A fast 23andMe genome text file parser, now superseded by arv
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
py-dnatraits Adds accessor to SNPS, new benchmarks Sep 12, 2016
util Adds 2016 to copyright Sep 3, 2016
CMakeLists.txt Adds unfinished test target Sep 5, 2016
LICENSES Reorganizes directory structure Nov 16, 2014
Makefile Adds repl target Sep 7, 2016
TODO Updates TODO Nov 24, 2014
build-make.sh Times execution Sep 4, 2016


Project has moved!

This project has been abandoned in favor of the newer and better https://github.com/cslarsen/arv — which is faster, installable via pip and works on both Python 2 and 3.

The dna-traits 23andMe parser library

This is a very fast 23andMe raw genome file parser, letting you lookup SNPs from RSIDs in Python:

import dna_traits as dt

genome = dt.parse("genome.txt")

gender = "man" if genome.y_chromosome else "woman"
complexion = "light" if genome.rs1426654 == "AA" else "dark"

eye_color = dt.unphased_match(genome.rs12913832, {
    "AA": "brown",
    "AG": "brown or green",
    "GG": "blue"})

print("You are a {gender} with {eye_color} eyes and {complexion} skin.".format(
    gender=gender, eye_color=eye_color, complexion=complexion))

In my case, this little program produces

You are a man with blue eyes and light skin.

The need for speed

On my machine, a 2010-era MBP with SSD, I can parse a 24 Mb file using Python's csv module and create a dictionary in 2.5 seconds. Pandas takes around 2.1 seconds, and I've seen some parsers take up to 8.

In comparison, dna_traits takes only 0.12 seconds (g++ 5.3), and uses dramatically less memory (six bytes per SNP, containing its nucleotide pair, chromosome and position). A more recent 2013-era, Intel Xeon Linux box does the same in a blazing 0.07 seconds!

While I love Pandas for its power and generality, this library is fast because it is meticulously specialized: The finely-tuned C++ backend memory maps the file and never scans backwards — every single byte is touched exactly once. The SNPs are stored in a memory efficient packed struct and stored in Google's dense hash map, keyed by its 32-bit RSID.

The Python API

The project is split up into two parts: A C++ API, and a Python 2 module front-end. It should be straight-forward to create bindings for other languages as well.

Here are some examples illustrating various facets of the Python API:

>>> import dna_traits as dt
>>> genome = dt.parse("genome.txt")
>>> genome
<Genome: SNPs=949905, y_chromosome=True, orientation=1>
>>> genome.male
>>> genome["rs123"]
SNP(genotype='AA', rsid='rs123', orientation=1, chromosome=7, position=24966446)
>>> snp = genome.rs123
>>> snp.homozygous
>>> ~snp
SNP(genotype='TT', rsid='rs123', orientation=1, chromosome=7, position=24966446)

More information can be found in py-dnatraits/README.md

Current issues

  • The Python module is not ready for PyPi

  • Depends on GNU Make, but an incomplete CMake configuration is in progress.

  • The Python API is currenty somewhat limited and inconsistent, but still very much usable!

  • Doesn't parse 23andMe internal IDs yet.

  • Although loading the file is fast, whole genome iteration is insanely slow in Python. It's because I don't really expose an iterator from C++ to Python yet.


The main build system is GNU Make, but I am slowly transitioning to CMake.


  • A C++11 compiler

  • Google sparse hash map

  • Genome files in 23andMe format. Many people have uploaded theirs on the net for free use. See for example OpenSNP. If you're a 23andMe customer, you can download your own from them.

  • Python development files, if you want to build the Python module.

GNU Make

If Google dense hash map is located in /usr/local/include, build everything, including the Python API, with:

$ make -j all CXXFLAGS=-I/usr/local/include

Build the check target to check if everything works.

$ make check


I'm transitioning to CMake, but it's currently not working properly. See build-ninja.sh and build-make.sh to test out the current status.

Example of inferring phenotypes

SNPedia contains the gs237 criteria for determining whether a person has blue eyes. At http://snpedia.com/index.php/Gs237/criteria the rule set says:


In C++ this would be:

static bool gs237(const Genome& genome)
  return genome[ 4778241] ==  CC
      && genome[12913832] ==  GG
      && genome[ 7495174] ==  AA
      && genome[ 8028689] ==  TT
      && genome[ 7183877] ==  CC
      && genome[ 1800401] == ~CC;

The only thing to note is each SNP's orientation. 23andMe uses positive orientation, while SNPedia has varying orientation. That's why we flip the orientation in the last check for the rs1800401 SNP

In Python, this can be done in any number of ways, but one way is to use the Genome.match function:

all(genome.match((("rs4778241",  "CC"),
                  ("rs12913832", "GG"),
                  ("rs7495174",  "AA"),
                  ("rs8028689",  "TT"),
                  ("rs7183877",  "CC"),
                  ("rs1800401",  "GG"))))

Example reports

I've implemented some of the 23andMe health reports, traits and conditions. You can see the content here:

Note that there might very well be errors in the above reports, so use with care! In particular, the odds ratios are most likely very incorrect. The best bet is the traits. These should be correct, as well as some simple health reports. So, the main goal is for this to be educational. In other words, explore on your own.


Copyright (C) 2014, 2016 Christian Stigen Larsen

Distributed under GPL v3 or later. See the file COPYING for the full license. This software makes use of open source software; see LICENSES for details.

Additional disclaimer

In addition to the GPL v3 license terms, and given that this code deals with health-related issues, I want to stress that the provided code most likely contains errors, or invalid genome reports. Results from this code must be interpreted as HIGHLY SPECULATIVE and may even be downright INCORRECT. Always consult a medical expert for guidance. I take NO RESPONSIBILITY whatsoever for any consequences of using this code, including but not limited to loss of life, money, spuses, self-esteem and so on. Use at YOUR OWN RISK.

The indended use is for casual, educational purposes. If this code is used for research purposes, I would be happy if you should cite me. If so, beware that the parser code may contain serious errors. I would advise you to check for updates, and also to double-check results with other parsers.

Places of interest