Supplied input data

1) Gene sets

In the folder example/ you will find the following 4 files. The first file is a file that can be used by the user to index their own networks and disease/test gene lists. The other 2 are an example of input train_node_labels and test_node_labels files for our disease-example, the primary immunodeficiencies (PIDs). The last one is a file with extra, qualitative features (node_attributes.tsv). This file is of general use and it is not filtered for any specific disease.

• Protein coding genes
  our_protein_coding_gene_list_20760genes_all_synonyms_genenames_indexed.tsv
  Columns:

  • Gene: all HGCN names matching the 20760 genes (Ensembl ID-based gene total)
  • Gene_ID: GRCh37 Ensembl gene stable ID of the 20760
  • index: 0-20759, increasing number-coding of each Gene_ID

  Protein coding genes were extracted from GRCh37 Ensembl and HGCN genes, accessed on 19/7/2018. GRCh37 Ensembl gene stable IDs were filtered for protein coding genes (having a protein stable ID), and also for chromosomes 1 to 22, X, and Y. HGCN download included all protein coding genes and filtered for non-duplicated approved symbols, and for chromosomes 1 to 22, X, and Y. The Ensembl gene stable IDs were mapped to the approved HGCN gene names. We then added the Ensembl gene stable IDs for any synonymous gene names from our HGCN table that matched the Ensembl gene names.
  The resulting table (our_protein_coding_gene_list_20760genes_all_synonyms_genenames_indexed.tsv) contains the gene index number, Ensembl stable IDs, approved HGCN names with all synonyms (58749 HGCN gene names to 20760 protein coding Ensembl gene IDs) forming the list of our protein coding genes. We end up with our protein coding gene table of 20760 gene stable IDs.

  The user can provide their own full list of genes, indexed anew, with the index respected in all other input data files (edgelists, disease and test set gene lists).

• Disease gene list
  Disease gene lists are first filtered for the genes that exist in the protein coding gene list file.
  train_node_labels.tsv: train disease gene list example
  Columns:

  • node: indexed gene name of the 20760 protein coding genes
  • label: 1 or 0 according to whether it is on your disease gene list or not

• Test set gene list
  Test set gene lists are first filtered for the genes that exist in the protein coding gene list file.
  test_node_labels.tsv: test disease gene list example
  Columns:

  • node: indexed gene name of the 20760 protein coding genes
  • label: 1 or 0 according to whether it is on your test set gene list or not

2) Gene features

Following the concept of incorporating as much relevant available data for the genes as possible into the learning process, Tiresias offers the option to add gene information not included in networks. Quantitative features are introduced as node attributes. The features provided are pLI scores, GERP scores, RVIS scores, dn/ds scores, gene age, in vivo essentiality of genes and gene damage index.

Attributes	Source
gene age	Popadin et al., 2014
GDI	Itan et al., 2015
GWAS_Mpval	MacArthur et al., 2016
in vivo essentiality	Chen et al., 2016
ncGERP	Davydov et al, 2010
pLI	Lek et al., 2016
RVIS percentile	Petrovski et al., 2015
slr dnds	Rausell et al., 2014

• Extra features file
  node_attributes.tsv: extra features example. NAs are curated to the mean of column values, and standard scaling of all columns has been performed. This is an optional file. Users can introduce their own node_attributes file.
  Columns:
  • node: indexed gene name of the 20760 protein coding genes for the genes that have at least one of the features present
  • _ attribute_0_: in vivo essentiality
  • _ attribute_1_: GWAS_Mpval
  • _ attribute_2_: pLI
  • _ attribute_3_: ncGERP
  • _ attribute_4_: RVIS percentile
  • _ attribute_5_: slr dnds
  • _ attribute_6_: GDI
  • _ attribute_7_: gene age

3) Networks

• Protein-Protein Interactions (PPI) network
  Based on the idea that molecular interactions can be physical or functional, with physical interactions coming from experiments with a proven binary interaction whereas functional interactions may be coming from homology or other experimental studies, Tiresias provides a PPI network of physical interaction experimentally validated. Tiresias PPI network is constructed from publicly available resources which were merged, filtered to remove duplicates, and then matched to the provided protein coding gene list of 20760 genes. This state-of-the-art PPI network was inspired from Huang et al., where a protein-protein association network compendium was produced, namely Network Data Exchange. After filtering that very network compendium for protein-protein networks that include experimentally-based interactions and are in an edgelist form, we created our own PPI network that, in addition, includes an updated version of Didier’s original update, and the STRING experiments network filtered for only the high confidence (>700) interactions. In total, our PPI was created based on 25 databases for binary, experimental interactions, and removing weights (all data accessed on 07.05.2019).

• Co-expression network
  Tiresias provides the STRING database co-expression network, filtered for duplicates and for the genes existing in the provided protein coding gene list.

• Regulatory networks
  Regulatory networks signifying the transcription-factor-to-gene impact (TF->gene) were collected from three different sources: Marbach et al., Saha et al., and Sonawane et al.. The first comprehensive resource (Marbach) was explored for its 32 high-level tissue- and cell- specific networks, the second (Saha) for all 16 tissue-specific networks, and the third (Sonawane) for its 38 tissue-specific networks. All regulatory networks in Tiresias are directed and weighted networks. Saha and Sonawane networks also being signed, are turned into simply weighted and directed by removing the negative sign (-). We suggest to make a customized choice of input regulatory networks according to the disease of interest, and the affected tissues and cells of the targeted disease(s).