The goal of this experiment is the identification of gene networks involving human genes of medical relevance for two broad families of human pathologies: motor neuron diseases (https://en.wikipedia.org/wiki/Motor_neuron_disease) and hematopoietic tumors (https://en.wikipedia.org/wiki/Tumors_of_the_hematopoietic_and_lymphoid_tissues).
Human gene network expansion will take advantage of the comprehensive gene expression dataset provided by the FANTOM project (http://fantom.gsc.riken.jp/5/data/).

I just wanted to add some technicalities: The dataset we are using right now is very big, a floating point matrix 87554x1829, it's size on the hard disk is ~0.5Gb and contains, for many genes, different isoforms/transcripts. The computational time for an algorithm's iteration, because of its 'complexity', is longer than the previous experiments.

Any single gene expansion is packed up into 294 workunits. Any workunit you receive has a name like this: 142041_Hs_T155402-MAP1B_wu-(1 to 294), a counter, Hs, an internal T code, a mnemonic for the gene/isoform. The T code is just a shortcut for the gene coordinates, for example T155402 is chr5:71403265..71403276,+

At this moment we don't know how many experiments we will make with this dataset, that's why I didn't update the 'Science status' page with the usual counters.

Moving to the FANTOM project dataset has been a big challenge for us. The original input contains a slightly different type of data (RNAseq instead of micro-array contrasts) and has 10 times the variables of the biggest data matrix we used so far.

Human is known to have around 20.000 protein-coding genes and 5.000 non-coding ones, but the technology employed in the FANTOM project allowed them to analyze expression at a higher resolution, reporting data also for alternative splicing. The raw dataset reports expression levels for more then 200.000 transcripts in 1829 conditions, wich would been too big to be explored, so we had to filter it.
A strongly related challenge comes from the highly specificity of certain transcripts, that are expressed only in few conditions. This is translated in lines of data with very few non-zero entries that are very likely to correlate one with another.

We are currently testing two version of the input dataset. The first one is a general filtering on the lines that have no HGNC code (www.genenames.org). The second is a more fine grained filtering, where we have removed also transcripts with less then 10% non-zero entries and gene-isoforms too similar one with another.

Most of the work in this moment is focused toward understanding which one of the two dataset is performing "better".

Thank you for all your support.