75% for LDA), sensitivity (82% vs. 72%), and specificity (85% vs. 75%). When decreased liver weight was targeted, CBA scored better accuracy (86% vs. 73%) and sensitivity (22% vs. 6%), while LDA marked CHIR-99021 ic50 better specificity (90% vs. 95%). We also compared between CBA and CBA-DR (CBA without Default Rule), our modified version of the original CBA (Table 2). CBA-DR does not predict if a sample does not match any rule except the default rule in a classifier, and, in turn, return a ‘hold’. When increased liver weight was targeted, CBA-DR marked lower accuracy (83% for CBA vs. 79% for CBA-DR) and specificity (85% vs. 29%) and higher sensitivity
(82% vs. 100%). When decreased liver weight was targeted, CBA-DR marked lower sensitivity (22% for CBA Tofacitinib vs. 0% for CBA-DR) and higher accuracy (86% vs. 95%) and specificity (90% vs. 100%). We compared
the form of generated classifiers between CBA and LDA (Figure 1), when all the records were used as a training set for increased liver weight. CBA tells us a set of rules, arranged in order of confidence. Each rule consists of an antecedent, which is an itemset in the form of (non-class attribute, its discretized value), and a consequence in the form of (class attribute, its class label), shown after “- > ” here. On the other hand, LDA tells us a single discriminative function (fd), which is a polynomial of non-class attribute values with their coefficients. Coefficients in a discriminative function of LDA reflect discriminative power of each non-class attribute (gene, here), with higher positive values and lower negative values meaning larger contributions to each corresponding class label of a class attribute (liver weight, here). To look
into how biologically reasonable the CBA-generated classifier is, we conducted the canonical pathway analysis for the set of genes selected in the classifier when all the records were used as a training set for increased liver weight (Table 3) (for brevity, only top 10 pathways in order of -logp are shown). Because LDA itself, in contrast to CBA, does not explicitly select a set of genes in building a classifier, we did not compare CBA with LDA here. We could assume that the most Non-specific serine/threonine protein kinase significant pathways involved with the genes in our classifier were mainly drug metabolism-related ones, such as Xenobiotic Metabolism Signaling, LPS/IL-1 Mediated Inhibition of PXR Function, PXR/RXR Activation etc. Figure 2A is an excerpt around the NRF2 molecule from the illustration of the Xenobiotic Metabolism Signaling pathway, exported from IPA. NRF2 is a key modulator of oxidative stress responses. In response of oxidative stress, NRF2 is released into the nucleus and up-regulates downstream antioxidant enzymes, mainly drug metabolism enzymes.