AIDE: annotation-assisted isoform discovery with high precision

Abstract

Genome-wide accurate identification and quantification of full-length mRNA isoforms is crucial for investigating transcriptional and posttranscriptional regulatory mechanisms of biological phenomena. Despite continuing efforts in developing effective computational tools to identify or assemble full-length mRNA isoforms from second-generation RNA-seq data, it remains a challenge to accurately identify mRNA isoforms from short sequence reads owing to the substantial information loss in RNA-seq experiments. Here, we introduce a novel statistical method, annotation-assisted isoform discovery (AIDE), the first approach that directly controls false isoform discoveries by implementing the testing-based model selection principle. Solving the isoform discovery problem in a stepwise and conservative manner, AIDE prioritizes the annotated isoforms and precisely identifies novel isoforms whose addition significantly improves the explanation of observed RNA-seq reads. We evaluate the performance of AIDE based on multiple simulated and real RNA-seq data sets followed by PCR-Sanger sequencing validation. Our results show that AIDE effectively leverages the annotation information to compensate the information loss owing to short read lengths. AIDE achieves the highest precision in isoform discovery and the lowest error rates in isoform abundance estimation, compared with three state-of-the-art methods Cufflinks, SLIDE, and StringTie. As a robust bioinformatics tool for transcriptome analysis, AIDE enables researchers to discover novel transcripts with high confidence.

Figure 1.

Workflow of the stepwise selection in the AIDE method. Stage 1 starts with a single annotated isoform compatible with the most reads, and all the other annotated isoforms are considered as candidate isoforms. Stage 2 starts with the annotated isoforms selected in stage 1, and all the possible isoforms, including the unselected annotated isoforms, are considered as candidate isoforms. In the forward step in both stages, AIDE identifies the isoform that mostly increases the likelihood, and it uses the LRT to decide whether this increase is statistically significant. If significant, AIDE adds this isoform to its identified isoform set; otherwise, AIDE keeps its identified set and terminates the current stage. In the backward step in both stages, AIDE finds the isoform in its identified set such that the removal of this isoform decreases the likelihood the least, and it uses the LRT to decide whether this decrease is statistically significant. If not significant, AIDE removes this isoform from its identified set; otherwise, AIDE keeps the identified set. After the backward step, AIDE returns to the forward step. AIDE stops when the forward step in stage 2 no longer adds a candidate isoform to the identified set.

Figure 2.

Gene-level isoform discovery and abundance estimation results of AIDE, Cufflinks, StringTie, and SLIDE on simulated RNA-seq data with 10 × coverage. Each box gives the first quantile, median, and third quantile of the gene-level accuracy given each of the nine sets of synthetic annotations. (A) Precision rates in isoform discovery; (B) recall rates in isoform discovery; and (C) error rates (defined as one-half of the sum of the absolute differences between the true and estimated isoform proportions) in abundance estimation.

Figure 3.

Comparison between AIDE and the other three isoform discovery methods in simulation. Given each synthetic annotation set, we applied AIDE, Cufflinks, StringTie, and SLIDE for isoform discovery and summarized the expression levels of the predicted isoforms using fragments per kilobase million reads mapped (FPKM) units. Then the precision-recall curves were obtained by thresholding the FPKM values of the predicted isoforms. The AUC of each method is also marked in the plot. The results shown are based on RNA-seq data with a 10 × coverage.

Figure 4.

Comparison of AIDE and the other three methods using real data. (A) Exon-level accuracy in the human ESC samples; (B) exon-level accuracy in the mouse BMDM samples; (C) transcript-level accuracy in the human ESC samples; and (D) transcript-level accuracy in the mouse BMDM samples. The gray contours denote the F₁-scores, as marked on the right of each panel.

Figure 5.

Evaluation of isoform discovery methods based on long reads. The F₁-score, precision, and recall of the four discovery methods were calculated at the base, exon, and transcript levels. (A) Evaluation based on isoforms identified by ONT. (B) Evaluation based on isoforms identified by PacBio.

Figure 6.

Experimental validation of isoforms predicted by AIDE and Cufflinks. Isoforms of genes MTHFD2 (A), NPC2 (B), RBM7 (C), CD164 (D), FGFR1 (E), and ZFAND5 (F) were validated by PCR and Sanger sequencing. The isoforms to validate (yellow) are listed under each gene (dark gray), with +/− indicating whether an isoform was/was not identified by PCR or a computational method. The forward (F) and reverse (R) primers are marked on top of each gene. For each gene, the agarose gel electrophoresis results show the molecular lengths of PCR products.

Figure 7.

AIDE identifies isoforms with biological relevance. (A) PCR experiments validated the expression of FGFR1-238 in breast cancer cell lines MCF-7, SUM149, BT474, SK-BR-3, MDA-MB-231, and BT549. (B) Long-term colonegenic assay with Lipofectamine 3000 controls (“siControl”) and FGFR1-238 knockdowns. Tumor growths relative to the siControl were quantified by the ImageJ software (Schneider et al. 2012). (C) FGFR1 isoforms identified by AIDE and Cufflinks. (D) Long-term colonegenic assay with siControl (negative control), si-FGFR1-238 (positive control), si-FGFR1-205, and si-FGFR1-C1. Tumor growths relative to the siControl were quantified by the ImageJ software. (E) NRAS isoforms in the GENCODE annotation, reported by Eisfeld et al. (2014) and discovered by AIDE, Cufflinks, or StringTie in three melanoma BRAF inhibitor–resistant cell lines: M229R, M263R, and M395R.

Figure 8.

Spearman's correlation coefficients between the estimated isoform expression and the benchmark NanoString counts. (A) For every probe, the sum of the expression levels of its corresponding isoforms is used in the calculation. (B) For every probe, the maximum of the expression levels of its corresponding isoforms is used in the calculation.