Skip to main content

Genomic Smoothing

For some commands, like cfdna midpoints, you may want all genomic regions to contribute approximately the same "mass" to the features, for example to reduce the effect of copy number alterations.

cfdna comes with two scaling weight calculation commands:

  • cfdna fragment-count-weights
  • cfdna coverage-weights

Conceptually, the main difference is that coverage weights longer fragments higher than shorter fragments just due to them covering more positions. When that is not specifically beneficial, we recommend using cfdna fragment-count-weights where all fragments contribute the same mass independently of length.

Technically, the only difference in their calculations is that where cfdna coverage-weights counts each fragment as 1.0 in all covered positions (before the optional GC-bias correction). cfdna fragment-count-weights instead counts 1.0 / num_countable_bases (usually just the fragment length) in the covered positions. When counting in large windows, this approximates fragment counts very closely (the few fragments not fully contained in their windows are weighted by their overlap).

When applied in downstream feature extractions, each fragment contribution is multiplied by the scaling factor of its genomic scaling bin.

Step 1. Build per-sample scaling factors​

cfdna fragment-count-weights --help
cfdna coverage-weights --help

cfdna fragment-count-weights \
  --bam <sample>.bam \
  --output-dir <sample_directory>/count_weights \
  --output-prefix <sample_id> \
  --n-threads 12 \
  --blacklist <path>/hg38-blacklist.v2.bed \
  --blacklist <path>/<another_blacklist>.bed

cfdna coverage-weights \
  --bam <sample>.bam \
  --output-dir <sample_directory>/coverage_weights \
  --output-prefix <sample_id> \
  --n-threads 12 \
  --blacklist <path>/hg38-blacklist.v2.bed \
  --blacklist <path>/<another_blacklist>.bed

Tip: use the same blacklist set you use in your other per-sample commands.

Step 2. Apply smoothing in feature extraction commands​

For midpoints, where the signal is specifically fragment counts, not a length-weighted coverage, we prefer the fragment count-based weights:

cfdna midpoints \
  --bam <sample>.bam \
  ... \
  --scaling-factors <sample_directory>/count_weights/<sample_id>.scaling_factors.tsv

The same --scaling-factors input pattern works for fcoverage and lengths.

Combine with GC-bias correction​

Genomic smoothing and GC-bias correction can be used together.

The only requirement is that they are used together consistently, meaning that the GC-bias correction must be used when calculating the scaling factors:


cfdna fragment-count-weights \
  --bam <sample>.bam \
  --output-dir <sample_directory>/gc_corrected_count_weights \
  --output-prefix <sample_id> \
  --n-threads 12 \
  --blacklist <path>/hg38-blacklist.v2.bed \
  --blacklist <path>/<another_blacklist>.bed \
  --gc-file <sample_directory>/gc_bias/gc_bias_correction.npz \
  --ref-2bit <path>/hg38.2bit

cfdna midpoints \
  --bam <sample>.bam \
  ... \
  --gc-file <sample_directory>/gc_bias/gc_bias_correction.npz \
  --ref-2bit <path>/hg38.2bit \
  --scaling-factors <sample_directory>/gc_corrected_count_weights/<sample_id>.scaling_factors.tsv