Discovery of functional motifs by association to clinical features using Graph Neural Networks.
We use conda. The environment file below has the versions we tested (Python 3.11, PyTorch 2.1.1 with CUDA 12.1, PyTorch Geometric 2.4.0):
conda env create -f pygem_environment.yml
conda activate py_gem_3Training needs a GPU; it stops with an error if PyTorch cannot see one. On NixOS, PyTorch only finds the driver when you set:
export LD_LIBRARY_PATH=/run/opengl-driver/libEach dataset lives in data/<dataset>/.
| Option | What it does | Default |
|---|---|---|
--dataset_name |
JacksonFischer or METABRIC |
JacksonFischer |
--en |
experiment name, used for the output folders | test_experiment |
--folds_json |
the folds file (required, except with --full_training) |
|
--full_training |
train one model on all patients instead of cross-validating, for example to explain it afterwards. The number of epochs is chosen on 20% of the patients, then the model is retrained on everyone. Saves models/<en>/<session>_SD.mdl and <session>.json |
off |
--params_json |
hyperparameter file; its values replace the ones given on the command line | |
--cv_setting |
nfold or heldout (see above) |
nfold |
--feature_normalise |
raw: use the protein abundances as they are; log1p_z: log-transform them and scale them with the training-fold mean and SD |
raw |
--seed |
fold k is trained with seed seed + k |
42 |
--gpu_id |
which GPU to use | 0 |
You can set these on the command line, but usually they come from the hyperparameter file.
| Hyperparameter | Meaning | Default |
|---|---|---|
model |
graph layer type: GATV2, GATConv, PNAConv, TransformerConv or GINConv |
GATV2 |
num_of_gcn_layers |
number of graph layers | 2 |
gcn_h |
size of each graph layer | 128 |
heads |
attention heads (GAT layers only) | 1 |
num_of_ff_layers |
number of fully connected layers after the graph layers | 2 |
fcl |
size of the fully connected layers | 128 |
dropout |
dropout rate | 0.2 |
lr |
learning rate | 0.01 |
weight_decay |
L2 penalty | 0.001 |
bs |
batch size (images) | 32 |
epoch |
maximum number of epochs | 5 |
factor, patience, min_lr |
the learning rate is multiplied by factor after patience epochs without improvement, down to min_lr. Training stops after 2 × patience epochs without improvement |
0.5, 5, 0.00002 |
aggregators, scalers |
aggregation settings for PNAConv |
sum mean, amplification identity |
use_batch_norm |
batch normalisation after each graph layer (file only) | false |
pooling |
how cells are summarised into one image vector: mean or mean_max (file only) |
mean |
augment |
randomly perturb the graphs during training (--augment on the command line) |
false |
min_epochs |
smallest epoch that can be kept (--min_epochs) |
0 |
To train on a second GPU, make only that card visible and keep --gpu_id 0:
CUDA_VISIBLE_DEVICES=1 python train_test_controller.py ... --gpu_id 0Classification tasks (for example tumour grade) use a separate script with its own options:
python bin/train_test_controller_classification.py --helpGNU General Public License v3.0, see LICENSE.