amethyst-facet 1.2.4

Compute window aggregations and alter contents of Amethyst HDF5 files

Facet is an efficient utility for computing window aggregations on Amethyst HDF5 files produced via the premethyst pipeline.

Create environment

Install facet.py dependencies using mamba:

mamba create -n facet pip && mamba activate facet && pip install amethyst-facet

Ingest base-pair-resolution .parquet and .cov files

facet calls2h5 will ingest base-pair-resolution methylation observations in the Scale Bio .parquet format as well as the legacy plaintext .cov format to the HDF5 format used by Amethyst. This can then be used to compute window aggregations using facet agg. Context and barcode can be flexibly parsed from the filename. Run facet calls2h5 --help for examples.

Compute Window Aggregations

facet agg will add window aggregations to an existing HDF5 file in version 2.0.0 (see below for information on file format conversion).

Example:

facet agg -u 500 -u step_1000=1000:250 -w special_fancy_windows=windows.tsv -p 55 *.h5

This computes several types of windows.

• -u 500 computes uniform non-overlapping 500bp windows. These will be stored in /[context]/[barcode]/[window_size] by default. A custom name can be chosen by prepending -u [dataset_name]=500.
• -u step_1000=1000:250 computes 1000bp windows with a 250bp step, so intervals will be computed at $[0, 1000), [250, 1250), ...$. This example uses a custom name of step_1000. The default is to use [window_size]_by_[step_size], which in this case would have been 1000_by_250.
• -w special_fancy_windows=windows.tsv computes aggregations over custom windows defined in a CSV-like file. The headers chr, start and end are required but the file format is sniffed by DuckDB (csv, tsv etc are allowed). Intervals are left-closed right-open, i.e. $[start, end)$ and may be overlapping and gapped.

The -p 55 option parallelizes the computation using 55 worker cores. All HDF5 files retrieved via *.h5 will have windows computed in this case. Multiple globs can be specified, i.e. -glob path1/*.h5 -glob path2/*.h5.

Other options are described in facet agg --help.

Help

The options for facet.py can be explored at the command line by appending --help.

Example:

$ facet --help
Usage: facet.py [OPTIONS] COMMAND [ARGS]...

Options:
  --help  Show this message and exit.

Commands:
  agg      Compute window sums over methylation observations stored in...
  convert  Convert an old Amethyst HDF5 file format to v2.0.0 format
  delete   Delete contexts, barcodes, or datasets from an Amethyst 2.0.0...
  version

You can also call --help on subcommands. Example:

facet agg --help

Convert old Amethyst HDF5 file format to version 2.0.0

File format conversion is necessary prior to computing window aggregations using facet.py for Amethyst HDF5 files produced using earlier scripts.

Example:

facet convert old_format.h5 new_format.h5

Explanation and schema comparison:

The old Amethyst HDF5 format stored datasets under a cell barcode under a context group:

/[context]/[barcode]

context values are typically CH and CG. The barcode values are unique identifiers attributed to single cells. Typically each value of barcode is found in both the CH and CG contexts.

The schema of barcode was chr, pos, pct, c, t, with chr the chromosome name, pos the bp position of the observation, pct equal to c/(c+t), and c and t the methylated and unmethylated count at that position.

This gave no clear way to store window aggregations alongside the bp-resolution observations. We therefore altered the schema to:

/[context]/[barcode]/[dataset]

The bp-resolution observations are stored under the dataset 1 by default. Window aggregations are stored under their context and barcode under other names. The schema for window aggregations is chr, start, end, c, t, c_nz, t_nz. The start and end values denote the interval $[start, end)$. The c and t values store the sum of c and t counts for observed positions on that interval. Intervals with no observations are not reported. The c_nz and t_nz fields store the count of positions where c >= 1 or t >= 1 respectively.

Delete datasets

Examples:

facet delete context CH *.h5
facet delete barcode AGCGAGCGAGCAHHCAHH *.h5
facet delete dataset 1 *.h5