kevingroup/emogi-reusability

emogi-reusability

Cancer gene prediction with co-expression network

To reproduce the results for cancer gene prediction using co-expression network, you should install the EMOGI model based on the original document (https://github.com/schulter/EMOGI). Then you can use the following parameters for training

python train_EMOGI_cv.py -e 5000 -s 1 -wd 0.005 -hd 20 40 -lm 90 -cv 10 -seed ${seed} -d ./co-expression/CPDB_coexp_multiomics.h5

Additional information: For the training process you would need a GPU to finsish in a reasonable time (it takes about 6 hours with one NVIDIA V100 GPU for 5000 epochs and 10-folds cross validation on the CPDB dataset, it would take a very long time running with CPU). The install time mostly depends on the network speed and normally can finish within 10 minutes. The expected outputs are the resulting model and the predictions, the training details with evaluation metrics can be found in the output log.

Essential gene prediction

The input h5 data files for the essential gene prediction can be found in the essential_gene folder. The following command can be used to reproduce our results

python train_EMOGI_cv.py -e 5000 -s 1 -wd 0.005 -hd 300 100 -lm 5 -cv 10 -seed ${seed} -d ./essential_gene/CPDB_essential_multiomics.h5

Additional information: For the training process you would need a GPU to finsish in a reasonable time (it takes about 6 hours with one NVIDIA V100 GPU for 5000 epochs and 10-folds cross validation on the CPDB dataset, it would take a very long time running with CPU). The install time mostly depends on the network speed and normally can finish within 10 minutes. The expected outputs are the resulting model and the predictions, the training details with evaluation metrics can be found in the output log.

Cancer gene prediction with Graph Attention Network (GAT)

Here we provided two versions of GAT (one based on the original tensorflow GAT library and the other based on pytorch DGL library).

For the tensorflow implementation from the original GAT paper, you could follow the document to install the required packages (https://github.com/PetarV-/GAT). And then you also need to install the other required packages with pip install h5py numpy scipy

After installing the packages, you can use the command below to train the GAT models

python train_GAT.py -e 5000 -lr 0.01 -lm 90 -hd 8 -ah 8 1  -do 0.25 -wd 0.0 -seed 1 -d ${PATH_TO_INPUT_H5_FILE}

For the DGL version, to make it consistent with the tensorflow version here we also used cuda 10.2. You need to first install the pytorch library with conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch as well as the dgl libray with conda install -c dglteam dgl-cuda10.2 . Then the other required packages can be installed with pip install h5py numpy scipy.

After installing the packages, you can use the command below to train the GAT models based on the DGL library

python dgl_gat_main.py \
    --num_epochs=1000 \
    --hidden_dims=64 \
    --heads=4 \
    --dropout=0.2 \
    --loss_mul=1 \
    --sample_filename=${PATH_TO_INPUT_H5_FILE} \
    --lr=0.001 \
    --seed=1  \
    --cuda \ # if you have GPU available

Additional information: For the training process you would need a GPU to finsish in a reasonable time, it would take a very long time running with CPU. For the tensorflow version, it takes about 3 hours with one NVIDIA V100 GPU for 5000 epochs on the CPDB dataset. For the DGL version, it takes about 10 minutes with one NVIDIA V100 GPU for 1000 epochs on the CPDB dataset. The install time mostly depends on the network speed and normally can finish within 10 minutes. The expected outputs are the training details with evaluation metrics in the log.