Associating Bphs-Related Genes Using FNTM

We associated genes with Bphs-related functions as described in Tyler et al.^{34 (link)}. Briefly, we used the connectivity weights in the Functional Network of Tissues in Mouse (FNTM) as features for training support vector machines^{75 (link)}. Each feature consisted of the connection weights from a given gene to genes in the functional module. To improve classification and reduce over-generalization we clustered each functional gene set into modules, each <400 genes^{45 (link)}. For each of these modules, we trained 100 SVMs to classify the module genes from a balanced set of randomly chosen genes from outside the module. We used 10-fold cross validation and a linear kernel. We also trained each SVM over a series of cost parameters identified by iteratively narrowing the cost parameter window to identify a series of eight cost parameters that maximized classification accuracy. We then used the training modules to score each positional candidate gene in the Bphse locus. To compare scores across multiple trained models, we converted SVM scores to false positive rates.