Awesome
Generative molecular design in low data regimes
Table of Contents
- Description
- Requirements
- How to run an example experiment
- How to run an experiment on your own data
- What can be found in the results?
- How is the pipeline working?
- FAQ
- How to cite this work
- License
Description<a name="Description"></a>
Supporting code for the paper «Generative molecular design in low data regimes»
Access without a paywall via SharedIt
Preprint version (not up to date with the published version)
Abstract of the paper: Generative machine learning models sample molecules from chemical space without the need for explicit design rules. To enable the generative design of innovative molecular entities with limited training data, a deep learning framework for customized compound library generation is presented, aiming to enrich and expand the pharmacologically relevant chemical space with druglike molecular entities ‘on demand’. This de novo design approach combines best practices, and was used to generate molecules that incorporate features of both bioactive synthetic compounds and natural products, which are a primary source of inspiration for drug discovery. The results show that the data-driven machine intelligence acquires implicit chemical knowledge and generates novel molecules with bespoke properties and structural diversity. The method is available as an open-access tool for medicinal and bioorganic chemistry.
Keywords: drug design; generative model; language model; chemical virtual libraries; natural product
Requirements<a name="Requirements"></a>
First, you need to clone the repo:
git clone https://github.com/ETHmodlab/virtual_libraries
Then, you can run the following command which will create a conda virtual environement and install all the needed packages (if you don't have conda, you can fist install it by following the instructions here: https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html):
If you are working with a Mac:
cd virtual_libraries/
bash install_mac.sh
If you are working with a linux system:
cd virtual_libraries/
bash install_linux.sh
Once the installation is done, you can activate the virtual conda environement for this project:
conda activate mollib
Please note that you will need to activate this virtual conda environement every time you want to use this project.
How to run an example experiment<a name="Run"></a>
Now, you can run an example experiment, e.g., the one with five dissimilar molecules from MEGx in the paper:
cd experiments/
bash run_morty.sh ../data/paper_five_dissimilar_MEGx.txt
The results of the analysis can be found in experiments/results/paper_five_dissimilar_MEGx/
Please be aware that you should have one experiment running at the same time for a given transfer learning set.
How to run an experiment on your own data<a name="OwnData"></a>
To apply transfer learning on your own set of molecules, you will need a .txt file with one SMILES string per line. Your file should be in the folder virtual_libraries/data/
Then, you can just run the following command:
bash run_morty.sh ../data/{your_file}.txt
You will find the results of your experiment in experiments/results/{name_of_your_file}/
What can be found in the results folder?<a name="Results"></a>
1. A figure displaying the results, as in the paper:<a name="Results1"></a>
(in experiments/results/{name_of_your_file}/resume/resume_0.7.jpg)
a. Fréchet ChemNet Distance (FCD) plot showing how the generated molecules evolved during transfer learning with respect to the source space (ChEMBL24)
and the target space (in the paper: MEGx natural products).
b. Boxplot displaying the evolution of the CSP3 fraction during transfer learning.
c. Uniform Manifold Approximation and Projection (UMAP), which displays molecules from the training set (ChEMBL24) used to pretrain the
chemical language model (CLM),
molecules from the target set (MEGx), molecules generated from the pretrained CLM, molecules generated at the last epoch of transfer learning and finally
molecules of the transfert learning set.
d. The five most common scaffolds for which molecules were generated, along with their scaled Shannon entropy (SSE)
and the time during which they were present (in percent with respect to all sampled molecules at the given epoch).
2. An interactive UMAP plot that can be opened in the browser:<a name="Results2"></a>
When the example run has finished, the interactive map of chemical space can be found there: experiments/results/paper_five_dissimilar_MEGx/umap/interative_umap.html or in experiments/results/{name_of_your_file}/umap/interative_umap.html if you ran an experiment on your own data). You can open this html file with a web browser. This interactive map allows you to zoom in our out and to see some properties of the molecules when you'll pass over them with your pointer. Finally, if you click on a molecule, it will open it on http://molview.org where you can see its structure.
3. The generated novo molecules by the chemical language model:<a name="Results3"></a>
You can find the list of molecules (.txt file with one SMILES string per line) in experiments/results/{name_of_your_file}/novo_molecules/. Each of those molecules are grammatically valid and not found in the training set. Molecules were sampled every 10 epochs (10, 20, 30 and 40) during transfer learning.
Note: if you change the number of epochs or the period at which the pretrained chemical language models are saved in experiments/parameters.ini, the UMAPs and the resume will not be computed.
How is the pipeline working?<a name="Pipeline"></a>
In experiments/ you can find all the files used to process the data, train the chemical language model and analyze the results.
All do_{something}.py files are where the actual code resides. All those files are run from a bash script called run_morty.sh,
which executes other bash scripts called run_do_{something}.sh. Each of those run_do_{something}.sh files takes care of the compute pipeline,
namely (i) preprocessing, (ii) training, (iii) data generation, and (iv) data analysis. You can use any of these scripts seperately if needed.
The parameters of the neural network, how to do the processing, etc ., are defined in the file parameters.init.
If you want, you can play with it. However, keep in mind that you will need access to a GPU if you wish to (re)train a language model,
which is already provided here (otherwise, it will take a very long time to run the full pipeline).
Note that there are some additional parameters and helper functions defined in src/python/.
FAQ<a name="FAQ"></a>
Do I need a GPU to run this code?
No, you don't. We already trained the chemical language model, which is the part for which you really need a GPU. Even though it might take some time to
sample new molecules, you can do it on your own computer – which was one of our goals.
I have never used a terminal, nor do I know what Git really is. What should I do?
If your goal is just to use our project to generate molecules, then we suggest that you use the release we made on
Code Ocean (https://doi.org/10.24433/CO.0753661.v1). There, you have the possibility to run the code without
having to install anything.
Can I use this code if I work with Windows?
We did not test our code with Windows. If you try, please let us know what was the modifications you had to make to the requirements such that
everybody can benefit from your experience!
How to cite this work<a name="Cite"></a>
@article{Moret2020,
title={Generative molecular design in low data regimes},
author={Moret, Michael and Friedrich, Lukas and Grisoni, Francesca and Merk, Daniel and Schneider, Gisbert},
journal={Nature machine intelligence},
volume={2},
number={3},
pages={171–180},
year={2020},
doi={10.1038/s42256-020-0160-y},
url={https://doi.org/10.1038/s42256-020-0160-y}
}