How is TAD calling affecting by read depth?

[blmoore]

Correction 1 from examiner's report:

The Chi-squared tests for calling TADs are potentially strongly affected by read depth (increasing power to call a TAD with increasing sequencing depth). Given the differences in sequencing depth between samples, this issue should be discussed. Does the HMM method help to mitigate this? The thesis should include some discussion of the impact of read depth and statistical power on TAD calling (e.g. in reference to Fig. 16 – though it may be preferable to include a figure showing TAD sizes in the different cell lines to illustrate this effect more explicitly)

[blmoore]

TAD size distribution plot added here: https://github.com/blmoore/3dgenome/commit/222b9b12ad00912bde9ce847d034ef6bb8ad375d

[blmoore]

Figure + sentence added in 8583d87cd754bb9b6617894f97d0f3877df104f9

Paragraph now reads:

The Dixon et al.[2] method of calling TADs relies on the detection of boundaries,[4] thus it is affected by sequencing depth: experiments with sparser contact matrices may not contain enough for a sufficiently high degree of bias to allow a boundary call. This is evident in our datasets even after normalisation, with the deeply-sequenced H1 hESC cell type having approximately 50% more TADs called than in the GM12878 cell type (Fig. 9). The increased power to detect TAD boundaries also results in smaller domains, on average, in the H1 hESC cell line (Fig. 10). This effect could have been mitigated by down-sampling reads in the H1 cell type, but at a cost of reducing the quality of the best dataset under study. Instead this disparity should just be noted as a potential cofounder in downstream TAD analysis; at lower-resolution such as that used to calculate compartment eigenvectors (1 Mb) this sensitivity to sequencing depth is not evident (Figs. 2, 3).