It is not computationally feasible to compute all possible gene clusters from the dataset with GR, so heuristics are necessary to find tightly co-regulated gene clusters. Our extensions permit genome-wide use of the GR by automatically selecting putative co-regulated genes, running the algorithm, and then recursively modifying the query using leave-one-out cross validation (LOOCV) as a scoring function until the cluster converges at a point where all query genes contribute approximately equally to the cluster. The procedure is as follows. First, using each gene as a seed for a potential cluster, initial predictions of co-regulated genes are made using Spearman's rank correlation. If the number of highly correlated genes is less than 5 or more than 20, the putative cluster is expanded or trimmed to the most correlated 5 or 20 genes, respectively. Next, the Gene Recommender algorithm is run with this initial gene cluster, and then rerun using the top hits (by s.g.i, the Gene Recommender normalized correlation metric) from the initial run. If the seed gene scores highly after the second run (i.e., within the top 50 genes
