Diploid Copy Number Estimation

Pipeline step: compute_dipcn (Step 4)

Module: grid.utils.compute_dipcn

Raw read counts in the target VNTR region depend on both the true copy number and each sample’s sequencing depth. To isolate the copy number signal, GRiD normalizes each sample’s counts relative to its nearest neighbors — individuals with similar genome-wide depth profiles (and therefore similar technical characteristics). This step is fully locus-agnostic.

Notation

For sample \(i\):

  • \(r_i\) — read count in the target VNTR region

  • \(s_i\) — depth scale (mean mosdepth coverage, stored in the neighbors file)

  • \(\mathcal{N}_i\) — set of top-\(k\) nearest neighbors (default \(k = 200\))

For each neighbor \(j \in \mathcal{N}_i\):

  • \(r_j\) — neighbor’s read count

  • \(s_j\) — neighbor’s depth scale

Depth-Scaled Normalization

Each read count is first divided by its sample’s depth scale to remove the effect of sequencing depth:

\[\tilde{r}_i = \frac{r_i}{s_i}\]

The neighbor mean of depth-scaled counts provides a cohort-matched reference:

\[\bar{r}_{\mathcal{N}_i} = \frac{1}{|\mathcal{N}_i|} \sum_{j \in \mathcal{N}_i} \frac{r_j}{s_j}\]

Diploid Copy Number Estimate

The diploid copy number estimate is the ratio of the sample’s depth-scaled count to the neighbor mean:

\[\widehat{\mathrm{dipCN}}_i = \frac{\tilde{r}_i}{\bar{r}_{\mathcal{N}_i}} = \frac{r_i / s_i}{\dfrac{1}{|\mathcal{N}_i|} \displaystyle\sum_{j \in \mathcal{N}_i} r_j / s_j}\]

Under the assumption that the average neighbor has diploid copy number equal to the cohort mean, \(\widehat{\mathrm{dipCN}}_i\) is a relative estimate of copy number normalized to the cohort. When used with the LPA realignment step, this is computed separately for each exon classification (e.g. 1A, 1B-KIV2, 1B-KIV3); for other loci a single count per sample is sufficient.

Output: A TSV file per exon type with sample ID and normalized diploid copy number estimate.