fastatacular

A small, dependency-free library for reading and writing FASTA sequence files in Python. It's built for proteomics and genomics pipelines that need fast, predictable FASTA parsing without pulling in a bioinformatics megapackage.

It understands UniProt-style description keys (OS=, OX=, GN=, PE=, SV=) and pipe-delimited identifiers (sp|P12345|EX_HUMAN, gi|12345|ref|NP_000001.1|) out of the box, so you get structured fields instead of a header string to parse yourself.

Highlights

• Zero dependencies — pure Python, nothing else to install.
• Two ways to read — read_fasta for the whole file at once, FastaReader to stream entries lazily without loading everything into memory.
• Compressed input — .gz, .bz2 and .xz files are read directly (detected from the file's magic bytes).
• UniProt headers parsed for you — accession, organism, gene name, protein existence, and sequence version come back as typed fields, not a string you have to split yourself.
• Round-trip safe — entries produced by read_fasta write back out byte-for-byte compatible headers.
• Actionable parse errors — FastaParseError reports the offending line number and surrounding context.
• Shares its API shape with pefftacular, the PEFF (PSI Extended FASTA) sibling library, so switching formats doesn't mean relearning the interface.

Install

pip install fastatacular

Dev install:

just install

Quick start

read_fasta — load everything into memory at once:

from fastatacular import read_fasta

entries = read_fasta("proteins.fasta")
for entry in entries:
    print(entry.identifier, len(entry.sequence))

FastaReader — iterate lazily without loading the full file:

from fastatacular import FastaReader

with FastaReader("proteins.fasta") as reader:
    for entry in reader:
        process(entry)

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")

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

Non-standard KEY=value pairs are captured in entry.extra. Headers with no KEY=value tokens leave description and pname populated and extra empty.

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-mode file object.

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.
• .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. Entries with no sequence raise (samtools skips them). 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.

Tables with pandas or polars

to_records(source) returns one plain dict per entry, so any data-frame library can take the result directly. fastatacular does not ship or require pandas or polars; install whichever you use. (The test suite runs these examples only when the library is installed.)

import pandas as pd
import polars as pl

from fastatacular import to_records

records = to_records("human.fasta")   # a path, an open text handle, or entries
df = pd.DataFrame(records)
human = pl.DataFrame(records).filter(pl.col("ncbi_tax_id") == 9606)

FastaReader.to_records() and SequenceEntry.to_record() give the same dicts. Every record has these keys, in this order (fastatacular.RECORD_KEYS):

The same fields in pefftacular records have other names where each package follows its own model. Rename these to put both in one frame:

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