pyrsx

Python bindings for rsx: a high-performance streaming toolkit for RAD-seq sex determination.

Installation

pip install pyrsx

Usage

import pyrsx

# Process FASTQ files into marker depth table
pyrsx.process("reads/", "markers.tsv", threads=4, min_depth=5)

# Compute distribution with Fisher's exact test + FDR
pyrsx.distrib("markers.tsv", "popmap.tsv", "distrib.tsv",
              test="fisher", correction="fdr")

# Extract significant markers with Bayesian output
pyrsx.signif("markers.tsv", "popmap.tsv", "signif.tsv",
             test="fisher", correction="fdr", bayes=True)

# Streaming PCA
pyrsx.pca("markers.tsv", "pca_results/", n_components=10)

# Merge tables (bounded memory, handles 75M+ sequences)
pyrsx.merge(["table1.tsv", "table2.tsv"], "merged.tsv")

Features

• All rsx commands accessible from Python
• 3.14x geometric-mean speedup on the tracked Slurm literature comparison panel
• Bounded-memory streaming for arbitrarily large datasets
• Multiple statistical tests: chi-squared, Fisher's exact, G-test
• Multiple corrections: Bonferroni, Benjamini-Hochberg FDR
• Bayesian sex-linkage classification (Bayes Factor + posterior)
• Streaming PCA via Tucker mode-2 decomposition
• K-mer based marker deduplication

High-level API & backend agnosticism (recommended)

The low-level functions above are thin wrappers. For most users the MarkerTable + result objects (in pyrsx.api) are the idiomatic entry point:

import pyrsx as rsx

table = rsx.MarkerTable.from_path("markers.tsv")   # or from_dataframe(...)
result = table.triage(popmap="popmap.tsv", min_depth=10)

# Everything is a narwhals DataFrame under the hood → backend agnostic
print(result.df)                    # stays in whatever backend you prefer
df = result.to_polars()             # or .to_pandas(), to_dataframe(backend=...)

How outputs are read (no forced pandas fallback): Internal TSVs produced by rsx core commands are read with pyarrow.csv (handling the leading #Number of markers comment via skip_rows=1) and then wrapped with to_narwhals(...). The exposed objects are always narwhals DataFrames (concrete backend = pyarrow by default for efficiency). You only pull in pandas/polars if you ask for that backend later. This is the standard narwhals approach used throughout the high-level API (see _adapters.py, _read_core_tsv, and the detailed docs in the Rust extension).

See the docstrings of MarkerTable, the various *Result classes, and _read_core_tsv for the full rationale.