Computational technique for improvement of the position-weight matrices for the DNA/protein binding sites

Abstract

Position-weight matrices (PWMs) are broadly used to locate transcription factor binding sites in DNA sequences. The majority of existing PWMs provide a low level of both sensitivity and specificity. We present a new computational algorithm, a modification of the Staden-Bucher approach, that improves the PWM. We applied the proposed technique on the PWM of the GC-box, binding site for Sp1. The comparison of old and new PWMs shows that the latter increase both sensitivity and specificity. The statistical parameters of GC-box distribution in promoter regions and in the human genome, as well as in each chromosome, are presented. The majority of commonly used PWMs are the 4-row mononucleotide matrices, although 16-row dinucleotide matrices are known to be more informative. The algorithm efficiently determines the 16-row matrices and preliminary results show that such matrices provide better results than 4-row matrices.

Figure 1

The schematic presentation of TP, FP, TN and FN. The number N_orig in the left rectangle represents the amount of sites recognized by the original matrix as respective TFBS among all considered sites N_total ×l_w in the given window l_w in all promoter sequences N_s from the training dataset. The number TP is the amount of sites recognized by the new matrix among N_orig so the rest of N_orig is FN. The number of sites recognized by the new matrix but not included in N_orig is FP. Finally, the number TN is the total number of considered sites N_total minus TP, FP and FN.

Figure 2

The flowchart of optimization process.

Figure 3

The occurrence frequency (the percentage of sequences having a considered motif centered at particular position) distribution of the GC-box sites found by the original matrix. The distribution is based on scanning of DBTSS (magenta, positive strand; red, negative strand; dark blue, both strands) and EPD (green, both stands) sequences. The value at each position is an 11 point sliding average. The TSS is placed at position +1. The straight horizontal line depicts the average amount of GC-box sites found in both strands of the randomly generated sequence with the same percentage of each of 4 nt as in the training set of promoter sequences, namely 20.6% for A and T, and 29.4% for C and G. The shadow rectangles indicate SD calculated based on 1871 random sequences (short rectangle) and on 8973 random sequences (long rectangle), respectively.

Figure 4

The sensitivity/specificity ratios for the original and new matrices. The averaged occurrence frequency of GC-box sites found by the original matrix (circle at the left upper corner) and two sets of new 4-row (squares) and 16-row (diamonds) matrices in the randomly generated sequence with the same percentage of each of four nucleotides as in the training set of promoter sequences versus sensitivity. The x-axis is the percentage of recognized sites from a control set of experimentally defined sites.

Figure 5

The occurrence frequency distribution of the GC-box sites found by the 16-row PWM with maximal sensitivity. The occurrence frequency distribution of the GC-box sites based on scanning of DBTSS (magenta, positive strand; red, negative strand; dark blue, both strands) and EPD (green, both stands) sequences. The rest is as in Figure 3.

Figure 6

The occurrence frequency of GC-box sites versus occurrence frequency of known genes in chromosomes of human genome. The OF of sites were obtained by 16-row matrix with maximal sensitivity. The diamonds and squares show the averaged OF of each chromosome in whole and conserved sequences, respectively.