Skip to content

Signal chunking

For ordinary inference, keep the default auto mode. This reference explains when to select another method and how boundaries affect model inputs.

Chunk-boundary selection

Data-consuming commands expose --chunking_mode auto|mv|hierarchical|dtw|inferred. Omission defaults to auto: use move-based decoding when mv exists, hierarchical anchors for DNA without moves, and mean6 DTW for RNA without moves. Auto does not hide errors in malformed present move tags.

Explicit mv requires move tags. Explicit hierarchical or dtw forces that method for either DNA or RNA, even when moves exist.

Alternatively, pass --chunking_config /path/to/chunking.json with:

{"chunking_mode": "dtw"}

This is a separate, run-wide data-processing JSON file, not a BAM manifest or model configuration. An explicit --chunking_mode overrides the JSON value. It applies to pretraining, supervised training, classifier inference, canonical-null workflows, prototype extraction/scoring and pairwise-comparison. Dataset/factory Python APIs accept the same modes, defaulting to chunking_mode="auto".

Both inferred methods support primary/unmapped DNA/RNA A/C/G/T reads with the appropriate 9-mer current model. They use positive integer ts as the signal start and ns as the exclusive end. A zero or absent ts triggers start estimation; an absent ns triggers end estimation. If both are missing, both endpoints are estimated. Character-valued ts:A alignment tags are ignored. Invalid numeric bounds raise an error. Secondary/supplementary records are excluded. Tags always describe acquisition-time coordinates; RNA signal is reversed after trimming to follow original read sequence order. This does not depend on the alignment strand. Other k-mer model sizes remain supported by mv only.

Both methods use six-sample non-overlapping means and quantile-based affine calibration for matching. The defaults are:

  • hierarchical: 192-base interior anchor queries, approximately 256-base anchor spacing, endpoint queries up to 384 bases, and piecewise interpolation. RNA search, duration and cost bounds scale with estimated signal dwell.
  • dtw: whole-read one-to-many DTW with a sloped half-band of 1,000 averaged bins (6,000 raw samples), squared-current error, rolling cost rows and byte traceback pointers. Known ts/ns endpoints bypass anchor searches. Missing endpoints are estimated by the hierarchy before DTW. The traceback is capped at 256 MiB per read; oversized or unreachable alignments raise an error, without silently switching methods. Hierarchical is available for very long reads that exceed this allocation limit.

The model receives original signal samples at its configured row width. Fewer than one row of terminal samples may be dropped to preserve that width. Inferred boundaries do not supply move tables or measured per-base dwell times.