CNV frequencies halved in gene_cnv_frequencies_advanced with nobs_mode="fixed"

[rehanxt5]

Summary

When gene_cnv_frequencies_advanced is called with nobs_mode="fixed", all returned CNV frequencies are exactly half the correct value. The nobs denominator multiplies by 2 when it shouldn't.

The Problem

In cnv_frq.py lines 587-588:

if nobs_mode == "called":
    nobs[:, cohort_index] = np.repeat(cohort_n_called, 2)
else:
    assert nobs_mode == "fixed"
    nobs[:, cohort_index] = cohort.size * 2  # ← BUG: should not multiply by 2

The issue: count represents number of samples with amp/del (line 580-581):

count[::2, cohort_index] = np.sum(cohort_is_amp, axis=1)   # sample count
count[1::2, cohort_index] = np.sum(cohort_is_del, axis=1)  # sample count

So frequency = count / nobs becomes sample_count / (samples × 2) — exactly half.

Why this is wrong

1. The "called" mode proves the bug. It uses:

   nobs[:, cohort_index] = np.repeat(cohort_n_called, 2)

   np.repeat(x, 2) interleaves each value twice — it produces [10, 10, 20, 20, 30, 30], NOT [20, 40, 60]. So nobs = number of called samples, not doubled.

2. The basic version confirms it. gene_cnv_frequencies (non-advanced) computes:

   frequency = amp_count_coh / called_count_coh  # sample count / sample count ✅
   nobs = called_count_coh                        # NO multiply by 2

3. CNV calls are per-sample, not per-allele. Unlike SNPs where each diploid sample contributes 2 alleles, CNV CN_mode gives one copy number classification per sample.

Impact

• All CNV frequencies from gene_cnv_frequencies_advanced(..., nobs_mode="fixed") are exactly half the true value
• Downstream confidence intervals from _add_frequency_ci are artificially narrow (inflated nobs inflates precision)
• Any spatial/temporal/population analyses using these frequencies show systematically deflated CNV rates

Example

10 samples with amp, nobs_mode="fixed":

• Current (wrong): frequency = 10 / (10 × 2) = 0.50 → suggests 50% frequency
• Correct: frequency = 10 / 10 = 1.0 → suggests 100% frequency (all samples have it)