This directory contains the code to compare models on molecular data. The code is written in Python and uses the RDKit/Datamol library\cite{} to handle molecular data. The datasets are imported from the Therapeutic Data Commons (TDC) plateform\cite{}.
To install the required packages,you need an environement with "pytorch", "pytorch-geometric" and "torch-scatter" already installed, you can use the following command,
pip install -r requirements.txt
cd external_repo/pre-training-via-denoising
pip install -e .Various descriptors use a 3-dimensional representation of the molecules. To generate these representations, and preprocess the datasets, use the script precompute_.py.
python precompute_molf_descriptors.py --dataset <dataset_name> --descriptors <[optional] molecular descriptors to compute>To evaluate the Information Sufficiency, you can use the following command:
python main.py --dataset <dataset_name> --X <models_names> --Y <models_names> --out-dir <output_dir>This command will train the models specified in the --X and --Y arguments on the dataset <dataset_name>, and save the results in the directory <output_dir>.
Once the Mutual information is computed, the results can be visualized using the notebooks
The various models considered can be trained on downstream tasks using the script main_downstream.py.
python main_downstream.py --dataset <dataset_name> --embedders <models_names> --n-runs <number_of_runs> --split-method <random/scaffold> --hidden-dim <hidden_dim> --n-layers <n_layers> --n-epochs <n_epochs> --d-rate <dropout_rate> --lr <learning_rate> --batch-size <batch_size>This command will train the models on the single instance tasks using the dataset <dataset_name>, and save the results in the directory <output_dir>.
For DTI tasks, the script main_downstream_dti.py can be used.
python main_downstream.py --dataset <dataset_name> --embedders <models_names>