In the GeneTrack software we have used HDF to store values for each genomic base. Its main advantage over other storage systems was that it was able to return consecutive values with minimal overhead.

For example it is extremely fast (ms) in retrieving say 100,000 consecutive values starting with a certain index.We used the Python bindings to HDF. An added advantage of these bindings is that they will return the data back as numpy arrays (very fast numerical operations).

Here is the relevant code that deals with HDF only: hdf.py

The HDF schema is set up in a different module, but in the end it simply something like:

class MySchema( IsDescription ):
    """
    Stores a triplet of float values for each index.
    """
    ix = IntCol  ( pos=1 )  # index
    wx = FloatCol( pos=2 )  # values on the W (forward) strand
    cx = FloatCol( pos=3 )  # value on the C (reverse) strand
    ax = FloatCol( pos=4 )  # weighted value on the combined W + C strands

What I do have is a netCDF-3 based Java application that I could show you. NetCDF-3 is basically the same idea as HDF, but quite more limited as it cannot do compound datatypes among other limitations.

But here's a small test code example to toy with:

Use the above in the following way:

Hello Pierre!

I have been talking with the BioHDF guys and from what they tell me, their work will be centered around a number of command-line APIs, written in C, that will address some areas of usage which for now do not seem to overlap.

I have seen this example on their site: http://www.hdfgroup.org/projects/biohdf/biohdf_tools.html Don't know if that helps.

I have been talking with them to see if we can achieve an API for saving genotype data. Don't know yet where that will lead me.

If you are looking for something more versatile, you will probably have to delve in the official HDF5 C code ( http://www.hdfgroup.org/HDF5/Tutor/ ), which seems to be the only one that offers all the functionality and goodies of that impressive storage system.

There is also a Perl binding to HDF5: PDL::IO::HDF5

http://search.cpan.org/~cerney/PDL-IO-HDF5-0.5/ This requires the Perl Data Language (PDL) package. The way, data-structures can be handled, sub-ranges of data can be defined an data can be manipulated is actually very elegant in PDL such that computational code can profit from PDLs vectorized style of writing expressions.

The same is true for R and the hdf5 package: http://cran.r-project.org/web/packages/hdf5/index.html

Code examples are in the package documentations of both, the R-hdf5 package documentation is quite little though.

Both of these language bindings might be a very efficient way to read and write HDF5 files.

There are also APIs in Fortran, Java, Python, Matlab, C, or C++. So it might make sense to select the language and define the type of data you wish to store first.

I've been talking a bit with one of the devs behind BioHDF (being at UIUC, up the road from The HDF Group doesn't hurt). I believe a publication is on the way describing it along with some implementation details.

Not BioHDF5 but probably readable and maintained:

HDF5 for Python http://code.google.com/p/h5py/

Unfortunately I don't have any example to shows you yet. I don't know how to program in C/C++ so I have been looking at two hdf5 wrappers in python, PyTables and H5PY.

PyTables has a database-like approach in which HDF5 is used as a sort of hierarchical database, in which a column can be a table itself, allowing to store nested data. For example, you have a table called 'SNPs' with two columns, 'id' and 'genotypes'; the column 'genotypes' contains a nested table, with the columns 'individual' and 'genotype'; and so on.

H5Py is basically a re-implementation of numpy's arrays, so you can store and access arrays/matrixes as you would do with numpy (it is similar to arrays and matrixes in matlab, R, and any other language with this data type) and they are stored in an HDF5 file so the access is faster.

Our Genomedata system stores multiple tracks of 1-bp resolution genomic data in a HDF5 array. Documentation and full source code is available on that page. It has a Python (PyTables) interface for reading the data. For originally loading it into HDF5, we wrote a C loader for added speed.

Another answer: I've been looking at the IGV ( Integrative Genomics Viewer ) sources, it use the JAVA bindings for HDF5 to store the data about the genomic tracks.