AI in drug discovery

Back in 2011, IBM’s Watson supercomputer won the US TV game show Jeopardy and captured the public’s imagination.

At that time, smartphones had no built-in Siri or Google Assistant and few people would have believed how commonplace devices such as Google Home and Amazon Alexa would become. Watson's win immediately made AI advances tangible for millions of people. 

There are many different types of AI, but the most common is ‘machine learning’. In machine learning, an input is added – for example, an image. A complex function is then applied to connect that image with an observation – for example, ‘cat’. In the case of the example, you would then have the ability to automatically sift through a collection of images to find images of cats.  

AI and machine learning are now used in many applications, from the example of image classification above [3] to autonomous driving. But one area of AI research where there are untapped opportunities is looking at how we can apply it to healthcare and drug discovery.  

Due to an explosion in knowledge about the molecular mechanisms behind different diseases, more promising targets for the design of new drugs are being identified than ever before.  

But the processes behind drug discovery can be challenging. Can AI and machine learning help alleviate some of the traditional bottlenecks so we can get new and more effective drugs to patients, faster?

A central goal of drug discovery research is to identify medicines that act beneficially on the body – in other words, they can help prevent or treat a particular disease. 

Although there are several different types of drugs, many are small chemically synthesized molecules that can specifically bind to a target molecule – usually a protein – involved in a disease. 

To find these molecules, researchers traditionally carry out large screens of libraries of molecules to identify one with the potential to become a drug. They then go through numerous rounds of tests to develop this into a promising compound.

Recently, more rational structure-based drug design approaches are becoming increasingly common. These avoid the initial screening stages but still require chemists to create potential new drugs by designing, synthesizing and evaluating numerous compounds.

Because it’s generally unknown which chemical structures will have both the desired biological effects and the properties needed to become an effective drug, the process of refining a promising compound into a drug candidate can be both expensive and time-consuming. Latest figures show that the cost of bringing a new drug to market now averages US$2.6bn (2013 dollars).[1]

In addition, even once a new drug candidate shows potential in laboratory testing, it may still fail when it is moved into clinical trials. In fact, only about 10% of drug candidates evaluated in phase I made it to market.[2]  

Considering this, it’s not surprising that experts are now looking to the unparalleled data processing potential of AI systems as a way to accelerate and reduce the cost of discovering new drugs. According to market research firm Bekryl, AI has the potential to offer over US$70 billion in savings for the drug discovery process by 2028.[4] 

The sheer size of the libraries used to screen for new drug candidates means it’s now practically impossible for individual researchers to review everything themselves - and that’s where AI and machine learning can help.

These sophisticated techniques allow researchers to extract hidden insights from huge datasets. The benefits of doing this are numerous: 

• Predicting the properties of a potential compound, meaning that only compounds with desired properties are chosen for synthesis – saving time and money by preventing work on compounds that are unlikely to be effective.  
• Generating ideas for entirely novel compounds, where the ‘invented’ molecule is predicted to have all the desired properties required for success – which could hugely accelerate the discovery of effective new drugs.
• Alleviating the need for repetitive tasks, such as the manual analysis of thousands of histology images – saving hundreds of person-hours in the laboratory.These are just a few of the potential advantages, looking at the early end of the drug discovery pipeline.

These are just a few of the potential advantages, looking at the early end of the drug discovery pipeline.

To understand how AI is being applied in practice, let’s take a look at some of the innovative research we currently have underway. 

“Predictive models are central to our work,” says Friedrich Rippmann, Director, Computational Chemistry & Biology at Merck KGaA, Darmstadt, Germany. “These are statistical models that predict whether a compound idea – a not-yet-synthesized molecule – will produce a desired activity.”  

“The technologies we’re using mostly relate to machine learning. In particular, we’re using various types of deep neural networks. But we’ve also explored other more classical statistical techniques, with funny names like random forests and support vector machines.”  

Deep neural networks started to become particularly popular around 2012, when researchers from the University of Toronto won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC).[1] 

In recent years, this subset of machine learning techniques has revolutionized several fields such as computer vision, natural language processing, and game playing.[5] And now it’s showing exciting promise in identifying novel compounds.

 “We’ve driving innovative research in the area of interpretable deep neural networks,” continues Rippmann. 

 “Our predictions no longer only tell us, ‘yes, this will work’ or ‘no, it won’t work’. Now they can also tell us why they will or won’t work, by highlighting the areas in a molecule that are responsible for a certain activity. This gives an immediate indication to chemists of what to do to remove a certain unwanted activity.”

 Their findings are described in greater detail in a recent, well-received article, Interpretable Deep Learning in Drug Discovery.[6]

So far, in terms of practical benefit, the research carried out by Friedrich Rippmann’s team and the company’s partners has made available approximately 300 new models for assessing the properties of a compound, which can help predict their ability to bind to a specific disease-relevant target. 

 “These models are already being used by our chemists to judge their compound ideas before deciding on whether to synthesize them,” adds Rippmann.  

Making advances in this field is, by nature, a collaborative process. 

 “We rarely develop novel methods entirely by ourselves,” says Rippmann. “All our work is done in collaboration with leading academic groups and start-up companies.”

So far, notable collaborations include working with Professor Sepp Hochreiter at Universität Linz in Austria – who leads one of the top academic groups in this research area. 

In addition, we’re also collaborating with Cyclica in Toronto, using their predictive technology that identifies which human proteins a chemical compound is likely to target (for example, in the deconvolution of phenotypic screens).  

And separately, we’re working with Iktos, a start-up based in Paris, using their technology to automatically ‘invent’ novel molecules that hit predefined requirements for a drug, including binding to the drug target, lack of specific toxicities, solubility, and much more. 

But collaboration to enhance AI in drug discovery is also happening at a larger scale.

AI analysis is only as good as the quality of the datasets in use. For this reason, the pharmaceutical industry is increasingly seeking to collaborate by pooling data.  

A recent initiative to facilitate, the MELLODDY Project, involves the EU Innovative Medicines Initiative and no fewer than ten pharmaceutical companies – including Merck KGaA, Darmstadt, Germany. It aims at improving predictive models through so-called ‘federated learning’, by using a novel blockchain system to store data on a secure ledger while protecting the trade secrets of individual companies.[7]

The biggest opportunity for the future, according to Friedrich Rippmann, is the acceleration of drug discovery and reduction of attrition rates, ultimately making more novel drugs available to patients, faster.

“There are so many opportunities to apply AI in drug discovery,” he says. “But at the moment we’re hindered at times by the prohibitive costs involved. As more competition emerges though, we will see costs coming down – opening up exciting possibilities for new discoveries in diverse fields.”

_{[1] https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
[2] https://www.sciencedirect.com/science/article/abs/pii/S0167629616000291?via%3Dihub
[3] https://www.bio.org/sites/default/files/legacy/bioorg/docs/Clinical%20Development%20Success%20Rates%202006-2015%20-%20BIO,%20Biomedtracker,%20Amplion%202016.pdf
[4] https://bekryl.com/industry-trends/ai-artificial-intelligence-in-drug-discoverymarket-size-analysis
[5] https://www.theverge.com/2019/3/6/18222203/video-game-ai-future-procedural-generation-deep-learning
[6] https://link.springer.com/chapter/10.1007/978-3-030-28954-6_18
[7] https://www.imi.europa.eu/projects-results/project-factsheets/melloddy}