Diseases Gene Discovery Using Artificial Neural Networks

I. Gene Expression Datasets (GED)

Gene expression datasets for different Illumina platforms is collected from multiple organs, development stages and perturbation experiments.

It would be useful to get 1K data set i.e. 1024 different data (i.e. columns) for main human organs/tissues (stages) such as eye, kidney, craniofacial, etc.

It may also be useful to use datsets from different models to account for inter-species differences

Datasets/Papers

1. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6658352/ [Bulk RNA-seq 7 tissues multiple stages]

   • E-MTAB-6769 (chicken),
   • E-MTAB-6782 (rabbit)
   • E-MTAB-6798 (mouse, annotated with Ensembl 71)
   • E-MTAB-6811 (rat)
   • E-MTAB-6813 (rhesus)
   • E-MTAB-6814 (human)
   • E-MTAB-6833 (opossum)

2. https://www.ebi.ac.uk/gxa/sc/experiments/E-ENAD-15/results/tsne [Single-cell multiple tissues]

3. https://lincsproject.org/LINCS/tools/workflows/find-the-best-place-to-obtain-the-lincs-l1000-data [L1000]

II. Processing GED For Gene Vectors

GED are processed for normalized expressions counts. These are likely normalized that sum of normalized expression is scaled between 0 and 1. These represent as gene expression vectors.

1. Compute Expression Counts

2. Impute Missing Expression Counts

Basic Imputation: Replace all missing FPKM counts to 0.
KNN: https://www.bioconductor.org/packages/devel/bioc/manuals/impute/man/impute.pdf
GAN: https://www.biorxiv.org/content/10.1101/2020.06.09.141689v1
AE: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7144625/

III. Generate Training Set

Labels could be fecthed for genes from DisGeNet site. Model as multi-label classification experiment by providing multiple diseases labels to each gene (using some cutoff).

Features

1. Gene expression
2. Kmer-embeddings from promotor regions
3. Keywords or sentiments associated with genes