Skip to content

Copy Calls Were Incorrectly Harmonized #506

Closed

Description

@jjacobson95

I was trying to clarify the definitions of the copy numbers for the paper and was looking into the distributions / data quality when I found some discrepancies which led to a deep dive and a discovered a pretty widespread issue.

Current: current 2.4 data
Harmonized: Modified the 2.4 data with the changes listed in New Conversion below. I'll apply these changes and do a fresh build and replace the current 2.4 (but keep it in place).

Image

copy_number is read in on different scales across datasets, but copy_call is assigned in each by a uniform set of cutoffs (0.521 / 0.731 / 1.214 / 1.422). The cutoffs are on DepMap's log2(relative CN + 1) scale, where 0, 1, 2, 3 and 4 copies are 0, 0.585, 1.0, 1.32 and 1.585.

In the table below, x is DepMap's raw value, s is a raw log2 ratio (0 = normal, −1 = one copy lost), and c is a raw absolute copy count (2 = normal).

Dataset(s) Raw input Current conversion New conversion
DepMap (ccle, ctrpv2, fimm, gcsi, nci60, prism) log2(rel+1) 2^x / 2 none
Sanger (gdscv1, gdscv2) log2 ratio 2^s log2(2^s + 1)
mpnst log2 ratio 2^s log2(2^s + 1)
cptac log2 ratio none (2^s in call) log2(2^s + 1)
pancreatic log2 ratio none, own cutoffs log2(2^s + 1)
liver log2 ratio 2·2^s log2(2^s + 1)
bladder, colorectal seg.mean (log2 ratio) 2·2^(s/2) log2(2^s + 1)
novartis absolute copies log2(c) log2(c/2 + 1)
hcmi absolute copies (ASCAT) log2(c) log2(c/ploidy + 1)

Simulated on the 2.4 build, every dataset has a median of about 1.0 after conversion, with 71–95% of genes called diploid.

Notes

  • DepMap: the 26Q1 source is already on DepMap's scale, so the 2^x / 2 step isn't needed.
  • bladder/colorectal: inputs are log2 ratios, so the /2 should be dropped.
  • hcmi: divide by per-sample tumor_ploidy (GDC ASCAT metadata) to match the ratio sources, which are relative to each sample's own baseline.

Package Bug too: coderdata/dataset/dataset.py recomputes copy_call with pd.cut using bin edges [0, …, 2] rather than using the stored column. Values outside that range get no label and come back as NaN. Once harmonized, that means genes above 2 (more than about 6 copies, the strongest amplifications) are dropped. Switching the outer edges to 0 and inf. Adding @ymahlich for visibility.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Type

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions