Predicting synergism scores from highly incomplete data via cyclical set projection.

A: To improve screening efficiency further, we introduce a projection-based predictor of synergism scores. An initial guess of a synergism score matrix is projected first onto the set , which corresponds to known interaction scores, then onto the set , which contains matrices of approximately low rank, and finally onto , holding the matrices consistent with known functional similarity. The projections are applied cyclically until convergence to a final prediction of is reached, which is guaranteed due to convexity of the three sets (here illustrating convergence in one iteration). B: Prediction accuracy in five glioblastoma cell lines and reference data sets. Comparison between our projection-based method and two state-of-the-art methods for interaction score imputation methods, LLS and EMDI. Generally, set based projections outperform the other methods (predictions correlate more with true values), especially when the screened fraction is small.