Skip to main content

Using autoencoders for molecule generation

Some time ago we found the following paper https://arxiv.org/abs/1610.02415 so we decided to take a look at it and train the described model using ChEMBL.

Lucky us, we also found two open source implementations of the model; the original authors one https://github.com/HIPS/molecule-autoencoder and https://github.com/maxhodak/keras-molecules. We decided to rely on the last one as the original author states that it might be easier to have greater success using it.

What is the paper about? It describes how molecules can be generated and specifically designed using autoencoders.

First of all we are going to give some simple and not very technical introduction for those that are not familiar with autoencoders and then go through a ipython notebook showing few examples of how to use it.

  1. Autoencoder introduction


Autoencoders are one of the many different and popular unsupervised deep learning algorithms used nowadays for many different fields and purposes. These work with two joint main blocks, an encoder and a decoder. Both blocks are made of neural networks.
In classical cryptography the cryptographer defines an encoding and decoding function to make the data impossible to read for those people that might intercept the message but do not have the decoding function. A classical example of this is the Caesar’s cipher https://en.wikipedia.org/wiki/Caesar_cipher .

However using autoencoders we don’t need to set up the encoding and decoding functions, this is just what the autoencoder do for us. We just need to set up the architecture of our autoencoder and the autoencoder will automatically learn the encoding and decoding functions minimizing a loss (also called cost or objective) function in a optimization algorithm. In an ideal world we would have a loss of 0.0, this would mean that all data we used as an input is perfectly reconstructed after the encoding. This is not usually the case :)

So, after the encoding phase we get a intermediate representation of the data (also called latent representation or code). This is why it’s said that autoencoders can learn a new representation of data.

Two most typical scenarios using autoencoders are:

  1. Dimensionality reduction: Setting up a bottleneck layer (layer in the middle) with lower dimensionality than the input layer we get a lower dimensionality representation of our data in the latent space. This can be somehow compared using classic PCA. Differences between using autoencoders and PCA is that PCA is purely linear, while autoencoders usually use non-linear transfer functions (multiple layers with relu, tanh, sigmoid... transfer functions). The optimal solution for an autoencoder using only linear transfer functions is strongly related to PCA. https://link.springer.com/article/10.1007%2FBF00332918

  1. Generative models: As the latent representation (representation after encoding phase) is just an n-dimensional array it can be really tempting to artificially generate n-dimensional arrays and try decode them in order to get new items (molecules!) based on the learnt representation. This is what we will achieve in the following example.

  1. Model training


Except RNN, most machine/deep learning approaches require of a fixed length vector as an input. The authors decided to take smiles no longer of 120 characters for a further one hot encoding representation to feed the model. This only left out less than 3% of molecules in ChEMBL. All of them above 1000 dalton.

We trained the autoencoder using the whole ChEMBL database (except that 3%), using a 80/20 train/test partition with a validation accuracy of 0.99.

  1. Example


As the code repository only provided a model trained with 500k ChEMBL 22 molecules and training a model against ChEMBL it’s a quite time expensive task we wanted to share with you the model we trained with whole ChEMBL 23 and a ipython notebook with some basic usage examples.

To run locally the notebook you just need to clone the repository, create a conda environment using the provided environment.yml file and run jupyter notebook.

cd autoencoder_ipython
conda env create -f environment.yml
jupyter notebook


The notebook covers simple usage of the model:

  • Encoding and decoding a molecule (aspirin) to check that the model is working properly.
  • Sampling latent space next to aspirin and getting auto-generated aspirin neighbours(figure 3a in original publication), validating the molecules and checking how many of them don’t exist in ChEMBL.
  • Interpolation between two molecules. Didn’t worked as well as in the paper.

  1. Other possible uses for this model


As stated on the original paper, this model can be also used to optimize a molecule given a desired property AlogP, QED...

Latent representations of molecules can be also used as structural molecular descriptors for target predictions algorithms.
Most popular target prediction algorithms are using fingerprints. Fingerprints have an obvious loss of structure information; molecule can’t be reconstructed from its fingerprint representation. As latent representation is saving all 2D molecule structural information in most cases (0.99 accuracy on ChEMBL test set)  we also believe that it may be used to improve fingerprint based target prediction algorithms accuracy.

Hope you enjoy it!

Eloy.


Comments

Popular posts from this blog

UniChem 2.0

UniChem new beta interface and web services We are excited to announce that our UniChem beta site will become the default one on the 11th of May. The new system will allow us to better maintain UniChem and to bring new functionality in a more sustainable way. The current interface and web services will still be reachable for a period of time at https://www.ebi.ac.uk/unichem/legacy . In addition to it, the most popular legacy REST endpoints will also remain implemented in the new web services: https://www.ebi.ac.uk/unichem/api/docs#/Legacy Some downtime is expected during the swap.  What's new? UniChem’s current API and web application is implemented with a framework version that’s not maintained and the cost of updating it surpasses the cost of rebuilding it. In order to improve stability, security, and support the implementation and fast delivery of new features, we have decided to revamp our user-facing systems using the latest version of widely used and maintained frameworks, i

A python client for accessing ChEMBL web services

Motivation The CheMBL Web Services provide simple reliable programmatic access to the data stored in ChEMBL database. RESTful API approaches are quite easy to master in most languages but still require writing a few lines of code. Additionally, it can be a challenging task to write a nontrivial application using REST without any examples. These factors were the motivation for us to write a small client library for accessing web services from Python. Why Python? We choose this language because Python has become extremely popular (and still growing in use) in scientific applications; there are several Open Source chemical toolkits available in this language, and so the wealth of ChEMBL resources and functionality of those toolkits can be easily combined. Moreover, Python is a very web-friendly language and we wanted to show how easy complex resource acquisition can be expressed in Python. Reinventing the wheel? There are already some libraries providing access to ChEMBL d

LSH-based similarity search in MongoDB is faster than postgres cartridge.

TL;DR: In his excellent blog post , Matt Swain described the implementation of compound similarity searches in MongoDB . Unfortunately, Matt's approach had suboptimal ( polynomial ) time complexity with respect to decreasing similarity thresholds, which renders unsuitable for production environments. In this article, we improve on the method by enhancing it with Locality Sensitive Hashing algorithm, which significantly reduces query time and outperforms RDKit PostgreSQL cartridge . myChEMBL 21 - NoSQL edition    Given that NoSQL technologies applied to computational chemistry and cheminformatics are gaining traction and popularity, we decided to include a taster in future myChEMBL releases. Two especially appealing technologies are Neo4j and MongoDB . The former is a graph database and the latter is a BSON document storage. We would like to provide IPython notebook -based tutorials explaining how to use this software to deal with common cheminformatics p

ChEMBL 30 released

  We are pleased to announce the release of ChEMBL 30. This version of the database, prepared on 22/02/2022 contains: 2,786,911 compound records 2,157,379 compounds (of which 2,136,187 have mol files) 19,286,751 activities 1,458,215 assays 14,855 targets 84,092 documents Data can be downloaded from the ChEMBL FTP site: https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/releases/chembl_30/ Please see ChEMBL_30 release notes for full details of all changes in this release:  https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/releases/chembl_30/chembl_30_release_notes.txt New Deposited Datasets EUbOPEN Chemogenomic Library (src_id = 55, ChEMBL Document ID CHEMBL4689842):   The EUbOPEN consortium is an Innovative Medicines Initiative (IMI) funded project to enable and unlock biology in the open. The aims of the project are to assemble an open access chemogenomic library comprising about 5,000 well annotated compounds covering roughly 1,000 different proteins, to synthesize at least

Multi-task neural network on ChEMBL with PyTorch 1.0 and RDKit

  Update: KNIME protocol with the model available thanks to Greg Landrum. Update: New code to train the model and ONNX exported trained models available in github . The use and application of multi-task neural networks is growing rapidly in cheminformatics and drug discovery. Examples can be found in the following publications: - Deep Learning as an Opportunity in VirtualScreening - Massively Multitask Networks for Drug Discovery - Beyond the hype: deep neural networks outperform established methods using a ChEMBL bioactivity benchmark set But what is a multi-task neural network? In short, it's a kind of neural network architecture that can optimise multiple classification/regression problems at the same time while taking advantage of their shared description. This blogpost gives a great overview of their architecture. All networks in references above implement the hard parameter sharing approach. So, having a set of activities relating targets and molecules we can tra