My impression is that small InDel (a couple of bp) is identified through cigar string in BAM...
Well, yes, by gaps in the alignment, easily detected by examining the CIGAR string.
...and typical CNV (at least thousands of bp) is detected through read depth.
Also by deviations from expected read pair distance or orientation.
What about InDel or CNV with size these them, say hundreds of bp? These CNV is too long to be covered by read and too short to be detected by statistically comparing the difference of read depth. Do we have consensus that this kind of CNV is hard to detect?
Yes, what you describe has certainly been recognized and acknowledged in the literature. It was one of the primary motivations for a recently published paper I contributed to on de novo (germline) variant discovery. Here's an excerpt from that paper:
In a reference-mapping context, calling indels with confidence requires accurate mapping of each read spanning the indel, with all gaps arranged consistently. This is possible only for short indels and tends to be prone to error and misalignment. Thus prediction of indels with length 10 bp has proved to be very challenging and accompanied by high false-positive and false-negative rates. Furthermore, the prediction of SVs via read mapping is only possible through indirect signatures such as alterations in read depth or read-pair signatures. These signatures can be quite noisy and result in high rate of false-negative and false-positive prediction.
As for your question:
is there a common strategy of detection for them?
Numerous mapping-free k-mer based methods for variant detection are coming up in the literature, and all seem to perform pretty well. The trick is how to define which k-mers are interesting—then you grab those k-mers (or the reads containing them) and assemble them into variant-spanning contigs. A few examples:
- In the Kevlar paper, we defined "interesting" k-mers as those with high abundance in an individual of interest but zero (or very low) abundance in the individual's parents.
- The HAWK paper, which describes a k-mer based GWAS-like method, defines "significant" k-mers as those significantly over- or under-represented in one class of samples versus a second class.
- The NovoBreak paper compares paired tumor and normal WGS data to find k-mers that span mutatino breakpoints in cancer genomes.
Do we generally think them having impact on protein function (especially in oncology background)?
There is definitely potential that indels can have a substantial impact on protein function. But since their detection has been so difficult for so long, a lot remains unknown about their actual collective impact on function.