Predicting Chemotherapeutic Response Using Optimally Discriminative Subnetworks

Wednesday, February 29, 2012 - 2:30pm - 2:45pm
Keller 3-180
Phuong Dao (Simon Fraser University)
Although molecular profiles of tumor samples have been widely applied to the classification of clinical outcomes, the lack of reproducibility and generalizability has limited use of the clinical use of markers identified in these studies. In particular, finding robust predictors of treatment response has proved to be especially challenging. Recent studies integrating protein-protein interaction (PPI) networks with gene expression profiles have demonstrated that subnetwork (SN) markers offer clear advantages in classification problems, where these SN markers consist of synergistic genes whose aggregate expression discriminate groups of samples. However, existing methods for constructing SNs do not guarantee an optimal solution and have yet to be applied towards predicting individual patient response to drug treatment.

We developed a novel and efficient randomized algorithm to identify optimally discriminative SNs for classification of samples from different classes. Our algorithm is based on a color coding paradigm, which allows for identifying the optimal discriminative SN markers for any given error probability. When the maximum size of a SN marker is k = O(logn) where n is the size of the network, we have a polynomial time algorithm with a fixed error probability. We demonstrated our method on a large published dataset of drug response comprising two independent cohorts of breast patients treated with a chemotherapy regimen. We compared the classification performance of our optimally discriminative SN markers against SN markers identified using heuristics and as well as numerous single gene (SG) markers. On average, the optimally discriminative SN markers show both the best and most stable classification performance in cross-dataset validation. We also show that our subnetwork method produces predictive markers that are more reproducible across independent cohorts and offer valuable insight into biological processes underlying response to therapy.