Gaussian process regression performs better than DFT and Karplus calculations for 1JCαH Spin-spin Scalar Coupling Constants

OSVALDO MARTIN; AGUSTINA ARROYUELO; JORGE VILA; HAROLD SCHERAGA; OSVALDO MARTIN; AGUSTINA ARROYUELO; JORGE VILA; HAROLD SCHERAGA

San Luis

Congreso; XLVIII Reunión Anual de la Sociedad Argentina de Biofísica; 2019

In the present work we explore three different approaches for the computation of theone-bond spin-spin coupling constants (SSCC) 1JCαH in proteins: DFT calculation,Karplus-like equation and Gaussian Process regression. The main motivation of this workis to select the best method for a fast and accurate computation of the 1JCαH SSCC, forits use on everyday applications in protein structure validation, refinement and/ordetermination. Initially, we observed a poor agreement between the DFT computed1JCαH SSCC values and the experimental ones. With a detailed analysis of the causes ofthis observation we ruled out the choice of functional, basis set, solvent effects andpeptide-geometry, as possible sources of error. However, we found that the DFTcomputed 1JCαH SSCC values depend on both the ψ, φ and χ1 torsional angles and theenvironment of the peptide model used to compute it. All of which leads to theunderstanding that the model chosen for the DFT computations is inappropriate. Hence,we discuss different models for solving this issue as well as the non-transferability of saidsolutions. Additionally, we show that Karplus-like equation and Gaussian Processregression provide faster and more accurate results than DFT-based calculations.