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.
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:
- 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
- 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.
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.
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.
conda env create -f environment.yml
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.
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!