Success in the DREAM3 signaling response challenge using simple weighted-average imputation: lessons for...

Our group produced the best predictions overall in the DREAM3 signaling response challenge, being tops by a substantial margin in the cytokine sub-challenge and nearly tied for best in the phosphoprotein sub-challenge. We achieved this success using a simple interpolation strategy. For each combination of a stimulus and inhibitor for which predictions were required, we had noted there were six other datasets using the same stimulus (but different inhibitor treatments) and six other datasets using the same inhibitor (but different stimuli). Therefore, for each treatment combination for which values were to be predicted, we calculated rank correlations for the data that were in common between the treatment combination and each of the 12 related combinations. The data from the 12 related combinations were then used to calculate missing values, weighting the contributions from each experiment based on the rank correlation coefficients. The success of this simple method suggests that the missing data were largely over-determined by similarities in the treatments. We offer some thoughts on the current state and future development of DREAM that are based on our success in this challenge, our success in the earlier DREAM2 transcription factor target challenge, and our experience as the data provider for the gene expression challenge in DREAM3.