Machines Learn to Discover Drugs and Medicines

     

Will your new chemistry lab partner be made of circuits?

Experimental sciences without experiments, lab geeks without labs, chemical reactions without chemicals. Is this science apocalypse or a godsend? In fact, it’s the latter. Only, the manna doesn’t come from heaven but from a computer, thanks to A.I. But let’s take a step back.

In two previous blog posts, we discussed how machine learning is a powerful ally in pesticide development and in predicting the environmental impact of chemicals. However, in silico testing can be applied to other chemical studies, too, such as drug development in the pharmaceutical industry: if you can investigate the biocidal properties of a chemical, then you can assess its curative properties, too.

If you want to know other ways machine learning can be applied in medicine,  download our whitepaper on the topic

Were you shocked by the amount of resources - in terms of time and money - invested in plant, crop, or wood preservatives development? You’d be surprised to know that even more resources are used in the development of novel therapies: $1 billion or more, and over 10-15 years for development, testing and clinical trials. Besides, all these efforts are not a guarantee of success, as the chance to fail is rather high: in the US, only 1 in every 5,000-10,000 potential compounds investigated is granted FDA approval.

How come we never tried to make the process more efficient? Actually, we did.

A.I. and Drug Development: About to Reach their Golden Wedding Anniversary

The relationship between drug discovery and artificial intelligence is not a publicity stunt. In fact, their rapport dates back to 1973, when artificial neural networks were used to determine whether a group of organic compounds were physiologically active or not, i.e. whether they had an effect on living organisms. Despite this initial success, computer tools were still (and remained so for decades) too inaccurate for routine chemical predictions, and laboratory experiments were the one and only source of information. But this didn’t discourage researchers from both fields, which kept improving the in silico methods.

In 2012, Merk sponsored a Kaggle competition to “help develop safe and effective medicines by predicting molecular activity” with machine learning. Three years later, the first deep convolutional neural network was applied to predict molecular binding, i.e. whether two molecules would get together or not. The software was AtomNet, and the company that created it was none other than AtomWise, which partnered up with Monsanto. In addition, in only 2 years, AtomWise launched 27 drug discovery projects on a series of diseases, and ready-to-test treatments for Ebola and multiple sclerosis are already available.

Likewise, GSK recently funded Exscientia to use A.I. to identify molecules that can treat up to 10 diseases. The result may not be a cure-all, but it is close enough.

As you can see, this doesn’t mean lab work will cease to exist, but experimental workload and failed experiments will certainly decrease, thanks to smarter in silico testing. Besides saving resources, A.I.-based testing will reduce the amount of time required to develop an effective treatment. One that can save or improve people’s lives. A small step for a machine, one giant leap for mankind.

Download our latest whitepaper on Medical Imaging and A.I.

About The Author