escpod classify¶

Read-level classification against a model bundle, from POD5 signal plus an aligned BAM. Today that means the tRNA charging (aminoacylation) classifier: for each read it asks whether the tRNA was charged, and writes the probability onto the BAM.

escpod classify reads.pod5 -b aln.bam -r ref.fa -m bundle/ -o out.bam

Ships in the default build — no extra Cargo feature.

Usage¶

escpod classify [OPTIONS] -b <BAM> -r <FASTA> -m <BUNDLE> -o <BAM> <POD5>

Arguments¶

Options¶

Output¶

The cl tag is written as uint8, round(P(charged) · 255), onto every record of each classified read — no modbase ML → cl round-trip. Reads that could not be scored carry no tag, and the run reports why each was skipped: unmapped/filtered, low mapq, reference without junction, missing mv/ns tags, junction not aligned, query outside move table, non-UUID read name.

--tsv additionally emits one row per read as read_id, reference, p_<positive class>, cl, reason — the convenient form for plotting or thresholding outside a BAM reader, and the place the skip reason is recorded per read rather than only in the summary.

Getting a model bundle¶

Bundles come from escapepod-models. A bundle is a directory holding the scorer plus a metadata.json describing the feature recipe.

The recipe travels in the bundle, never in flags — feature order and offsets, the k-mer table pinned by sha256, the operating point. That is deliberate: a caller computing the features differently gets a wrong answer rather than an error, so there is nothing to configure.

The metadata schema is closed. Every key in the file is a rule the model was built with, so a key this runtime does not implement is refused at load rather than silently dropped. (provenance, metrics and caveats are exempt and free-form.)

That cuts both ways, and it is worth knowing which escpod you are running: a bundle from a newer builder fails to load rather than loading with its new rule quietly ignored. Refusing to answer is the recoverable half of the trade, but it does mean a bundle and a runtime have to be released in that order.

Which basecaller called the corpus¶

A bundle may declare the basecalling model its training corpus was called with:

"basecaller": {"model": "rna004_sup@v6.0.0", "dorado_version": "2.1.1+d66c17c"}

escpod prints it at load and does not check it — it would have to read the BAM's @PG to know what called the reads it was handed. Checking it is worth the trouble: the charging feature set is mean + z-scored k-mer residual and the expected level is predicted from the read's own basecall, so a charging model substantially detects how the basecaller fails at the aminoacyl adduct. Scoring the same reads called with a different model loses ~0.010 AUROC, loses ~3 pp of TPR while gaining ~0.4 pp of FPR at the shipped set point (so no threshold recovers it), and flips 3.9% of per-read calls — while the aggregate charged fraction moves 0.04 pp. The one number anyone would check reads "no change" while one read in 26 answers differently.

Bundles published before this key exists do not carry it, and load fine; escpod says so rather than assuming comparability.

Two scorers, one feature space¶

A bundle carries either a gradient-boosted tree model (gbm, which routes NaN natively) or a small ONNX network over the same columns (feature_model). Which one a directory holds is a property of the bundle, never a flag — everything upstream of the final scoring step is shared verbatim. Both load in the default build.

The network is what escapepod-models ships, and it is the better model: 0.727 of reads callable at 99% precision, against the GBM's 0.449 on a held-out flowcell.

A third variant: the signal window¶

waveform_model reads a signal window rather than a column vector: the normalised current and its k-mer residual, the sequence k-mer context scattered along the signal axis, and 12 per-base dwell/level rows — three tensors, one BCE logit out.

It is a second pipeline over the same two files, not a second scorer on the first one. It also reads the reference differently: the sequence a read is scored against is rebuilt from that read's MD tag, the way the training corpus builds it, not sliced out of the FASTA. Those disagree wherever the FASTA carries an N — the alignment recorded a real base there — and since levels are looked up per 9-mer, one N blanks nine of them. A read with no MD tag is skipped (no MD tag) rather than silently scored off the FASTA.

The bundle says which source it was built from, in waveform_model.preprocessing.reference_source — absent meaning md, which is what every published bundle was built with. That is a different question from motif_reference, which names where the motif is searched (the FASTA, in reference coordinates, which is right); reading one as an answer to both is how this runtime got it wrong. A bundle naming a source escpod does not assemble is refused at load, naming both, rather than scored off MD anyway — the ambiguity is permanent (it is an ordered degenerate position, so no substituted base is right for more than 55% of reads), and both sources score every read without erroring.

The base-to-signal map is walked through the CIGAR into reference coordinates, the anchor is the motif +2 rather than +3, the spans are refined by a banded DP before any feature is taken, and the signal frame comes from the bundle instead of a vote (so --orientation is ignored, with a warning).

Two things follow from that:

• The bundle version decides whether it loads at all. It goes through the statically linked tract like every other ONNX graph escpod runs, so a stock release binary can run one — but only since charging_tcn_rna004@v0.1.1. The first export, @v0.1.0, defeated tract's shape inference two ways at once — a value_info entry per intermediate carrying the batch axis as a symbol, so pinning the batch failed at the first convolution; and adaptive_avg_pool1d(390 → 11) open-coded into a rank-8 GatherND — and no graph rewrite fixed either. The fix was the re-export (rnabioco/escapepod-models#96, and #97's build-time gate so a graph the shipped runtime cannot load never registers again). Loading @v0.1.0 fails at into_optimized with tract's own analysis error, naming the file.
• Its shipped Platt calibration is carried, not applied. The operating point beside it is stated on the uncalibrated probability the graph emits, so calibrating would move the scale out from under the very threshold that ships with the model. escpod says so at load; a caller who wants calibrated probabilities must re-derive the threshold too.

How a read is anchored¶

1. Locate the CCA–aa junction in reference coordinates (the CCAGGC motif, +3; the windowed variant anchors at +2, one base earlier, and says so in its own metadata).
2. Map reference → query through the CIGAR.
3. Map query → signal through the move table, Remora convention (move_pos * stride + ts).
4. Compute per-base dwell / mean / std plus the z-scored k-mer residual, with everything before the common arm masked.
5. Score.

--orientation¶

Some runs index reversed signal in their move tables. Getting this wrong silently mirrors every window, so escpod detects it per run by vote rather than assuming — but the vote needs ≥ 50 informative reads and a 95% consensus to commit.

Leave it at auto for a normal run. Use --orientation time or --orientation reversed to force the frame on a batch too small for the vote to resolve.

GPU acceleration¶

Only the waveform_model variant has a GPU path — the windowed TCN, batched through tract-cuda rather than onnxruntime. GBM and feature_model bundles have no device path at all; --device gpu on one of those is a no-op with a one-line log note, never an error. Build with --features gpu (see demux's GPU section for the Cargo/pixi setup — one gpu feature covers every GPU-capable stage in the binary, classify included).

tract-cuda needs more than the demux GPU stages do. detect --method cnn and the CRF encoder go through onnxruntime, which only dlopens a prebuilt libonnxruntime — the CUDA driver (libcuda.so, part of the NVIDIA driver, always present on a GPU node) is enough for that check. classify's TCN instead builds a CUDA context and compiles its own kernels at run time through NVRTC, and needs two more things the driver alone does not supply:

• The CUDA runtime library, libcudart, discoverable by name at dlopen time — not just the driver. A node can have a perfectly good driver and still fail here if the libcudart it finds is the wrong major version (see rnabioco/escapepod-rs#347's post-mortem below).
• CUDA headers (cuda_fp16.h, the CCCL headers) on disk at run time, for NVRTC to compile against — nothing else in this binary needs headers past build time.

The pixi gpu environment (pixi run install-gpu, then pixi run -e gpu … — see demux's runtime-libraries section) already pins both: cuda-cudart/cuda-cudart-dev for the runtime library and cuda-cccl for the headers. Run escpod classify --device gpu the same way you would run any other GPU-capable command — inside pixi run -e gpu, or with that environment's LD_LIBRARY_PATH reproduced by hand. Running the released -gpu binary bare, with no CUDA environment set up by the caller, means libcudart resolution falls through to whatever the node happens to expose on its default library path — see the warning below.

Notes¶

• A bundle's abstain rule is carried and warned about, not applied — if your bundle declares one, the calls here are unabstained and you should apply the rule downstream.
• Parity with the training-corpus implementation (escapepod_models.charging) is pinned by golden-vector tests, and the ONNX arm additionally by bit-exact reference-vector parity plus a real-weights run over a real corpus.
• BAM file I/O lives in the CLI; the classifier itself is the escapepod-classify crate, which owns every definition of the model's input.