CNN model based on RBD-ACE2 sequence, predicting binding affinity change upon mutations.
The code currently only supports CUDA-supported machine.
Pre-trained models are based on these packages, and errors/warnings may occur when using different versions.
Install Anaconda for Linux on the node following the instructions.
conda env remove -n cnn_seq # Remove existing conda env
conda create -n cnn_seq python=3.8.5 # Create new conda env
conda activate cnn_seq # Activate the cnn_seq env
conda install numpy matplotlib scikit-learn # Basic tools
conda install pytorch=1.8.1 torchvision torchaudio cudatoolkit=10.2 -c pytorch # PyTorch with CUDA support
conda install biopython -c conda-forge # Biopython[./]
cnn_kfold.py is the main code for training, prediction and analysls.
var_blind_test_set.csv contains 1677 variants, constituting a completely blind test set.
var_chk.csv contains all variants to be predicted.
var_chk_circ.csv contains many variants (including circulating variants), which can serve as an example of how to prepare the list.
[./Mater]
[./Mater/expt_multi_mutant] Contains experimental data for training purposes.
[./Mater/python] Contains python modules called by cnn_kfold.py.
[./Mater/species] Contains basic information used by cnn_kfold.py for each hosts.
[./Models] Place where trained models are stored. Note that freshly-trained models will overwrite existing ones.
[./Results] Temporary place where predicted results are processed.
Modify or create the file var_chk.csv.
See var_chk_circ.csv for examples.
Note that for non-human hosts, the residue indices are 318 smaller than the expected one. For instance, deer_166K corresponds to E484K for deer, while humancov2_484K or simply 484K corresponds to E484K for human.
./cnn_kfold.py 5
./cnn_kfold.py 3