pymisha 0.4.0

Python interface for misha genomic databases with C++ streaming backends

PyMisha Python interface for misha genomic databases. PyMisha provides full read/write access to misha track databases with C++ streaming backends for genome-scale operations.

Features

• 1D and 2D track support: Dense, sparse, and 2D (rectangle/point) tracks with full CRUD operations.
• C++ streaming backends: Extraction, summary, quantiles, distribution, lookup, segmentation, Wilcoxon tests, correlation, and sampling all stream through C++ for performance.
• Virtual tracks: Computed-on-the-fly track views with filtering, shifting, and 30+ aggregation functions.
• Interval operations: Union, intersection, difference, canonicalization, neighbors, annotation, normalization, random generation, and liftover.
• Sequence analysis: Extraction, k-mer counting, PWM/PSSM scoring, and Markov-chain synthesis (gsynth).
• Database management: Create, link, convert, and manage misha-compatible genomic databases.
• R misha compatibility: Reads and writes the same on-disk formats as R misha (123/145 R exports covered).

Installation

pip install pymisha

Pre-built wheels are available for Linux (x86_64) and macOS (x86_64 and arm64), Python 3.10-3.12.

To install from source (requires a C++17 compiler and numpy):

pip install -e ".[dev]"

Quick start

PyMisha ships with a built-in examples database so you can start exploring immediately -- no external data needed:

import pymisha as pm

# Option 1: one-liner to load the bundled examples database
pm.gdb_init_examples()

# Option 2: equivalent explicit form
pm.gsetroot(pm.gdb_examples_path())

# List available tracks and extract data
print(pm.gtrack_ls())
print(pm.gextract("dense_track", pm.gintervals("chr1", 0, 1000)))

To connect to your own misha database, use gsetroot:

import pymisha as pm

# Initialize the database
pm.gsetroot("/path/to/misha_db")

# Create intervals and extract data
intervals = pm.gintervals_from_strings(["chr1:0-1000", "chr1:2000-2600"])
out = pm.gextract("track1", intervals, iterator=100)

# Filter and summarize
filtered = pm.gscreen("track1 > 0.5", intervals)
stats = pm.gsummary("track1", intervals)

Thread safety

PyMisha inherits R misha's single-threaded design. Keep the following constraints in mind:

• Not thread-safe. All module-level state (_GROOT, _UROOT, _VTRACKS, CONFIG) is process-global and unsynchronized. Do not call PyMisha from multiple threads concurrently.
• One database per process. You cannot have two databases open simultaneously; gsetroot() replaces the active database globally.
• CONFIG is global. Changing settings like max_processes affects every subsequent operation in the process.
• Multiprocessing uses fork(). The C++ backend parallelizes via fork() with shared memory (mmap) and semaphores. This is transparent to the caller but means PyMisha should not be used inside already-forked worker processes or with fork-unsafe libraries.

Examples

Using the built-in example database:

import pymisha as pm

# Quickest way to get started
pm.gdb_init_examples()

# Or equivalently, using gsetroot with the examples path
pm.gsetroot(pm.gdb_examples_path())

print(pm.gtrack_ls())
print(pm.gextract("dense_track", pm.gintervals("chr1", 0, 1000)))

Creating a genome database

PyMisha ships prebuilt genome databases for common assemblies. Download and set up with a single call:

import pymisha as pm

# Download a prebuilt genome (mm9, mm10, mm39, hg19, hg38)
pm.gdb_create_genome("hg38", path="/data/genomes")   # creates /data/genomes/hg38/
pm.gsetroot("/data/genomes/hg38")

pm.gchrom_sizes()  # verify it worked

To build a database from your own FASTA files (e.g. a custom assembly):

pm.gdb_create("/data/my_genome", "genome.fa.gz", verbose=True)
pm.gsetroot("/data/my_genome")

See the Creating Genome Databases tutorial for UCSC download workflows and advanced options.

Optional dependencies

• pyBigWig: For BigWig import in gtrack_import.
• pyreadr + Rscript: For loading R-serialized big interval sets.
• PyYAML: For richer gdataset_info metadata parsing.

Using pymisha with an LLM agent

LLM coding agents (Claude Code, Copilot, Cursor) writing pymisha analysis code can pre-load these reference docs into context for fewer hallucinated APIs and more idiomatic recipes:

• agent-guides/pymisha-core.md — concepts, chooser tables, everyday recipes.
• agent-guides/pymisha-advanced.md — 2D/Hi-C, PWM, import/export, new genomes, gsynth.
• agent-guides/pymisha-anti-patterns.md — silent footguns cross-referenced from the above.
• agent-guides/skills/importing-tracks/SKILL.md — full track-import reference.

Drop-in prompt (no clone needed). Paste the block below into your agent at the start of a pymisha task. It points the agent at the raw files on GitHub, so it works without a local checkout:

Before writing any pymisha code, fetch and read:

- https://raw.githubusercontent.com/tanaylab/pymisha/main/agent-guides/pymisha-core.md  (mandatory: concepts + everyday recipes)
- https://raw.githubusercontent.com/tanaylab/pymisha/main/agent-guides/pymisha-anti-patterns.md  (silent footguns; cross-referenced from core)
- https://raw.githubusercontent.com/tanaylab/pymisha/main/agent-guides/pymisha-advanced.md  (consult on demand: 2D/Hi-C, PWM, import/export, new genomes)

Follow the conventions in those files. When you hit a recipe with an
"Avoid:" block, treat it as a hard rule.

Pin to a release tag for stability by replacing main with any tag that contains agent-guides/. The skills/importing-tracks/SKILL.md guide listed above is load-on-demand; pull it in only when the task specifically calls for track import.

The guides mirror the equivalent set in R misha — same section numbering, same recipes, translated to the pymisha API.

Missing features

Compared to R misha, the following are not yet implemented:

• Track Arrays: gtrack.array.* and gvtrack.array.slice.
• Legacy Conversion: gtrack.convert (for migrating old 2D formats).

License

MIT. See LICENSE for details.