February 20, 2017
Deep Neural Networks for drug discovery
Deep Neural Networks (DNNs) have been in the spotlight for last couple of years. This technique has been applied with great success to traditional Machine Learning problems such as Image and Speech recognition. Moreover, it has been applied to other learning problems such as mastering the game of Go being able to beat a human professional of this game for the first time in history.
A DNN is a system made of biologically-inspired neurons that are organized in layers with increasing levels of abstraction. Take for instance a DNN trained to classify paintings by their author. In such system the first layers will be in charge of detecting colors and edges. Subsequent layers will use that information to assess more complex patterns such as shapes, and recurring motives. Finally, the last layers will take all that information into consideration in order to predict the artist that produced a given painting.
DNNs' ability to extract information at different levels of abstraction can also be used to perform creative tasks. For instance, work has been done to use DNNs for speech synthesis, the design of new encryption algorithms, and to apply a given pictorical style to an image.

DNNs have been used in the drug discovery field for different tasks such as Quantitative Structure-Activity Relationship classification outperforming traditional Random Forests approaches [1], structure-based binding affinity prediction [2], and toxicity prediction [3].
At Intelligent Pharma we are using DNNs to improve our virtual screening tools, learning from millions of freely available data points in order to predict the activity/inactivity of a small molecule against known targets.
Research on the applications of DNNs to the drug discovery is an active and thriving field with many open possibilities. One could even imagine taking advantage of DNNs' creative qualities and apply them to lead/hit optimization or drug repurposing.


Ondrej Svoboda
2017-04-19 09:27:48
Image and speech learning are both operating with relatively uniform space of data. There are many combinations but the process itself is not necessarily that complicated. However, drug discovery is complex, you have few hits and lots of misses. I can imagine you fit some neural network which identifies possible new hits according to some molecular similarity. However, I guess those hits will be structurally similar to hits we already know. In other words what if you have an unknown active compound which is structurally very different from known hits? Will you find it using ML?

Leave a comment