aeruginosa laboratory strain PAO1 was included in the dataset. The microarray dataset was prepared as matrix X which contains n (26) samples and m (5900) columns. We modeled the whole gene expression in a cell as a mixture of independent biological process

by using FastICA method [15]. The P. aeruginosa microarray data matrix X was decomposed by FastICA into latent variable matrix A (26 × 26) and gene signature matrix S (26 × 5900). Figure 1 Isolate sampling points and patient life span. P. aeruginosa isolates were collected from eleven different CF patients during a 35-y time period. Bacterial isolates are represented by the different symbols and patient life span is represented Mdivi1 solubility dmso gray bars. This figure is adapted from Yang et al., 2011 [8]. ICA improved clustering patterns of P. aeruginosa microarray data Unsupervised hierarchical clustering was applied to the original Tideglusib supplier normalized data, the outputs of ICA (latent variables) and the outputs of PCA (principle components), respectively. For the original data, the P. aeruginosa isolates were grouped into three distinct groups: an early stage infection group, a late stage infection group and a mucoid strain group (Figure 2). The early stage infection isolates were grouped together with the PAO1 strain, which indicates that they have not gained extensive adaptations. However, the clustering

did learn more not fully discriminate the early stage isolates (CF114-1973, CF105-1973 and CF43-1073, strain names marked in red color) of Yang’s study [8] from the early stage isolates (B12-0, B12-4, B12-7, B38-1, B38-2NM, B6-0 and B6-4, strain names marked in green color) from Rau’s study [5]. In contrast, the clustering dendrogram from ICA outputs showed better separation of the early stage isolates from the two different studies (Figure 3A). The CF114-1973 was clustered together with the CF105-1973 and CF43-1973 from the ICA outputs (Figure 3A). This indicates that these two groups of early stage isolates have distinct physiology. Clustering dendrogram from PCA outputs (Figure 3B) generated the same pattern as the one generated from the original data (Figure 2). These results showed

Etomidate that ICA is better than PCA in filtering noisy and extracting important features from microarray data. Figure 2 Hierarchical clustering of the normalized raw data using Euclidean distances. Red/green blocks represent signal increase/decrease respectively. Figure 3 Hierarchical clustering of the ICA and PCA outputs. (A) Hierarchical clustering of the ICA outputs with the last ‘common’ components of matrix A removed. (B) Hierarchical clustering of the principle components, with the number of the principle components k = 26. ICA identified significant genes for adaptation of P. aeruginosa to the CF airways The ICA output matrix A contains the weight with which the expression levels of the m genes contribute to the corresponding observed expression profile.