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 pKa 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 pKa 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 pKa values.
tl;dr version
- use pKAI+ for pKa predictions arising from a single structure
- use pKAI for pKa 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.
| Method | RMSE | MAE | Quantile 0.9 | Error < 0.5 (%) |
|---|---|---|---|---|
| Null2 | 1.09 | 0.72 | 1.51 | 52.3 |
| PROPKA3 | 1.11 | 0.73 | 1.58 | 51.1 |
| PypKa4 | 1.07 | 0.71 | 1.48 | 52.6 |
| pKAI | 1.15 | 0.75 | 1.66 | 49.3 |
| pKAI+ | 0.98 | 0.64 | 1.37 | 55.0 |
[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.