Extending the knowledge of genotype-phenotype relationships with machine learning approaches
June 14 2013, 11AM
Recently, a tremendous amount of data for genomic variations in many living systems is being collected and utilized to understand their phenotypic characteristics. Machine learning approaches can make a significant contribution to the extension of the knowledge of these genotype-phenotype relationships. In the first half of this talk, the transcriptional response to copy number changes of Drosophila genes will be explained. We profiled expression in engineered flies where gene dose was reduced from two to one. While expression of most one-dose genes was reduced, the gene-specific dose responses were heterogeneous. Expression of two-dose genes that are first-degree neighbors of one-dose genes in novel network models also changed, and the directionality of change depended on the response of one-dose genes. In the second half of this talk, a probabilistic model for dissecting cancer heterogeneity will be presented. Here, individual cancer cases are modeled as mixtures of subtypes and phenotypic similarities are explained with the help of genotypic features (e.g., copy number variations, mutations, and microRNA levels) used to define the subtypes. We applied our probabilistic model to brain cancer data and successfully identified well-known subtypes along with their important features.