Reverse engineering partially-known interaction networks from noisy data

One of the most difficult challenges associated with the problem of inferring functional interaction networks from experimental data is that of dealing with the effects of measurement noise in the data used for reverse engineering. A second important challenge is that of taking full advantage of prior knowledge about some elements of the network to improve the results of the reconstruction process. This paper introduces a new inference algorithm, PACTLS, which addresses both of the above issues. The algorithm combines methods to exploit mechanisms underpinning scale–free networks generation, i.e. network growth and preferential attachment (PA), with a technique to optimally reduce the effects of measurement noise in the data on the reliability of the inference results, i.e. the Constrained Total Least Squares (CTLS) algorithm. The technique is assessed through numerical tests on in silico random networks and is shown to consistently outperform approaches based on Bayesian networks.