Gene prediction by spectral rotation measure: a new method for identifying protein-coding regions

Abstract

A new measure for gene prediction in eukaryotes is presented. The measure is based on the Discrete Fourier Transform (DFT) phase at a frequency of 1/3, computed for the four binary sequences for A, T, C, and G. Analysis of all the experimental genes of S. cerevisiae revealed distribution of the phase in a bell-like curve around a central value, in all four nucleotides, whereas the distribution of the phase in the noncoding regions was found to be close to uniform. Similar findings were obtained for other organisms. Several measures based on the phase property are proposed. The measures are computed by clockwise rotation of the vectors, obtained by DFT for each analysis frame, by an angle equal to the corresponding central value. In protein coding regions, this rotation is assumed to closely align all vectors in the complex plane, thereby amplifying the magnitude of the vector sum. In noncoding regions, this operation does not significantly change this magnitude. Computing the measures with one chromosome and applying them on sequences of others reveals improved performance compared with other algorithms that use the 1/3 frequency feature, especially in short exons. The phase property is also used to find the reading frame of the sequence.

Figure 1

Computing U_b(N/3) in the case f_minf(b,1).

Figure 2

Argument distributions for all experimental genes in all chromosomes in S. cerevisiae.

Figure 3

Argument distribution for noncoding regions in all chromosomes in S. cerevisiae.

Figure 4

(A) Argument distribution for all experimental genes in all chromosomes of S. cerevisiae. (B) Argument distribution for all genes in chromosomes 2 and 3 of S. pombe. (C) Argument distribution for all genes in chromosome 1 of Guillardia theta.

Figure 5

(A) Rotation and alignment of the vectors G(s) and T(s), when arg(T[s]) ≈ μ_T and arg(G[s]) ≈ μ_G. (B) Rotation and alignment of the vectors G(s) and T(s), when arg(T[s]) and arg(G[s]) are any random values.

Figure 5

(A) Rotation and alignment of the vectors G(s) and T(s), when arg(T[s]) ≈ μ_T and arg(G[s]) ≈ μ_G. (B) Rotation and alignment of the vectors G(s) and T(s), when arg(T[s]) and arg(G[s]) are any random values.

Figure 6

Probability density functions for Spectral Rotation (bold) and Spectral Content (fine) measures (solid lines represent exons and dashed lines represent noncoding regions).

Figure 7

Argument distribution of coding DNA strands of length 120 bp in S. cerevisiae.

Figure 8

Graphs of gene prediction applied on the gene SPBC582.08 in chromosome 2 of S. pombe, using a sliding window of 180 bp: (A) TG-Rotation measure; (B) Codon Usage measure. The horizontal segments represent the actual location of the three exons. To get the actual base location in the chromosome, add 300,000 bp to the numbers on the horizontal axis.

Figure 9

Graphs of the SR measure on the gene SPBC1685.08 in chromosome 2 of S. pombe, using a sliding window of 351 bp. (A) The measure; (B) arg(V). The horizontal segments represent the actual location of the exons. To get the actual base location in the chromosome, add 400,000 bp to the numbers on the horizontal axis.

Figure 10

Graphs of the SR measure on the gene SPBC1709.08 in chromosome 2 of S. pombe, using a sliding window of 351 bp. (A) The measure; (B) arg(V). The horizontal segment represents the actual location of the gene. To get the actual base location in the chromosome, add 1,000,000 bp to the numbers on the horizontal axis.