This comes back to the core purpose of MD tags: they only represent information about the read aligned to the reference. Immediately, the surprising part is that the MD tag represents 100 matches to the reference, but the read length is 101 bases and the CIGAR string is 101. Let’s look at an example: read seq length: 101 Note that insertions, since they don’t represent a loss of information about the reference, are not stored in MD flag. Here, there are 85 matches, 1 deletion from the reference (the reference has an A there and the read doesn’t), and then there are 16 matches. The MD tag stores this information too: 85^A16 Likewise, a reference would be necessary to know the deleted bases from the reference in an alignment. Note that 0s are use used to indicate positions of neighboring SNPs. A SNP looks like: 10A3T0T10ġ1 bases in from the aligned portion of the read, the reference has an A and the read has what ever base is at the 10th position (excluding softclips).ġ5 bases in there’s a T in the reference.ġ6 bases in there’s a T in the reference. Thus, this information is carried in the MD tag. A SNP’s alternate base is carried in the read, but without the MD tag or use of the alignment reference, it’s impossible to know what the reference base was. It does this by carrying information about the reference that the read does not carry, for a particular alignment. The MD tag is for SNP/indel calling without looking at the reference.

There’s not too much discussion online about this, so I took some notes as I went through examples. I needed to work with the MD tag in BAM/ SAM files for a recent project.

Bam file format