Figure 4

(A) Data preprocessing and feature extraction: A sample sequence is shown with its property annotations. Line (1) header of the protein (potassium channel Kcsa). Line (2) primary amino acid sequence. Line (3) Topology: nonmembrane ‘-’ and membrane ‘M’. Line (4) Charge: posi-tive ‘p’, negative ‘n’ and neutral ‘-’. Line (5) Polarity: polar ‘p’ and nonpolar ‘-’. Line (6) Aro-maticity: aromatic ‘R’, aliphatic ‘-’ and neutral ‘,’. Line (7) Size: small ‘.’, medium ‘o’ and large ‘O’. Line (8) Electronic property: strong acceptor ‘A’, weak acceptor ‘a’, neutral ‘.’, weak donor ‘d’ and strong donor ‘D’. Line (9): topology again. A window of width 16 residues is moved across the sequence from left to right, one residue at a time. At each position the different prop-erty-feature combinations (such as “charge-negative”, size “medium”) in the window are counted. The collection of these counts in a vector forms the feature at that position. In the ex-ample shown above, the window width is shown as 16 residues. In the analyses, the width used for HMM modeling is 6 residues and that for NN modeling is 16 residues. If the length of the protein is L residues, and window length is l residues, the number of feature vectors obtained is: L-l+1. The three shaded windows at positions 1, 23 and 50 have labels “completely non-TM”, “completely TM” and “mixed” correspondingly. (B) Feature vectors: Feature vectors of the sequence corresponding to each of the window posi-tions are shown. The 10 rows of property number correspond to the Cij list of Eqn. 2. The win-dow position refers to the residue number of the first residue in the window. Feature vectors cor-responding to the blue, red and yellow windows in (A) are shown in their corresponding color in the table. The class label of the feature vector is shown in the last row: completely nonmembrane -1, membrane 1.