Is there a point in recalibration of scores for variant calling?

Question

The most variant calling pipeline GATK include a Base Quality Score Recalibration (BQSR) which requires a list of known variants. Recently, some work has been done for reference-free recalibration of scores as well: Lacer and atlas, which is motivated by making the most for aDNA and low coverage datasets.

The importance for aDNA is explained in this lecture, but it is not clear to me if / how is important BQSR is for fresh DNA samples with decent (>15x) coverage. Especially when I work with non-model organisms and I can not simply use the standard tools.

How big an impact does recalibration of scores have on variant calling? Is there a rule of thumb for which it is / it is not worth the effort?

Answer 1

I personally don't think BQSR has a huge impact on variant calling, but you don't really need to guess. If you run GATK BQSR, it outputs a table and charts of exactly how much quality scores are adjusted. The adjustment will vary depending on the position in the read and genomic context (previous and following base). In my experience, the difference is a few points at most, but it's certainly noticeable.

GATK recommends BQSR for both genome and exome data, which is normally much higher than 15x.

Answer 2

That's a good question.

I'd say that you don't need to bother with variant recalibration for

• low number of samples (e.g., just two trios); I could not get GTAK recalibration of variant scores to work anyway
• high-coverage samples (e.g., X Ten genomes with 30x coverage) where the DNA samples themselves are of high, comparable quality and have been sequenced with consistent technology.

Generally, it is my impression that a lot of the thoughts and advanced statistical models built into GATK come from the earlier phases of the 1000 Genomes project. This means (1) low-coverage, (2) different coverage genomes (3) sequenced with varying technology versions by (4) different samples and (5) population sequencing.

If you are in a clinical setting where you do 30x sequencing on X Ten platforms only anyway, then variant recalibration will probably not help you that much.

On the other hand, if you are integrating many data sets from different data centers and machine versions etc., variant recalibration might be worth a shot.

A good check would be looking at genotype quality distributions and other variant/quality related metrics before and after recalibration.

Anyone: please correct me if I'm wrong!

Answer 3

Ideally these BQSR methods were made keeping in mind how technical errors will actually screw up the base quality calls and when the machines were still more on development phase while being used for the 1000G project. As of now machines are more powerful and strong where it will be unlikely to use it but still we use with listed SNPs to find the covariates and build a model around the data using the information with machine learning tricks to improve the quality of those base calls. Ideally it should be more appropriate when old machines from Illumina or other standard companies are being used but with new machines which are much powerful and having high throughput they should tend to go down. I do not recall if such tests have been made but obviously I know new sequencing machine always make such tests to show that they have reduced such errors but still recommend such BQSR for variant calls. Now the problem is the list of SNPs, this to me is the real problem since the list we use is far from being Gold standard and if that is not properly taken care of everything we infer about quality is still shaky. This link is pretty informative but it's an old one. I would really see improvements with new sequencers. However very less people care about such tests in academic research and also translational lab will really not invest time and money on such unless the facility has some bioinformaticians who always does such testing while buying a new sequencer for the institute. In terms of clinical genomics for finding variants I reckon most powerful and up-to-date sequencers should be used but not sure if they still use BQSR and if so what is the list they use to build model of covariation around the data.

Answer 4

In case BQSR is not an option (i.e. non-model organisms) it would be best to use some internal control sequence such as PhiX for illumina platform. Although this is supposed to be common practice some facilities ignore it. In principle, the machines should use these sequences as reference so that the scoring would be more accurate. In my experience the first 10-15 bases of the illumina reads always had lower quality. This can easily be seen in the nucleotide distributions. I would advise trimming the first 10-15 bases and quality based end-trimmingiIf the quality of the individual reads are important, such as low coverage resequencing or de-novo genome assembly applications.