Optimized prediction of extreme treatment outcomes in ovarian cancer

The TCGA ovarian cancer database shows that about 10% of patients respond poorly to platinum-based chemotherapy, with tumors relapsing in seven months or less. At the other extreme, another 10% or so enjoy disease-free survival of three years or more [1]. At present there are more than a dozen prognostic signatures that claim to predict the survival prospects of a patient based on her genetic profile. Yet, according to [2], none of these signatures performs significantly better than pure guessing. Accordingly, in this paper the objective is to propose and validate another gene-based signature. TCGA ovarian cancer data is analyzed using the “lone star” algorithm [3] that is specifically developed for identifying a small number of highly predictive features from a very large set. Using this algorithm, we are able to identify a biomarker panel of 25 genes (out of 12,000) that can be used to classify patients into one of three groups: super-responders (SR), medium responders (MR), and non-responders (NR). We are also able to determine a discriminant function that can divide patients into two classes, such that there is a clear survival advantage of one group over the other. This signature is developed using the TCGA Agilent platform data, and cross-validated on the TCGA Affymetrix platform data, as well as entirely independent data due to Tothill et al. [4]. The P-value on the training data is below machine zero, while the P-values on cross-validation are well below the widely accepted threshold of 0.05.