A new view of transcriptome complexity and regulation through the lens of local splicing variations

1) The manuscript would benefit from presenting the work in framework of previous developments in the field, comparing with them, and emphasizing the relative merits of the current contribution. The authors seem to imply that theirs is the first method that attempts to systematically classify and categorize AS events, beyond merely, exon skipping, alternative usage of splicing sites or intron retention. But, in my opinion, this is incorrect. The authors appear to ignore a large body of previous literature that actually attempts to address exactly this problem. Worth mentioning in this regard is the work of Nagasaki et al. (2005,

http://bioinformatics.oxfordjournals.org/content/22/10/1211.full) or Sammeth et al. (2008, http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1000147). In the latter work, in particular, analysis similar to those presented in Figure 3 in Vaquero-Garcia et al., regarding the frequency of complex LSVs depending on the annotation used and the species considered, are presented. Moreover the methods developed by Sammeth et al. have actually been used to categorize AS events in a number of publications (https://scholar.google.es/scholar?cites=8586199022962348855&as_sdt=2005&sciodt=0,5&hl=en). Also single source and single target LSVs are quantified in their recent work by Pervouchine et al. (2013, http://bioinformatics.oxfordjournals.org/content/29/2/273.long). The authors also claim that they obtain splicing information by combining the annotation and RNA-seq evidence (Figure 2A). It is not clear if they use only splice junctions, or also exons and transcript models. The latter is important for the definition of LSV. Indeed, if "the reference exon may include multiple 3' splice sites (ST-LSV) or 5' splice sites (SS-LSV)", as it is said in Materials and methods, it would be necessary to know novel exons as continuous units. If this is the case, a lot more has to be explained in Materials and methods regarding the procedure of exon discovery from RNA-seq. In summary, in contrast to the claims by Vaquero-Garcia et al., complex LSVs have received wide attention within the field.

The original manuscript lacked a section describing related work, which was mistakenly removed before submission. We apologize for this mistake and have now re introduced an extended section for this in the revised Discussion. We also added text in the Methods section to answer the question regarding how MAJIQ combines known annotation and de-novo junction to annotate putative de-novo exonic regions.

2) Another potential caveat is whether the results reported will persuade the general reader of the genuine potential of the approach to get biological insight.

Specifically:

A) The authors argue that complex LSVs represent over 30% of all transcripts variants. The general reader may wonder whether the switches in transcript isoforms within this category generally represent significantly large differences between tissues or biological conditions as to have biological impact. In other words, is biologically-relevant LSV likely to be limited to exceptional cases, or are the majority of complex variants likely to be functionally relevant? It would be useful, in this regard, to analyze the range and frequency distribution of ΔPSI values associated with complex LSVs compared to more standard alternative splicing events. Analysis of the conservation level of the predictably more relevant events could also be helpful.

The reviewers raise here an excellent point, which is also related to the concern raised below about disease relevance. To address the concern about significance of complex LSVs we followed the reviewers’ suggestion and created a new figure (Figure 5). Figure 5A is a meta analysis of over 30 datasets and 243 RNA-Seq experiments that shows the significant enrichment of complex LSVs in datasets that cover developmental stages, key splice factors, and human disease (AD) across diverse tissues and cell lines. Figure 5B shows the changes in LSV inclusion levels (dPSI) across these conditions, just as suggested by the reviewers. Finally, Figure 5C shows the suggested conservation analysis with a significant increase in intronic conservation for differentially spliced complex LSVs compared to their binary LSVs counterparts.

B) The authors predict that their method "will advance the ability to relate tissue-specific splice variation to genetic variation, phenotype and disease". They illustrate this by the identification of transcript variants possibly related to Alzheimer Disease, but the results are essentially descriptive and, considering the number of samples tested and the absence of validation in different cohorts, of limited value to extract rigorous biomedical conclusions. The manuscript would greatly benefit from additional efforts to more broadly illustrate the utility of the tool, for example to show how the approach is advantageous for explaining the impact of genetic variation using publicly available data from the GTEx project (e.g. pilot samples), ENCODE, modENCODE, etc.

To address the concern about the anecdotal nature of the disease related results in the original submission we performed the following: First, we followed the reviewers’ suggestion and validated these results on an independent data set (syn3157743 from the AMP-AD Target Discovery Consortium data portal) which is much larger: 157 samples from AS patient’s brains and 128 control samples, across three different brain sub regions. Second, we added a genome wide analysis of LSVs that are consistently differentially spliced between brain samples from patients and controls in the two datasets (included in Figure 5A described above). We very much agree this work is still just the tip of the iceberg and we believe much more can be done with the AD datasets and many others. Specifically, the suggested analysis of genetic variations with MAJIQ is a great direction to pursue in future work. However, we are very much aware that a careful analysis of this sort is beyond the scope of this paper.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

[…] The authors have done a good job discussing previous efforts by other scientists to identify complex alternative splicing decisions (main point #1 of the referees' report) and to illustrate the range and frequency distribution of complex LSVs compared to more standard alternative splicing events (point #2A). It would be helpful if the authors could more clearly state what was the validation rate of their disease-associated LSV changes in a second cohort of Alzheimer Disease patients (point #2B) (unless I have missed it, they only provide the number of LSVs reproducibly differentially spliced – p20 – but not what is the fraction of predictions validated in the two datasets). Another point: it is not clearly stated in the text/Figure 5A legend that the data corresponds to LSVs consistently differentially spliced between brain samples from patients and controls in the two datasets (as stated in the rebuttal).

We have worked hard to address the reviewers’ concerns and suggestions. In particular, to address the reviewers’ concern about the significance of complex LSVs we performed a whole new meta analysis of LSVs across 32 datasets covering diverse tissues, cell lines, developmental stages and splice factors knockdown. We added this to a conservation analysis as suggested by the reviewers, creating a new figure (Figure 5). In addition, to address the concern regarding the relevance of our findings to disease studies we added a genome wide analysis of differentially spliced LSVs in independent datasets of 285 samples from AD patients and controls, which validated our initial findings for the CAMK2 family and other genes. We also added a section detailing previous related work, which was erroneously omitted in the original submission, a mistake we apologize for. Finally, we made sure to add details in all sections indicated by the reviewers as unclear.