ep-touche 0.1.1

Tools for preprocessing and analyzing high-resolution chromatin contact data.

touche

Python API and CLI tools for analyzing enhancer-promoter contacts (touches) from high-resolution chromatin contact data, refactored from the Danko Lab E-P_contacts reference workflows.

touche starts from processed pairs files. Raw FASTQ processing, alignment, deduplication, and cooler generation should be handled by an external workflow such as distiller-nf.

Status

The current implementation includes CLI tools+API for:

• Micro-C pairs conversion, filtering, QC, and chromosome-sharded NPZ caches
• local-decay contact calling, pair-type assignment, and plotting
• Aggregated peak analysis (APA) aggregation and inter-sample APA comparison
• enhancer/promoter local-background counting and treatment comparison
• pipeline run wrappers that preserve intermediate outputs and write JSON manifests

Installation

Install with pip from PyPI (package name is ep-touche):

pip install ep-touche
touche --help

uv can also install and run the package:

uv add ep-touche
uv run touche --help

local-decay also supports a statsmodels-backed LOWESS path (lowess_backend="statsmodels") for exact reference comparisons (at the cost of being much slower than our custom implementation). Install the optional legacy extra only if you need that backend:

pip install ep-touche[legacy]
# or: uv sync --extra legacy

Older CPUs and Rosetta

Polars publishes an lts-cpu wheel for older CPUs and for x86-64 Python on Apple Silicon under Rosetta. Because polars-lts-cpu provides the same import polars module but is a different Python distribution, it does not automatically satisfy ep-touche's normal polars>=1.0 dependency. If the standard Polars wheel does not run on your machine, install ep-touche, then replace Polars in that environment:

pip install ep-touche
pip uninstall polars
pip install "polars-lts-cpu>=1.0"

CLI Overview

touche preprocess --help
touche local-decay --help
touche apa --help
touche background --help

Available command groups:

• touche preprocess: convert/filter pairs, write QC summaries, and build NPZ caches.
• touche local-decay: call observed/expected contacts, assign pair types, plot distributions, or run the full local-decay workflow.
• touche apa: aggregate APA matrices, compare treatment/control APAs, or run a paired APA workflow.
• touche background: count EP/background contacts, compare treatment ratios, or run the full EP/background workflow.

See the CLI reference for examples, common options, and expected outputs.

Typical workflow

Start from analysis-ready .pairs or .pairs.gz files produced by distiller-nf or an equivalent workflow.

1. Filter or convert pairs with touche preprocess.
2. For real or repeated local-decay runs, build a position-only cache with touche preprocess build-cache --no-metadata.
3. Use the run wrappers for end-to-end analyses: touche local-decay run, touche apa run, and touche background run.
4. Use individual subcommands such as local-decay call or background count when debugging or replacing one stage.

Many of the computation-heavy steps use numba acceleration and parallelism. Set NUMBA_NUM_THREADS before running to control CPU usage (default is to use all threads on machine):

NUMBA_NUM_THREADS=8 touche background run ...

Python API

For notebooks and custom scripts, import the provisional API surface:

import touche.api as tt

indexes = tt.build_contact_indexes("sample.nodups_30_intra.pairs.gz", source="touche")

The API is organized around reading pairs and anchors once, running in-memory compute functions such as compute_apa, compute_local_decay, and compute_ep_and_background, then displaying or saving returned Matplotlib figures as needed. Long-running CLI and API calls support optional progress bars and lightweight profiling. See the API guide for examples.

Documentation

Detailed usage notes live under docs/:

• Docs index: human-facing guides and documentation conventions.
• CLI reference: command groups, common options, examples, outputs, and run-wrapper manifests.
• Micro-C preprocessing: distiller-nf boundary, pairs format expectations, filtering, QC, and cache building.
• API: provisional in-memory APIs for notebooks, interactive analyses, and custom scripts.
• Reproducing reference plots: end-to-end commands for the reference local-decay, APA, and EP/background plots.
• Testing and publishing: CI, local checks, and PyPI release workflow for the ep-touche distribution.

Implementation plans, experiment logs, and agent-facing notes belong in notes/.

Development

Development is managed with uv.

git clone https://github.com/adamyhe/touche.git
cd touche/
uv sync --dev
uv run touche --help
uv run pytest