Electrostatic Potential Surface(ESP) calculation with GCNN #RDKit #chemoinformatics

ESP is a key feature in Drug Discovery. There are many publications discussing ESP in Drug Design. However getting accurate ESP is time-consuming because it needs high level QM calculations. To reduce the calculation cost of QM, predict quantum nature by using Deep learning is researched.

And some days ago, I found interesting article published by reseachers in Astex. URL is below.
https://pubs.acs.org/doi/10.1021/acs.jmedchem.9b01129

They build GCNN model with over than 100,000 diverse drug like molecules QM calculation (B3LYP/6-31G* basis set) data set. And their model showed good performance in ESP calculation, pKa prediction and binding affinity prediction.

And also GCNN based calculation is faster than QM based calculation, Fig 3 shows that their approach is 4 order faster than conventional DFT based calculation.

Luckly, author disclosed the code on github. ;)
https://github.com/AstexUK/ESP_DNN

….But,,,,,, the code is written for python 2.x.
Hmm…. I would like to use the code in python 3.x.

So I forked ESP_DNN and modify the original code. You can find the code here.
https://github.com/iwatobipen/ESP_DNN

My env for the ESP_DNN is below.
Python 3.7
numpy 1.17.2
rdkit 2019.03.4
xarray 0.14.0
tensorflow-gpu 1.14.0
keras-base 2.2.4

Install is very easy.

$ git clone https://github.com/iwatobipen/ESP_DNN.git
$ cd ESP_DNN
$ pip install -e . # or/ pip install .

Then run the example code which is introduced in README.md

ESP_DNN$ python -m esp_dnn.predict -m ligand -i examples/ligands/

After running the code forementioned, pqr files are generated in -i folder.
It can be visualized in http://nglviewer.org/ngl/ where is recommend in the document.

Open the URL with web browser and read a pqr file.
Then change and set visualize settings according to the document.
Finally I could get image with ESP. Following ESP is generated from provided GCNN model.

QM based approach is powerful for drug discovery however the calculation cost is high. On the other side, deep learning based approach is very fast but it requires high quality training data.

So I think the authors work is very useful because they provided not only their code but also high quality trained model. Awesome work!

If I have to say something, in the original repo seems that training dataset is not provided. I would like to check chemical space of training data.

Anyway, calculation of accurate ESP in short time is very useful for me. I would like to apply some prediction tasks ASAP.

Advertisement