Chemistry and signal preprocessing

Checkpoint-based commands infer sequence type, pore chemistry, and expected k-mer current from the checkpoint. Do not supply --seq_type, --pore_type, or --kmer_dict_path in ordinary inference, fine-tuning, or embedding commands. If supplied, these options assert compatibility and fail on disagreement. Pretraining from scratch requires --chemistry or an explicit custom profile.

Built-in profiles

Chemistry	Sequence	Expected-current table
dna_r10.4.1_e8.2_400bps	DNA	dna_r10.4.1_e8.2_400bps/9mer_levels_v1.txt
rna004	RNA	rna004/9mer_levels_v1.txt
rna002	RNA	rna_r9.4_180mv_70bps/5mer_levels_v1.txt

These files ship with PoreFM, pinned to nanoporetech/kmer_models commit 4e56daed7fbb79b538f58e41262d5c54b07356ea. Runtime use needs neither a separate checkout nor network access. Package metadata records upstream paths, file SHA-256 hashes, and hashes of the parsed float32 lookup values. Upstream license and copyright notices are included.

Identifiers and pore aliases are case-insensitive. RNA002 implies RNA; R9 plus seq_type="rna" also selects RNA002. Bare R9 is ambiguous and does not identify a table. Bare rna cannot distinguish RNA002 from RNA004. R10 plus DNA selects the registered 400-bps DNA profile.

RNA002 uses move-table decoding. Signal-based chunk inference remains restricted to 9-mer tables: provide BAM move tags and use mv, or auto with moves, for RNA002. Registering a chemistry does not make a pretrained backbone compatible with different signal widths or move strides.

Python readers and datasets

Custom readers can use these exports without adopting PoreFM's dataset classes:

from porefm.data.bam import resolve_preprocessing, build_lookup_table, get_read_info_lite

prep = resolve_preprocessing(chemistry="rna002")
table = build_lookup_table(chemistry="rna002")  # Cached, read-only float32 array.
read_fields = get_read_info_lite(bam_record, preprocessing=prep)

Alternatively resolve from a trusted checkpoint without constructing a model:

from porefm.preprocessing import load_preprocessing
prep = load_preprocessing("model.pt")

All signal iterable datasets accept preprocessing=prep or one chemistry="rna004" argument. Model loading also exposes the resolved contract:

from porefm.models import load_model
from porefm.data.reader import BamIterableDataset

model = load_model("model.pt", model_type="backbone")
dataset = BamIterableDataset(
    bam_list=bam_list, size=2048, step=1792, move_downsample_factor=1,
    preprocessing=model.preprocessing,
)

Resolve once before the read loop. Collators that batch decoded samples need no table or chemistry arguments. Built-in tables retain upstream sequence orientation; BAM decoding reverses RNA signal into read orientation.

Custom profiles and checkpoints

For an experiment with a custom table, provide its sequence type explicitly:

prep = resolve_preprocessing(
    chemistry="custom", seq_type="rna", pore_type="experimental",
    kmer_dict_path="my_levels.txt",
)

The headerless text format is KMER LEVEL, using a consistent odd k and ACGT encoding and finite current levels. build_lookup_table(path) returns a writable array; profile arrays are immutable. Custom checkpoints embed their exact table, so subsequent users do not need the original file.

When saving a custom model, include its resolved preprocessing profile and table so it can process signal on another machine. PoreFM checkpoint helpers use model.preprocessing. For a custom checkpoint writer, use attach_preprocessing(payload, prep) from porefm.preprocessing.