Multi-task analysis of gene expression data on cancer public datasets

Background There is an availability of omics and often multi-omics cancer datasets on public databases such as Gene Expression Omnibus (GEO), International Cancer Genome Consortium and The Cancer Genome Atlas Program. Most of these databases provide at least the gene expression data for the samples contained in the project. Multi-omics has been an advantageous strategy to leverage personalized medicine, but few works explore strategies to extract knowledge relying only on gene expression level for decisions on tasks such as disease outcome prediction and drug response simulation. The models and information acquired on projects based only on expression data could provide decision making background for future projects that have other level of omics data such as DNA methylation or miRNAs.

Results We extended previous methodologies to predict disease outcome from the combination of protein interaction networks and gene expression profiling by proposing an automated pipeline to perform the graph feature encoding and further patient networks outcome classification derived from RNA-Seq. We integrated biological networks from protein interactions and gene expression profiling to assess patient specificity combining the treatment/control ratio with the patient normalized counts of the deferentially expressed genes. We also tackled the disease outcome prediction from the gene set enrichment perspective, combining gene expression with pathway gene sets information as features source for this task. We also explored the drug response outcome perspective of the cancer disease still evaluating the relationship among gene expression profiling with single sample gene set enrichment analysis (ssGSEA), proposing a workflow to perform drug response screening according to the patient enriched pathways.

Conclusion We showed the importance of the patient network modeling for the clinical task of disease outcome prediction using graph kernel matrices strategy and showed how ssGSEA improved the prediction only using transcriptomic data combined with pathway scores. We also demonstrated a detailed screening analysis showing the impact of pathway-based gene sets and normalization types for the drug response simulation. We deployed two fully automatized Screening workflows following the FAIR principles for the disease outcome prediction and drug response simulation tasks.