Optimization of minimum set of protein-DNA interactions: a quasi exact solution with minimum over-fitting

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein-DNA Interactions

Background:

• Modeling protein interactions is challenged by difficulties in assessing training set over-fitting.
• Previous experimental approaches integrated crystallographic data and mutant binding data for C2H2 zinc finger-DNA complexes.
• An interaction code was previously defined without optimization for EGR finger I mutants.

Purpose of the Study:

• To develop a novel mixed integer programming (MIP)-based method for optimizing protein-DNA interaction codes.
• To minimize over-fitting and under-fitting in interaction modeling using a mathematical formulation.
• To demonstrate the robustness of physical parameters mapped by the optimized code.

Main Methods:

• A mixed integer programming (MIP) approach was employed to transform experimental data into an optimized interaction code.
• The MIP method minimizes cumulative binding energy across feasible interaction networks for EGR-DNA complexes.
• Scalability was used to probe against the elimination of related interactions, guarding against over-fitting.

Main Results:

• The MIP method achieved a high correlation coefficient of 0.9983 after eliminating five parameters from an initial set of twelve.
• Further parameter reduction led to under-fitting, indicating the optimal parameter set.
• The optimized code accurately predicted binding affinity changes in validation sets and identified context-dependent effects.

Conclusions:

• The developed MIP-based method effectively optimizes protein-DNA interaction codes, minimizing over- and under-fitting.
• The approach demonstrates the robustness of underlying physical parameters and improves predictive accuracy for binding affinity.
• Constraining predictions to pre-selected interactions limits error impact on low-affinity complexes.