I have a reference genome and now I would like to call structural variants from Illumina pair-end whole genome resequencing data (insert size 700bp).

There are many tools for SV calls (I made an incomplete list of tools bellow). There is also a tool for merging SV calls from multiple methods / samples - SURVIVOR. Is there a combination of methods for SV detection with optimal balance between sensitivity and specificity?

There is a benchmarking paper, evaluating sensitivity and specificity of SV calls of individual methods using simulated pair-end reads. However, there is no elaboration on the combination of methods.

List of tools for calling structural variants:

  • $\begingroup$ I will just add a comment, as it's not a complete answer. Check the Genome in a Bottle consortium. There are discussions now on how to determine the best caller(s) and definition on a standard set of calls for benchmarking and testing new approaches. In my work I had good results with Socrates, now replaced with GRIDSS. $\endgroup$
    – nuin
    May 18, 2017 at 15:12
  • 1
    $\begingroup$ @nuin - I didn't know about Genome in a Bottle consortium, looks interesting, but I could not find any public record of a discussion. Do you have a link? $\endgroup$ May 20, 2017 at 15:10

2 Answers 2


I think the best method or combination of methods will depend on aspects of the data that might vary from one dataset to another. E.g. the type, size, and frequency of structural variants, the number SNVs, the quality of the reference, contaminants or other issues (e.g. read quality, sequencing errors) etc.

For that reason, I'd take two approaches:

  1. Try a lot of methods, and look at their overlap
  2. Validate a subset of calls from different methods by wet lab experiments - in the end this is the only real way of knowing the accuracy for a particular case.
  • $\begingroup$ What do you mean by "Validate a subset of calls from different methods"? $\endgroup$ May 20, 2017 at 14:37
  • $\begingroup$ Funny, the review paper (sim data) kind of agree with you - all methods have lot of false positives - it is better to take overlap. However, the paper "An integrated map of structural variation" (Germain's answer) have done opposite - tweaking every software to reduce false positives and than taking them all. $\endgroup$ May 20, 2017 at 15:14
  • $\begingroup$ By validate, I mean go in and resequence a bunch of predicted variants. All of the bioninformatic methods will give lots of false positives (and lots of false negatives, though it's harder to find the negatives of course). So the only genuine way to know the accuracy of your calls is to get independent information e.g. from resequencing. $\endgroup$
    – roblanf
    May 21, 2017 at 2:58

In case if you are really dedicated to obtain perfect results, you can use strategy described there, in 1000GP 3rd Phase SV detection paper - use these tools, validate your calls with IRS test, merge calls into one callset.

If you do not wanna spend thousands human-hours as was spent during this paper preparation, from my experience, it is better to use 1 paired-end insert distance method and one read-depth based method. Each of them cover "different" regions in the genome. (even thou they have huge overlap, paired-end detection requires both SV breakpoints to be located within the regions with good mappability which is not always the case, but resolution of read-depth methods is lower in general, paired-ends works well for deletions/tandem duplications/inversions, but have troubles with non-tandem duplications).

Hope it helps.

  • $\begingroup$ I went though the supplement of the paper, it is really crazy - it seems that there is very little overlap of different computational methods, but the False discovery rate is for non inversions types of SVs pretty good (2 - 10%). However, they have used a huge population dataset, I have sequencing of 9 individuals including the reference. $\endgroup$ May 20, 2017 at 14:39
  • $\begingroup$ Yeap, the overlap is kinda small, but this is mainly because different tools look for different type of CNVs (as I explained above comparing Read Depth vs Insert Distance-based methods). This is the best accuracy you can get. Having only 9 individuals, you'd better exclude CNVs that happen with frequency >1% in the population (see the paper) and then use eg DELLY and cn.mops (just an example, not sure if this is an ideal choice, but both of them are quite good). $\endgroup$ May 20, 2017 at 15:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.