Figure 3

Block diagram of TMpro prediction. Primary sequence of protein (amino acid sequence) is input to the system. (A) maps it to 5 amino acid property se-quences. The output has size 5 x L (rows x columns). These 5 sequences form input to (B) which performs windowed analysis, and outputs a matrix of counts of 10 properties (C1 to C10) for each window position. This output has the size 10xL-l+1. The outputs from (B) for all proteins are collected together and singular value decomposition is performed by C. During testing phase, an SVD approximation is performed for the matrix of a single test protein. The output of this block (C) forms the final features used by neural network. (D) Features are evaluated by the NN model and a 2-state prediction is output for each residue (TM, non-TM). An analog value also is output ranging from -1 to 1 that indicates the closeness of the feature to non-TM or to TM correspondingly.