Abstract

Better diagnosis of cancer using topic models may hold the key to curing cancer. This is because cancer is not a single disease but a large number of related diseases. Different types of cancer respond better to different types of therapy. Targeted therapies would improve patient survival, but before we can do that, we need to be able to better sub-categorize cancer tumours. Right now, international consortia are spending hundreds of millions of dollars to sequence the genomes of 25,000 cancers in an effort to identify cancer-associated mutations and try to understand their role. However, many of these mutations do not impact the coding sequence of genes but instead they change the instructions that describe when the genes should be activated and how the gene’s messenger RNA (mRNA) should be processed. So, in order to understand how these mutations are associated with cancer, we need to be able to understand the genome regulatory code and we are far from being able to do that. On the other hand, we can (at least in theory) directly measure the mRNA levels and determine how they change in cancer. But cancer RNA levels are difficult to measure accurate because tumours are “contaminated” by widely varying amounts of normal, healthy tissue. Because RNA levels change rapidly once a tumour biopsy is taken, there’s simply not enough time to physically purify the sample. I will describe how to perform this purification statistically using topic models. Our methods deconvolve measured tumour RNA levels into those coming from healthy and cancerous cell populations. Our modeling framework is naturally extended to identifying the site of origin of metastatic cancers and to finding descriptive groupings of cancers into sub-types.