Application to the NIEHS (necrosis, cross-tissue) data set For the forward prediction, from blood to liver, the application of different batch effect removal selleck kinase inhibitor methods generally does not appear to affect the prediction performance. The backward prediction, from liver to blood, has poor prediction performance with or without the application of batch effect removal algorithms (Figure 8). This may in part be due to the fact that although blood gene signatures can be used to effectively predict liver necrosis, liver gene signatures do not have predictive power for necrosis measured in blood. This finding is consistent with the observations reported by Huang et al.

14 Application to the NIEHS (necrosis, cross-tissue-cross-platform) data set Without batch (group) effect removal, there is no predictive power either from blood (Agilent) to liver (Affymetrix), or vice versa, noting that all the MCC values are either zero or negative (Figure 9). In using data from blood (Agilent) to predict liver injury (Affymetrix), the applications of most batch effect removal methods enhance the prediction performance except the W-S combination for Refseq and Unigene mapping. When using data from liver (Affymetrix) to predict appropriate blood-based genes (Agilent), the application of batch effect removal methods yields both increased and decreased prediction performance. However, the two ratio-based methods consistently outperform the other methods. Similar to the cross-tissue results, we see that the blood samples have strong predictability of the liver necrosis with proper batch effect removal.

However, the predictability is much weaker for backward prediction, from liver to blood. In general, the sequence mapping slightly outperforms the other two mappings. Meta analysis To evaluate the general impact of batch effect removal in cross-batch (group) prediction performance, we calculate the increase of prediction performance value MCC after batch effect removal ��MCC=MCCAfter?MCCBefore. If ��MCC is greater than or lower than a threshold value, we say the batch effect removal has a positive or negative impact on the performance, respectively. If the difference in MCC after and GSK-3 before batch impact removal is less than the threshold, we say the impact is negligible. For simplicity, the threshold is chosen as 0.05. Figure 10 shows the percentages of cases with increased, decreased and unchanged predictive performances (with sum 100%) over the 120 cases explored. We find that, for each batch effect removal method, the number of cases with increased predictive performance is greater than that with decreased predictive performance, indicating that, in general, batch effect removal has a positive impact on prediction performance.