Motivation: A major limitation in modeling protein interactions is the difficulty of assessing the over-fitting of the training set. Recently, an experimentally based approach that integrates crystallographic information of C2H2 zinc finger-DNA complexes with binding data from 11 mutants, 7 from EGR finger I, was used to define an improved interaction code (no optimization). Here, we present a novel mixed integer programming (MIP)-based method that transforms this type of data into an optimized code, demonstrating both the advantages of the mathematical formulation to minimize over- and under-fitting and the robustness of the underlying physical parameters mapped by the code.
Results: Based on the structural models of feasible interaction networks for 35 mutants of EGR-DNA complexes, the MIP method minimizes the cumulative binding energy over all complexes for a general set of fundamental protein-DNA interactions. To guard against over-fitting, we use the scalability of the method to probe against the elimination of related interactions. From an initial set of 12 parameters (six hydrogen bonds, five desolvation penalties and a water factor), we proceed to eliminate five of them with only a marginal reduction of the correlation coefficient to 0.9983. Further reduction of parameters negatively impacts the performance of the code (under-fitting). Besides accurately predicting the change in binding affinity of validation sets, the code identifies possible context-dependent effects in the definition of the interaction networks. Yet, the approach of constraining predictions to within a pre-selected set of interactions limits the impact of these potential errors to related low-affinity complexes.
Supplementary information: Supplementary data are available at Bioinformatics online.