fastatacular

fastatacular is a pure-Python library with no dependencies for reading and writing FASTA sequence files. Each record becomes a frozen SequenceEntry whose header is parsed into typed fields: the identifier, the database prefix, accession and entry name of pipe-delimited identifiers (sp|P12345|EX_HUMAN, gi|12345|ref|NP_000001.1|), the UniProt description keys OS=, OX=, GN=, PE= and SV=, and the accession and description fields of NCBI, Ensembl and UniRef headers (NP_000549.1, ENSP00000322421.5, UniRef90_P69905). The library also reads compressed files, writes headers back unchanged, builds target-decoy databases, gives random access by identifier or accession through a samtools-compatible .fai index. It is intended for proteomics and genomics pipelines and requires Python 3.12 or later.

Documentation

• This README: installation, quick start and the reference for reading, writing, header parsing and indexing.
• docs/decoys.md: decoy methods, options, the Markov models and decoy quality on the human proteome.
• llms-full.txt: the complete usage guide, with every public name, signature and runnable example.
• CHANGELOG.md: release history.
• FASTA Toolkit: a web app built on fastatacular that makes decoys, checks decoy quality, cleans up, merges and converts FASTA files in the browser (files are not uploaded). To browse entries, use the FASTA Viewer.

Install

pip install fastatacular

Quick start

Write a FASTA file, then read it back:

from fastatacular import SequenceEntry, read_fasta, write_fasta

write_fasta(
    [
        SequenceEntry(
            identifier="sp|P12345|EX_HUMAN",
            sequence="MKTIIALSYIFCLVFA",
            pname="Example protein",
            os_name="Homo sapiens",
            ncbi_tax_id=9606,
            gname="EXMP",
            pe=1,
            sv=2,
        ),
        SequenceEntry(identifier="seq2", sequence="MPEPTIDEK"),
    ],
    "proteins.fasta",
)

entries = read_fasta("proteins.fasta")
for entry in entries:
    print(entry.identifier, entry.accession, len(entry.sequence))
# sp|P12345|EX_HUMAN P12345 16
# seq2 None 9

read_fasta returns a list. FastaReader yields one entry at a time without loading the whole file:

from fastatacular import FastaReader

with FastaReader("proteins.fasta") as reader:
    for entry in reader:
        print(entry.identifier, entry.sequence)

A str argument is always a file path. To parse FASTA text, wrap it in io.StringIO:

import io

from fastatacular import read_fasta

entries = read_fasta(io.StringIO(">a\nMKV\n>b\n  >c\nPE\n"))
[(e.identifier, e.sequence) for e in entries]
# [('a', 'MKV'), ('b', ''), ('c', 'PE')]

As the example shows, a header followed directly by another header (or by the end of the file) gives an entry with an empty sequence, and a header line with leading whitespace is still read as a header. write_fasta writes an entry with an empty sequence as its header line alone, so it reads back identically.

Compressed files are read transparently: gzip, bzip2 and xz, detected from the magic bytes, not the file name. Pipes, FIFOs and /dev/stdin work too:

entries = read_fasta("uniprot_sprot.fasta.gz")

Every reader and writer takes a pandas-style compression= keyword (type alias fastatacular.Compression): "infer" (default), "gzip", "bz2", "xz" or None. On read, "infer" sniffs the magic bytes, an explicit format is forced (input in another format raises FastaParseError) and None reads plain text. A binary handle is read as plain UTF-8 unless you pass an explicit format:

entries = read_fasta("download.bin", compression="gzip")
with open("proteins.fasta.xz", "rb") as fh:
    entries = read_fasta(fh, compression="xz")

A PEFF file also reads as plain FASTA: its # file header lines are skipped and each entry keeps its identifier, sequence and raw description. Use pefftacular to parse the PEFF annotations.

Data model

Each entry is a SequenceEntry:

UniProt-style headers

from fastatacular import read_fasta

[entry] = read_fasta("one.fasta")
# >sp|P12345|EX_HUMAN Example protein OS=Homo sapiens OX=9606 GN=EXMP PE=1 SV=2

entry.prefix         # "sp"
entry.accession      # "P12345"
entry.entry_name     # "EX_HUMAN"
entry.pname          # "Example protein"
entry.os_name        # "Homo sapiens"
entry.ncbi_tax_id    # 9606
entry.gname          # "EXMP"
entry.pe             # 1
entry.sv             # 2

Other KEY=value pairs are kept in entry.extra. A header without KEY=value pairs fills description and pname and leaves extra empty. When a key appears more than once (OS=Human ... OS=Mouse), the first value is used; description and raw_header still hold the full text.

NCBI, Ensembl and UniRef headers

Identifiers that are a bare accession are recognized by their shape, and the fields their database writes in the description are read into the same typed fields:

[ncbi] = read_fasta("refseq.fasta")
# >NP_000549.1 hemoglobin subunit alpha [Homo sapiens]
ncbi.accession       # "NP_000549.1"
ncbi.os_name         # "Homo sapiens" (the defline's trailing brackets)

[ensembl] = read_fasta("ensembl.pep.fasta")
# >ENSP00000322421.5 pep chromosome:GRCh38:16:176680:177522:1 gene:ENSG00000206172.8 \
#  transcript:ENST00000320868.9 gene_biotype:protein_coding gene_symbol:HBA1
ensembl.accession    # "ENSP00000322421.5"
ensembl.gname        # "HBA1" (from gene_symbol:)

[uniref] = read_fasta("uniref90.fasta")
# >UniRef90_P69905 Hemoglobin subunit alpha n=12 Tax=Primates TaxID=9443 RepID=HBA_HUMAN
uniref.accession     # "UniRef90_P69905"
uniref.os_name       # "Primates" (from Tax=)
uniref.ncbi_tax_id   # 9443 (from TaxID=)
uniref.extra         # {"n": "12", "RepID": "HBA_HUMAN"}

Only what each format defines is parsed, and an explicit UniProt key always wins:

• NCBI: RefSeq protein (NP_, XP_, YP_, WP_, AP_, ZP_) and versioned GenPept (AAA59174.1) accessions. The organism is the defline's trailing [...] only, so an NCBI [key=value] qualifier is left as text, and the brackets stay in pname as well. A defline carries no taxonomy id, so ncbi_tax_id stays None.
• Ensembl: protein ids of an Ensembl pep file (ENSP..., ENSMUSP...). Only gene_symbol: is a gene name; gene: and transcript: are Ensembl ids, which no field takes, and Ensembl pep headers name no organism.
• UniRef: UniRef50_, UniRef90_ and UniRef100_ clusters. The cluster size (n=) and the representative member (RepID=) stay in extra.
• A decoy or contaminant tag in front of the identifier keeps the target's accession (DECOY_NP_000549.1), as it does for pipe-delimited identifiers.
• Any other first word is an identifier only: accession stays None and no description field is guessed from it.

Writing

Construct entries and write them out:

from fastatacular import SequenceEntry, write_fasta

entries = [
    SequenceEntry(
        identifier="sp|P12345|EX_HUMAN",
        sequence="MKTIIALSYIFCLVFA",
        pname="Example protein",
        os_name="Homo sapiens",
        ncbi_tax_id=9606,
        gname="EXMP",
        pe=1,
        sv=2,
    ),
]

write_fasta(entries, "output.fasta")

dest accepts a path string, a pathlib.Path, or a text or binary file object (a binary one gets UTF-8 bytes; a handle is never closed). With the default compression="infer", a path ending in .gz, .bz2 or .xz (any case) is written gzip, bzip2 or xz compressed and anything else, including a handle, is plain text. compression="gzip", "bz2" or "xz" forces that format whatever the suffix (a handle must then be binary), and compression=None always writes plain text. gzip output has mtime 0 and no file name in its header, so the same entries give the same bytes.

write_fasta(entries, "output.fa", compression="xz")
with open("output.fasta.gz", "wb") as fh:
    write_fasta(entries, fh, compression="gzip")

Sequence lines wrap at 60 characters by default. Override with line_width= (pass 0 to disable wrapping):

write_fasta(entries, "output.fasta", line_width=80)
write_fasta(entries, "single-line.fasta", line_width=0)

If raw_header is set on an entry (as it is on every entry produced by read_fasta) and still matches the entry's structured fields, the writer round-trips it verbatim. If you changed a field (for example dataclasses.replace(entry, gname="XYZ")), or raw_header is empty, the header is rebuilt from the structured fields, so your edit is written.

Decoy databases

Build target-decoy databases for FDR estimation with reverse, pseudo_reverse, shuffle, debruijn (repeat-preserving, Moosa et al. 2020) or markov (order-2 chains shipped for human, mouse, yeast and E. coli, trained on UniProt, CC BY 4.0). Every method takes keep_residues (e.g. "KR"), keep_nterm and keep_cterm, and the same seed always gives the same decoys. Pure Python, no extra dependencies.

from fastatacular import is_decoy, make_decoys, read_fasta, write_decoy_fasta

write_decoy_fasta("human.fasta", "human_td.fasta", method="pseudo_reverse", seed=1)

decoys = make_decoys(read_fasta("human.fasta"), method="markov", model="human", keep_residues="KR", seed=1)
next(decoys).identifier   # "DECOY_sp|..."; is_decoy(entry) checks the prefix

See docs/decoys.md for every option, the Markov model data and per-method quality numbers on the human proteome.

Random access by identifier or accession

FastaIndex(path) reads the file once and keeps only each entry's key and byte range; index[key] then reads and parses just that entry.

from fastatacular import FastaIndex

index = FastaIndex("human.fasta")
entry = index["sp|P31946|1433B_HUMAN"]   # SequenceEntry, read from disk on demand
"sp|P31946|1433B_HUMAN" in index, len(index)   # no file access
index.write_fai()                        # human.fasta.fai, samtools-compatible
index = FastaIndex.from_fai("human.fasta")   # later: load the .fai instead of scanning

by_acc = FastaIndex("human.fasta", key="accession")
by_acc["P31946"]

• By default keys are full identifiers (the first header word, the .fai name, as in samtools). This works on target-decoy databases, where sp|P1|X and DECOY_sp|P1|X or Reverse_sp|P1|X share an accession.
• key="accession" keys entries by accession (P31946 for sp|P31946|1433B_HUMAN, or the whole identifier when it has no |). It needs unique accessions.
• A repeated key raises FastaError naming the first one. Real databases do repeat identifiers (IP2 exports, merged databases): FastaIndex(path, duplicates="first") keeps the first entry for each key and skips the rest, as samtools faidx does, and logs one warning with the number skipped. from_fai takes the same option.
• A missing key raises FastaKeyError, which is also a KeyError.
• The file must be a regular, uncompressed file. A FIFO, pipe or directory raises FastaError: save the input to a file first, or read it once with FastaReader. gzip (including bgzip), bzip2 and xz raise FastaError: decompress first (gunzip -k human.fasta.gz). bgzip/.gzi is not supported. compression="gzip"/"bz2"/"xz" raises FastaError; compression=None skips the magic-byte check.
• .fai caveats: the name column is the first header word (mapped to the accession on load with key="accession"). Like samtools, write_fai() needs every sequence line of an entry but the last to have the same length, and no comment or blank lines inside a sequence; otherwise it raises FastaError (rewrite the file with write_fasta first). from_fai checks each entry's header against the .fai name but does not re-count residues, so rebuild the .fai whenever the FASTA changes. An entry with no sequence gets a row with length 0. samtools and pysam refuse a FASTA file that starts with a byte-order mark, so the .fai of such a file works only with FastaIndex.

Error handling

Parse errors raise FastaParseError:

from fastatacular import FastaParseError, read_fasta

try:
    entries = read_fasta("malformed.fasta")
except FastaParseError as e:
    print(e.line)     # offending line number
    print(e.context)  # surrounding line content

Input that is not UTF-8, or a corrupt compressed file, also raises FastaParseError (chained to the underlying UnicodeDecodeError or OSError).

Write errors raise FastaWriteError, whose index names the bad entry. Every entry is validated before anything is written, so a failed write_fasta leaves no partial file. Both errors subclass FastaError (a ValueError), so except FastaError catches either.

SequenceEntry is frozen but not hashable (its extra field is a dict), so key sets and dicts by entry.identifier, not by the entry. A FastaReader is single-pass: iterate it once, or open a new one to read the file again.

Development

just install      # install dependencies
just test         # run tests
just test-v       # run tests (verbose)
just cov          # run tests with coverage
just lint         # ruff lint
just format       # ruff format
just check        # lint + type check + test
just build        # build the package
just clean        # remove cache files

Citation

If you use fastatacular in research, please cite the archived software release. Machine-readable citation metadata is available in CITATION.cff; GitHub's Cite this repository menu can render it as APA or BibTeX. DOI: 10.5281/zenodo.22926358.

License

MIT