vdjtools

Load a bundled model with vdjtools.model.load_bundled(locus) and call pgen_aa(model, sequence), optionally pinning V/J calls or allowing one mismatch. For many sequences, pgen_aa_batch runs thread-parallel with bitwise-identical results to the serial version.

Use vdjtools.io.read with fmt="auto" to sniff the format, or call read_mixcr, read_immunoseq, read_imgt, or read_trust4 directly. Every reader emits the same canonical clonotype frame with junction_nt and junction_aa columns following the AIRR junction convention.

Yes, model.from_germline builds a model from any germline DataFrame with allele, segment, and sequence columns, and reference.read_germline_fasta constructs that frame from your own FASTA files. reference.validate_germline audits anchor frames and naming before training.

Yes, vdjtools.sc.read_airr_cell reads CellRanger airr_rearrangement.tsv files and read_10x handles contig annotation CSVs across CellRanger versions 3 through 7. It also exports to scirpy, dandelion, and scRepertoire formats for downstream single-cell analysis.

Before version 3.3.0, collapse_alleles used an unordered polars group_by, so the collapsed table row order varied per process and the same seed drew different alleles. Seeds are reproducible across processes only from version 3.3.0 onward.

Amino-acid Pgen is a relative score only and is not normalized, so model_fit uses nucleotide Pgen for likelihood, AIC, and BIC. Also, the character X means wildcard only in pgen_aa_degenerate; in pgen_aa it matches literally and silently returns 0.0.