A precise classification of cancerous tissue and/or tumor subtype is difficult, but very important to maximize efficacy and minimize toxicity of treatment. Given the availability of efficient statistical methods, bio-molecular information from techniques such as gene expression microarrays could become as, or even more important for cancer diagnosis than traditional clinical factors. A challenging task is to find groups of genes whose collective expression is strongly predictive for an outcome variable of interest.

This talk presents Pelora, a novel type of algorithm that combines gene selection, gene grouping and sample classification in a single step. It aims at finding gene signature components whose centroids render the discrimination of the outcome variable y as simple as possible. This is achieved with an efficient grouping heuristic that operates in a penalized likelihood setting. Empirical results from different microarray datasets yield evidence that the algorithm identifies gene groups whose collective expression has very good predictive potential, and can thus be beneficial in medical diagnostics, as well as for providing insights into gene function and regulation.

Moreover, the challenge of combining various sources of genomic information is discussed. Pelora can be adapted to cope with multiple gene expression datasets, with clinical predictors and biological contraints. Finally, the talk addresses the issue of statistical inference in such situations, and presents methodology to for deciding where the most relevant predictive information comes from.