, 2011). Nevertheless, known marker genes (Lein et al., 2007) for layers 2/3, 4, 5, 6, and 6b demonstrated high concordance between individual samples and specific layers

(Figure 1B, Belgard et al., 2011). We compared our RNA-seq results with those previously obtained using microarrays for layer 6 and 6b from anterior cortex (putative S1) of postnatal day 8 mice (Hoerder-Suabedissen et al., 2009). RNA-seq levels for samples E and F were highly and significantly concordant with microarray expression levels for layers 6 and 6b despite methodological differences and the difference in age (Supplemental Experimental Procedures): 85% (147 of 173) of genes whose expression click here was found, with microarrays, to be significantly lower in layer 6 versus 6b also showed lower expression in sample E versus F; significant concordance was also found for 87% (385 of 441) of genes significantly lower in layer 6b versus 6, compared with sample F versus E (each test, p < 2 × 10−16, two-tailed binomial test relative to a probability of 0.5). We next predicted 6,734 “patterned” genes that are preferentially expressed in one or more layers and 5,689 “unpatterned” genes that were expressed more uniformly across all layers. For this, layer expression for 2,200 genes annotated from in situ hybridization images (see also Belgard

et al., 2011) was used for training a naive Bayes classifier for each layer 2–6b. (Annotated marker genes were insufficient to permit training of a reliable classifier for layer 1.) whatever These curations are generally consistent

with the literature and other SAHA HDAC in vivo expression data sets (Allen Institute for Brain Science, Top 1,000 Genes Analysis: Validation of Frequently Accessed Genes in the Allen Mouse Brain Atlas, http://mouse.brain-map.org/pdf/Top1000GenesAnalysis.pdf, 2010). A classifier was also constructed to separate patterned from unpatterned genes. Classifier generalization accuracies were assessed with 10-fold cross-validation (Figure 1C; Table 1; Figure S1), and smoothed calibration curves were constructed for the resulting predicted probabilities to arrive at accurate estimates of enrichment likelihood (Figure S2). Finally, these classifiers were applied to both known and previously unknown genes and transcripts (Table S2; Belgard et al., 2011). A total of 11,410 known genes (10,261 protein-coding) were expressed at sufficiently high levels for their layer patterning to be classifiable. Predicted layer expression patterns typically recapitulated both the literature (Figure 2A) and the results of the high-throughput curation-based approach (Table 1). Upon correcting for false positives and false negatives, we found that an estimated 5,835 of these 11,410 classifiable known genes (51%) were expressed preferentially in one or more layers (Table 1, Supplemental Experimental Procedures).