Integrating Multi-Omic Data with Penalized Cox Models for Improved Cancer Prognostic Signatures

Cancer research faces a significant challenge in identifying crucial genes for prediction, treatment, and amelioration. The vast array of potential genes complicates the identification of genomic markers that effectively predict outcomes while remaining interpretable. This thesis introduces a novel miRNA-targeting metric that scores genes based on their interactions with miRNAs. We use this metric in an integrated approach to elucidate a small number of important biomarkers in colorectal cancer (CRC) and small cell lung cancer (SCLC).

We developed a method integrating gene prioritization steps with miRNA targeting information and penalized Cox regression models to identify biomarkers that accurately predict cancer survival. Our approach, which combines multiple types of omics information, significantly outperforms existing methods in biomarker identification from large omics datasets.

For instance, in CRC, our method achieved a concordance index (a measure of predictive accuracy) of 0.75 compared to 0.65 for standard gene expression-based approaches. Similarly, in SCLC, our method showed improved performance over commonly used biomarker discovery tools such as DESeq2 and edgeR. Moreover, our approach demonstrated good performance across various other cancer types, suggesting its broad applicability.

The identified signatures show strong literature support, with many genes previously implicated in various cancers. Furthermore, these biomarkers are enriched in cancer-relevant pathways, including iron metabolism, NOTCH3 signaling, and ALK signaling. Our pipeline's multi-step approach allows for easy substitution of other model types, enhancing its flexibility while maintaining interpretability by using a limited number of genes in the final signatures.

This research contributes to the field by providing a computationally accessible, adaptable, and effective method for identifying interpretable biomarkers. These findings have potential implications for personalized cancer prognosis and treatment strategies, paving the way for more targeted therapeutic approaches.