A fast and interpretable deep learning approach to accurate electrostatics-driven pKa prediction

@article{pkai,
author = {Reis, Pedro B. P. S. and Bertolini, Marco and Montanari, Floriane and Machuqueiro, Miguel and Clevert, Djork-Arné},
title = {pKAI: A fast and interpretable deep learning approach to accurate electrostatics-driven pKa prediction},
note = {in preparation}
}

We recommend installing pKAI on a conda enviroment. The pKAI+ model will be downloaded on the first execution and saved for subsequent runs.

python3 -m pip install pKAI

pKAI <pdbfile>

It can also be used as python function,

from pKAI.pKAI import pKAI

pks = pKAI(pdb)

where each element of the returned list is a tuple of size 4. (chain, resnumb, resname, pk)

pKAI+ (default model) aims to predict experimental pK_a values from a single conformation. To do such, the interactions characterized in the input structure are given less weight and, as a consequence, the predictions are closer to the pK_a values of the residues in water. This effect is comparable to an increase in the dielectric constant of the protein in Poisson-Boltzmann models. In these models, the dielectric constant tries to capture, among others, electronic polarization and side-chain reorganization. When including conformational sampling explicitly, one should use a lower value for the dielectric constant of the protein. Likewise, one should use pKAI -- instead of pKAI+ -- as in this model there is no penalization of the interactions' impact on the predicted pK_a values.

tl;dr version

• use pKAI+ for pK_a predictions arising from a single structure
• use pKAI for pK_a predictions arising from multiple conformations

Change the model to be used in the calculation by evoking the model argument:

pKAI <pdbfile> --model pKAI

Performed on 736 experimental values taken from the PKAD database¹.

[1] Pahari, Swagata et al. "PKAD: a database of experimentally measured pKa values of ionizable groups in proteins." doi:10.1093/database/baz024

[2] Thurlkill, Richard L et al. “pK values of the ionizable groups of proteins.” doi:10.1110/ps.051840806

[3] Olsson, Mats H M et al. “PROPKA3: Consistent Treatment of Internal and Surface Residues in Empirical pKa Predictions.” doi:10.1021/ct100578z

[4] Reis, Pedro B P S et al. “PypKa: A Flexible Python Module for Poisson-Boltzmann-Based pKa Calculations.” doi:10.1021/acs.jcim.0c00718

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

Please submit a github issue to report bugs and to request new features. Alternatively, you may email the developer directly.