Comparison of Cutoff Strategies for Geometrical Features in Machine Learning-Based Scoring Functions

摘要

Countings of protein-ligand contacts are popular geometrical features in scoring functions for structure-based drug design. When extracting features, cutoff values are used to define the range of distances within which a protein-ligand atom pair is considered as in contact. But effects of the number of ranges and the choice of cutoff values on the predictive ability of scoring functions are unclear. Here, we compare five cutoff strategies (one-, two-, three-, six-range and soft boundary) with four machine learning methods. Prediction models are constructed using the latest PDBbind v2012 data sets and assessed by correlation coefficients. Our results show that the optimal one-range cutoff value lies between 6 and 8 A instead of the customary choice of 12 A. In general, two-range models have improved predictive performance in correlation coefficients by 3-5%, but introducing more cutoff ranges do not always help improving the prediction accuracy.