PIC-Me: paralogs and isoforms classifier based on machine-learning approaches

Abstract

Background: Paralogs formed through gene duplication and isoforms formed through alternative splicing have been important processes for increasing protein diversity and maintaining cellular homeostasis. Despite their recognized importance and the advent of large-scale genomic and transcriptomic analyses, paradoxically, accurate annotations of all gene loci to allow the identification of paralogs and isoforms remain surprisingly incomplete. In particular, the global analysis of the transcriptome of a non-model organism for which there is no reference genome is especially challenging.

Results: To reliably discriminate between the paralogs and isoforms in RNA-seq data, we redefined the pre-existing sequence features (sequence similarity, inverse count of consecutive identical or non-identical blocks, and match-mismatch fraction) previously derived from full-length cDNAs and EST sequences and described newly discovered genomic and transcriptomic features (twilight zone of protein sequence alignment and expression level difference). In addition, the effectiveness and relevance of the proposed features were verified with two widely used support vector machine (SVM) and random forest (RF) models. From nine RNA-seq datasets, all AUC (area under the curve) scores of ROC (receiver operating characteristic) curves were over 0.9 in the RF model and significantly higher than those in the SVM model.

Conclusions: In this study, using an RF model with five proposed RNA-seq features, we implemented our method called Paralogs and Isoforms Classifier based on Machine-learning approaches (PIC-Me) and showed that it outperformed an existing method. Finally, we envision that our tool will be a valuable computational resource for the genomics community to help with gene annotation and will aid in comparative transcriptomics and evolutionary genomics studies, especially those on non-model organisms.

Keywords: Alternative splicing; Gene duplication; Isoforms; Machine learning; Paralogs; RNA-Seq.

Fig. 1

Illustration of the five features. A Two amino acid sequences (sequence1 and sequence2) are aligned. Matches, mismatches, and gaps between the two sequences are colored in black, red, and yellow, respectively. Green underlining indicates the consecutive identical or non-identical blocks (CB). B Sequence similarity (SS) is the percentage of matched sequences in the aligned sequences. C Inverse count of consecutive identical or non-identical blocks (ICCB) is the inverse count of CB. D Match-mismatch fraction (MMF) indicates the overall number of consecutive matches and mismatches–namely, it is the sum of the sequence lengths minus one of all CB divided by the length of the alignment. E Twilight zone (TZ) is a range of sequence similarity; a 20% cut-off score was used here. F Expression level difference (ELD) is the difference in the expression levels between two genes. From the example alignment in A, the scores of the first three features were 0.625 for SS, 0.091 for ICCB, and 0.656 for MMF, which are detailly described in Additional file 1: Fig. S1

Fig. 2

Paralogs and isoforms are poorly classified using pre-existing sequence features. The distributions of SS, ICCB, and MMF are shown in panels A, B, and C, respectively. Panel D illustrates all three features at the same time. Samples derived from paralogs and isoforms in the human brain data are shown in blue and red, respectively. The same tests for the other datasets are shown in Additional file 2: Fig. S2–Additional file 5: Fig. S5

Fig. 3

Paralogs and isoforms, especially those with a short gene length, are indistinguishable using the SS, ICCB, and MMF features. A–C Scatterplots illustrating the combinations of the mean lengths of the pair sequences and each feature. D Expression level differences between the isoform and paralog groups. The central line and lower and upper edges of the box indicate the median and 25th and 75th percentiles, respectively. The whiskers extend to the furthest point within 1.5 times the interquartile range (IQR). P-values were calculated using the Mann–Whitney U test

Fig. 4

Classifying paralogs and isoforms using machine learning methods. a AUC comparison between the SVM and RF models using nine RNA-seq datasets from human, zebrafish, and wheat tissues. AL is the aleurone layer, TC is transfer cells, and WE is whole endosperm. b Performance assessment of our method, PIC-me, and a pre-existing method, IsoSVM. Accuracy, positive predicted value (PPV), negative predicted value (NPV), and MCC were calculated as follows: Accuracy = (TP + TN)/(TP + FP + TN + FN), PPV = TP/(TP + FP), NPV = TN/(TN + FN), and MCC = $\frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$, where TP and TN are true positive and true negative, respectively, and FP and FN are false positive and false negative, respectively. P-values were calculated using the Mann–Whitney U test. Error bars indicate the standard error of the mean