Is this what we are talking about:

http://en.wikipedia.org/wiki/Data_validation_and_reconciliation

or probably:

http://en.wikipedia.org/wiki/Missing_data

best regards

Johannes

Yes, I am referring to the "Missing data" problem as described in Wikipedia. In the event of a missing data point, there are several possible alternatives. One is to impute artificial data in order to have a complete dataset. Several data imputation methods have been proposed in the literature.

But, I am specifically asking about the problem of missing data when considered in combination with the network inference problem: i.e., when we want to apply a network inference method (e.g. ARACNe, CLR, etc) to a dataset with missing values. Usually network inference methods --like the ones I mentioned above-- need complete datasets, so the user would need to 'curate' those datasets, by applying some data imputation technique, before the network inference step. What I am asking is whether there are network inference methods available which can handle datasets with missing values automatically, i.e. without forcing the user to 'repair' the datasets before.

Hope this clarifies the question,

Alex

A suggestion is to use something like "multiple imputation", which tends to work with any analysis method which requires "full" datasets (see links for details).

The main difficulty with such approaches is to choose an appropriate "result pooling" approach. In the specific cases you mention (i.e. network inference from microarray transcriptomic data), it probably depends on the type of graphs provided by the algorithm (weighted/unweighted, directed/undirected). I guess a naive approach could be to just compare the different outputs of a multiple imputation analysis and see if the structure/property you are interested in (e.g. hubs/cliques, maximum/average distance between nodes, node degree) is mostly preserved for all imputations: if yes, then inference is probably feasible even in the presence of "missing values".

http://sites.stat.psu.edu/~jls/mifaq.html

http://www.stefvanbuuren.nl/mi/MI.html

ODE based methods should be able to handle missing data points? At least the method I am working on does. Also I had heard about modular response analysis (MRA) being quite flexible in this respect.

Weird place for a review like this, but you could also check out http://www.ncbi.nlm.nih.gov/pubmed/23293646.

My first thought was something bayesian to make the most of the data to hand, but I also came across this: http://www.comp-sys-bio.org/CMSB14/abstracts/cmsb2014_submission_48.pdf

Thanks for the link to the review, Kolja, I'll check it out.

@Jonathan: I'm one of the coauthors of the abstract you mention, so I'm already familiar with that work ;-) I'm actually looking for similar things!