Recent developments in computational techniques have advanced drug discovery and design. For example, standard databases with known chemicals and their modes of actions can be considered by machine learning to classify new drugs. Here, we were interested in the vectorized presentations of the structure of small molecules, a crucial first step towards any data analytics in computational chemistry. Vectorized presentations were inferred through the construction of autoencoders. We followed the current literature trends and used a combination of convolutional and recurrent layers. Experimental results show that our model is comparable to standard chemical fingerprints, where on some of the test databases even provides for improved accuracy. We published the code to infer the embedder in open source on the GitHub repository and included the embedder within the fingerprinting widget for Orange data mining suite.
|
