GID: trained graph convolutional autoencoder

Weights of the GID model trained on 1,007 cancer cell lines (1,998 genes as nodes, 5 omics layers, 5035 input features per gene). The code is at https://github.com/Mohamedema/GID.

Loading

from gid.train import load_model

model, config, checkpoint = load_model("model.pt")

The encoder has two graph convolutions (hidden dimension 400, latent dimension 190) with batch normalisation, PReLU, dropout (0.1518) and a linear skip connection; the decoder mirrors it. The node features of a gene are its values in every cell line and layer, so the model applies to this cohort and to the gene order of the networks repository. The input tables are listed in the Data availability section of the manuscript, and python -m gid.infer recomputes the networks from them.