Resources:

• README.md: this file.
• GSE18123GPL570.csv and GSE18123GPL6244.csv: two datasets
• result.csv: result for each method with different number of features. Measured by accuracy, auc, f1, and mcc.

source codes:

• conventional_feature_ranking.py: 12 conventional feature ranking methods plus AUC (baseline) and SFARI.
• gene_interaction_co_expression.py: create co-expression networks.
• gene_interaction_pathway.py: create pathway networks ('DO', 'GObp','GOcc', 'GOmf', and 'HPO' networks).
• gene_interaction_PPI.py: create PPI networks.
• gene_interaction_pubmed.py: create pubmed-embedding networks.
• gene_embedding.py: create word, hence gene name included, embedding, used as input for gene_interaction_pubmed.py
• geneRank.m: score genes based on their networks: co-expression, pathway, PPI, and pubmed-embedding networks
• score_2_index.py: turn the score returned by geneRank.m into the index of genes in the datasets.
• classify.py: running the classification task with the features ranked highest by the above methods, be it conventional or network-based.

folder pubmed/:

• pubmed.txt (unzip pubmed.7z): contain PubMed raw text as input to learn word/gene embeddings.
• gene_embedding.embed (unzip gene_embedding.7z): word/gene embeddings returned by running word2vec models on pubmed.txt.
• gene_name_ref.txt: mapping vocabulary (gene names) between gene_embedding.embed and the datasets.

folder mapping/ (unzip mapping.7z):

• hgncTo*.csv: ontology mapping from genes to DO, GO, and HPO.
• hippie_current.txt: protein-protein interactions, downloaded from http://cbdm-01.zdv.uni-mainz.de/~mschaefer/hippie/hippie_current.txt
• SFARI-Gene_genes_export09-06-2018.csv: SFARI list of genes, downloaded from https://gene.sfari.org//wp-content/themes/sfari-gene/utilities/download-csv.php?api-endpoint=genes

folder gene_interaction/ (upzip all *.7z files into *.csv files):

• store the output of gene_interaction_*.py, creating gene-gene networks.
• gene_interaction/score/: store the output of geneRank.m, scoring genes by their networks.

folder feat_ranking/ (unzip feat_ranking.7z):

Gene ranking, either return by conventional_feature_ranking.py or score_2_index.py, which is the network-based feature ranking. This will be the input for classify.py, telling which features to be used in the classification.

Step-by-step running:

Please unzip all *.7z files required before running any algorithms. Please note that some may take hours to run.

0. Installing Python libaries needed for

• conventional feature selections:

pip install skfeature-chappers

• training embbeded vectors:

pip install gensim

1. Running conventional feature ranking:

python conventional_feature_ranking.py index

where index is in the range of the list of conventional methods ['chi_square', 'cmim', 'f_score', 'fisher_score', 'gini_index', 'icap', 'jmi', 'll_l21', 'ls_l21', 'reliefF', 'rfs', 'trace_ratio', 'SFARI', 'AUC'] The ranking of genes will be stored at the folder of feat_ranking/

2. Building gene-gene networks:

• co_expression interactions:

python gene_interaction_co_expression.py

• pathway interactions:

python gene_interaction_pathway.py

The mapping of gene-ontologies is stored at mapping/hgncTo*.csv. There are five ontologies: 'DO', 'GObp','GOcc', 'GOmf', and 'HPO'.

• PPI networks:

python gene_interaction_PPI.py

The material for building PPI is stored at mapping/hippie_current.txt, downloaded from http://cbdm-01.zdv.uni-mainz.de/~mschaefer/hippie/hippie_current.txt

• PubMed based networks: For learning gene-gene interactions from PubMed, first the embedding for genes should be learned by running:

python gene_embedding.py

Raw data extracted from PubMed is stored at /pubmed/pubmed.txt. The embbed model is stored at pubmed/gene_embedding.embed. Then the network is learned through:

python gene_interaction_pubmed.py

• Running all these python programs result in gene-gene networks stored as gene_interaction/

3. Running network-based feature ranking for these networks by running:

geneRank.m

The score for genes returned by running geneRank.m is stored at /gene_interaction/score/. We convert the score to index for genes by running:

python score_2_index.py

The ranking of genes returned by running score_2_index.py will be stored at the folder of feat_ranking/

4. Now we have feature ranking for both conventional and network-based methods. Then the features with highest ranks will be used in the following classification:

python classify.py

The result for all these feature selections returned by running classify.py will be stored in result.csv.