba/3dkkiba.tar.tar.gz
tar zxvf 3dkkiba.tar.gz
python process_3dkkiba.py
[2] Training and testing MMCLKin on the dataset with low protein sequence similarity. For kinase-inhibitor binding affinity, three splitting strategies are provided. For example, to test on unseen kinases, run the following script.
python train_3dkkiba_affinity.py
For the selectivity of kinase inhibitors on datasets with low protein sequence similarity，execute the following command:
python train_3dkkiba_selectivity.py
[1] Download the PDBBindv2020 dataset and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/pdbbind2020/pdbbind2020.tar.gz
tar zxvf pdbbind2020.tar.gz
[2] extracting the features of PDBBindv2020 and CASF-2016 datasets.
python process_pdbbind2020.py
[3] Training and testing the generalization ability of MMCLKin for protein-drug binding affinity on the dataset with diversity protein structures.
python train_pdbbind2020.py
[1] Download checkpoints(~900MB) and a dataset (~5GB) composed of a 1:1 ratio of inhibitors targeting both wild-type LRRK2 and LRRK2 G2019S mutant kinases.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/finetune_lrrk2g2019s.tar.gz
tar zxvf finetune_lrrk2g2019s.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/mutant/lrrk2_g4.tar.gz
tar zxvf lrrk2_g4.tar.gz -C ./mutant
https://fca_icdb.mpu.edu.mo/DataResource/database/mutant/lrrk2_mw.tar.gz
tar zxvf lrrk2_mw.tar.gz -C ./mutant
[2] Fine-tune the MMCLKin model, and then perform prediction and validation on the inhibitors identified by our group. Execute the following command:
python finetune_lrrk2g2019s.py
Please submit GitHub issues or contact Huanxiang Liu([email protected]), Xiaojun Yao([email protected]), Yanan Tian([email protected]) for any questions related to the source code.
MMCLKin, an attention consistency-guided multimodal and multiscale contrastive learning framework for efficient and accurate prediction of kinase-inhibitor activity and selectivity
Python
10
203 commits
updated Oct 8, 2025
This repository has been tested on Linux operating systems.
Create a virtual environment to run the code of MMCLKin.
It is recommended to use conda to manage the virtual environment.The installation method for conda can be found here.
Make sure to install torch with the cuda version that fits your device.
This process usually takes few munites to complete.
git clone https://github.com/Yanara-Tian/MMCLKin.git
cd MMCLKin
export PYTHONPATH=$PWD:$PYTHONPATH
This package is tested with Python 3.8.15 and CUDA 11.0 on Ubuntu 20.04. Run the following to create a conda environment and install the required Python packages (modify torch+cu116 and dgl-cu116 according to your CUDA version).
conda create -n mmclk python=3.8.15
conda activate mmclk
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install --pre dgl-cu116 -f https://data.dgl.ai/wheels-test/repo.html
pip install matplotlib rdkit scipy Bio transformers sympy scikit-learn jupyterlab lifelines notebook
pip install pyg_lib torch_geometric torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.12.0+cu116.html
pip install networkx==2.7.1
Running the above lines of pip install should be sufficient to install all MMCLKin's required packages (and their dependencies). Specific versions of the packages we tested were listed in requirements.txt.
[1] Download checkpoints(~275MB)(~2mins) and dataset(~1.2GB)(~6mins) using the links, and transfer them to main directory.
https://1drv.ms/u/s!ApQsEObuotQkbxMGi7ezUTOd3CI?e=UYMaM3
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./pkls
https://1drv.ms/u/s!ApQsEObuotQkbsFbZlJ_l8RXMGg?e=fhUyvv
tar -xvf demo_3dkd_affinity.tar.xz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on a small samples from 3DKDavis dataset under the kinase cold-start splitting strategy, run the following script(~3 mins):
python test_3dkdavis_affinity_demo.py
If you would like to replicate the results published in this article, please run the scripts following the steps below.
[1] Download checkpoints(~275MB) and dataset(~15GB) using the links, and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkdavis_new_kinase_affinity.tar.gz
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkdavis_new_kinase_affinity.tar.gz
tar zxvf 3dkdavis_new_kinase_affinity.tar.gz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on 3DKDavis dataset. We offer dataset splits based on drug cold-start, kinase cold-start, or kinase-drug cold-start, ensuring that the model is tested on unseen kinases, unseen drugs, or both. To enable this option, set the --label argument to new_kinase, new_drug, or both_new. For example, to test on unseen kinases, run the following script(~3 mins):
python test_3dkdavis_affinity.py
MAE CI MSE PCC
0.295 0.853 0.284 0.741
[1] Download checkpoints(~300MB) and dataset(~6.8GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkkiba_new_kinase_affinity.tar.gz
tar zxvf 3dkkiba_new_kinase_affinity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkkiba_new_kinase_affinity.tar.gz
tar zxvf 3dkkiba_new_kinase_affinity.tar.gz -C ./test_datasets
[2] Test the prediction performance of MMCLKin for kinase-inhibitor binding affinity on low sequence similarity dataset of 3DKDavis. For example, to evaluate on unseen kinases, run the following script(~3 mins):
python test_3dkkiba_affinity.py
MAE CI MSE PCC
0.374 0.733 0.300 0.658
[1] Download checkpoints(~300MB) and dataset(~15GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkdavis_selectivity.tar.gz
tar zxvf 3dkdavis_selectivity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkdavis_new_drug_selectivity.tar.gz
tar zxvf 3dkdavis_new_drug_selectivity.tar.gz -C ./test_datasets
[2] To ensure comprehensive learning of human kinases, the predictive performance for kinase inhibitor selectivity of MMCLKin was evalueated under the drug cold-start splitting strategy (~1.5 h).
python test_3dkdavis_selectivity.py
[1] Download checkpoints(~300MB) and dataset(~6.1GB), and transfer them to main directory.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/3dkkiba_selectivity.tar.gz
tar zxvf 3dkkiba_selectivity.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/test_datasets/3dkkiba_new_drug_selectivity.tar.gz
tar zxvf 3dkkiba_new_drug_selectivity.tar.gz -C ./test_datasets
[2] Test the predictive performance for kinase inhibitor selectivity of MMCLKin on the low sequence similarity dataset of 3DKKIBA(~25 mins).
python test_3dkkiba_selectivity.py
To provide a more intuitive demonstration, we have also presented the virtual screening process and results of MMCLKin on the LRRK2 kinase target, which is based on a known experimental structure, using a Jupyter Notebook file.
examples/virtual_screening_lrrk2.ipynb
Additionally, we have also presented the virtual screening process and results of MMCLKin on the CRK12 kinase target, for which the crystal structure remains unknown, using a Jupyter Notebook file.
examples/virtual_screening_crk12.ipynb
[1] Download the new constructed 3DKDavis dataset(~1GB)to main directory and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkdavis/3dkdavis.tar.gz
tar zxvf 3dkdavis.tar.gz
[2] Feature extraction, encompassing the biochemical and conformational characteristics of kinase inhibitors, evolutionary information and the intricate spatial structural features of binding pockets and kinase domains.To expedite feature generation for each complex, we recommend executing the following command, which directly generates complex-level features using our pre-generated feature files for all kinases, binding pockets, and kinase inhibitors:
python process_3dkdavis.py
[3] Three splitting strategies (kinase cold-start, drug cold-start and kinase-drug cold-start) remain available for training and evaluating the predictive performance of MMCLKin on kinase-inhibitor binding affinity. For example, to train and test on unseen kinases, run the following script:
python train_3dkdavis_affinity.py
For the selectivity of kinase inhibitors, only the drug cold-start splitting method is provided，execute the following command:
python train_3dkdavis_selectivity.py
[1] Download the new constructed low sequence similarity dataset(~35GB)of 3DKDavis and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkkiba/30sm_3dkkiba_gra_seq.tar.gz
tar zxvf 30sm_3dkkiba_gra_seq.tar.gz
If you wish to obtain features for the entire 3DKKIBA dataset (approximately 500GB), download the 3dkkiba.tar.gz and then execute the following command:
https://fca_icdb.mpu.edu.mo/DataResource/database/3dkkiba/3dkkiba.tar.tar.gz
tar zxvf 3dkkiba.tar.gz
python process_3dkkiba.py
[2] Training and testing MMCLKin on the dataset with low protein sequence similarity. For kinase-inhibitor binding affinity, three splitting strategies are provided. For example, to test on unseen kinases, run the following script.
python train_3dkkiba_affinity.py
For the selectivity of kinase inhibitors on datasets with low protein sequence similarity，execute the following command:
python train_3dkkiba_selectivity.py
[1] Download the PDBBindv2020 dataset and extract its content.
https://fca_icdb.mpu.edu.mo/DataResource/database/pdbbind2020/pdbbind2020.tar.gz
tar zxvf pdbbind2020.tar.gz
[2] extracting the features of PDBBindv2020 and CASF-2016 datasets.
python process_pdbbind2020.py
[3] Training and testing the generalization ability of MMCLKin for protein-drug binding affinity on the dataset with diversity protein structures.
python train_pdbbind2020.py
[1] Download checkpoints(~900MB) and a dataset (~5GB) composed of a 1:1 ratio of inhibitors targeting both wild-type LRRK2 and LRRK2 G2019S mutant kinases.
https://fca_icdb.mpu.edu.mo/DataResource/database/pkls/finetune_lrrk2g2019s.tar.gz
tar zxvf finetune_lrrk2g2019s.tar.gz -C ./pkls
https://fca_icdb.mpu.edu.mo/DataResource/database/mutant/lrrk2_g4.tar.gz
tar zxvf lrrk2_g4.tar.gz -C ./mutant
https://fca_icdb.mpu.edu.mo/DataResource/database/mutant/lrrk2_mw.tar.gz
tar zxvf lrrk2_mw.tar.gz -C ./mutant
[2] Fine-tune the MMCLKin model, and then perform prediction and validation on the inhibitors identified by our group. Execute the following command:
python finetune_lrrk2g2019s.py
Please submit GitHub issues or contact Huanxiang Liu([email protected]), Xiaojun Yao([email protected]), Yanan Tian([email protected]) for any questions related to the source code.