CPAT: Coding-Potential Assessment Tool using an alignment-free logistic regression model

Abstract

Thousands of novel transcripts have been identified using deep transcriptome sequencing. This discovery of large and 'hidden' transcriptome rejuvenates the demand for methods that can rapidly distinguish between coding and noncoding RNA. Here, we present a novel alignment-free method, Coding Potential Assessment Tool (CPAT), which rapidly recognizes coding and noncoding transcripts from a large pool of candidates. To this end, CPAT uses a logistic regression model built with four sequence features: open reading frame size, open reading frame coverage, Fickett TESTCODE statistic and hexamer usage bias. CPAT software outperformed (sensitivity: 0.96, specificity: 0.97) other state-of-the-art alignment-based software such as Coding-Potential Calculator (sensitivity: 0.99, specificity: 0.74) and Phylo Codon Substitution Frequencies (sensitivity: 0.90, specificity: 0.63). In addition to high accuracy, CPAT is approximately four orders of magnitude faster than Coding-Potential Calculator and Phylo Codon Substitution Frequencies, enabling its users to process thousands of transcripts within seconds. The software accepts input sequences in either FASTA- or BED-formatted data files. We also developed a web interface for CPAT that allows users to submit sequences and receive the prediction results almost instantly.

Figure 1.

Score distribution between coding (red) and noncoding (blue) transcripts for the four linguistic features selected to build the logistic regression model; training data set containing 10 000 coding and 10 000 noncoding transcripts were used. (A) ORF size. (B) ORF coverage. (C) Fickett score (TESTCODE statistic). (D) Hexamer usage bias measured by log-likelihood ratio.

Figure 2.

Three-dimensional plot shows combinatorial effects of Fickett score, hexamer score and ORF size on 10 000 coding genes (red dots) and 10 000 noncoding genes (blue dots).

Figure 3.

Performance evaluation using 10-fold cross-validation. Dashed curves represent the 10-fold cross-validation; solid curves represent the averaged curve from 10 validation runs. (A) ROC curve. (B) PR curve. PPV = positive predictive value, TPR = true positive rate. (C) Accuracy versus cutoff value. (D) Two-graph ROC curve is used to determine the optimum cutoff value.

Figure 4.

Performance comparison between CPAT, CPC, PhyloCSF and PORTRAIT using ROC curves.

Figure 5.

Cumulative curves of coding-potential assessment score for (A) CPAT, (B) PORTRAIT, (C) CPC and (D) PhyloCSF.