The application of deep learning in drug discovery

We live in a world of “big data”, a term that is becoming widely used and familiar in the media. But big data, alone, is not particularly useful. What we need are clever ways of analysing those data and learning from them. Enter “deep learning“, a class of computer algorithms that promises to help meet this challenge. Deep neural networks are an example of this relatively new form of computation. These computer programs possess the ability to model highly complex and non-linear data using multiple so-called “hidden layers”.

Deep learning hit the headlines recently after Google’s alphaGo program beat a human champion at the ancient Chinese game “Go” – the first computer algorithm ever to do so. As well as being featured in the popular press, the details of this methodology also earned some serious science cred when it formed the basis of a paper in Nature earlier this year.

At the recent American Chemical Society meeting in San Diego, which had a focus on “Computers in Chemistry”, there were several presentations on deep learning and how it can be applied to help solve drug discovery problems. A timely, if rather technical, review of the area was published earlier this year by Gisbert Schneider and co-workers at the ETH in Zurich. In addition, several start-ups have been founded to try to capitalise on the technology and it’s unsurprising that Google is also taking an interest.

So, what kind of problems may deep learning help with in drug discovery? Here are just two examples:

  • Virtual screening – computationally sifting through vast collections of molecules in search of those that might exhibit a particular biological activity.
  • Quantitative structure-activity relationship (QSAR) modelling – trying to establish a mathematical relationship between chemical structures and biological activity data in order to predict the activities of molecules that have not yet been tested experimentally.

These tasks are already tackled by current algorithms, but the hope is that deep learning may lead to significant improvements in accuracy through their ability to model complex, non-linear phenomena.

Deep Thought” was, of course, the computer featured in “The Hitchhiker’s Guide to the Galaxy” and is famous for computing the answer to the meaning of life, the universe and everything (it was 42, in case you hadn’t heard). While it may be some time before deep learning can rival this achievement, it’s certainly making good progress!

How to cite:

Clark, David E., Deep Thought. Eureka blog. Apr 25, 2016. Available: http://eureka.criver.com/deep-thought/