AI in Drug Discovery — Part 1

Jonathan Hay
Nov 3 · 5 min read

Making sense of the crowded landscape

A lot of people have been talking about the Insilico Medicines paper published in Nature Biotechnology in September this year. The claim that a drug was developed in 21 days is the kind of thing that gets people’s attention and strikes FOMO into the hearts of pharma executives who know that AI is coming but are behind in understanding and adoption of new technologies. Andreas Bender, Ash Jogalekar and Derek Lowe all cut through the hype and analysed the paper. Bottomline: A nice piece of work showing some incremental improvements but not a massive breakthrough. So what would a massive breakthrough look like? Where are the real game changing applications of AI in drug discovery likely to come from? There are a lot of approaches being tried. Here is an interesting and regularly updated list of 158[1] start-ups using various forms of AI in drug discovery and related areas. We just led an investment in one of them: Turbine.ai. Where does Turbine.ai fit into that, and why are we so excited about what they are doing? I will answer these questions in two separate blogs. Before I introduce Turbine, let’s try to make sense of the evolving landscape of AI-powered drug discovery.

At a high level we might divide AI technologies in drug discovery into 1) those that speed up and/or improve the design of drugs/molecules, 2) technologies that generate new biological targets, and 3) technologies that define the right patient populations. More and more it is tasks 2) and especially 3) that can add the most value to a drug discovery program.[2]

There are of course many other applications of AI in drug discovery and related areas (e.g. tools and infrastructure companies like LifeBit, SevenBridges, DNAnexus and Grakn.ai). There are also companies that fit somewhere in between molecular design and target discovery e.g. those companies that are repurposing existing drugs for new disease areas. Healx is an example of a company in this area. But at the end of the day drug discovery is mainly about selecting the right biological target (usually a protein), designing a molecule to bind to that target, and then, if all the preclinical work validates the biology, getting it into exactly the right patient population i.e. the patients who are most likely to respond.

The really game changing applications in drug discovery are likely to involve whole new work flows where AI is integrated into all aspects of the drug discovery pipeline i.e. from generating new target ideas to designing molecules that are selective for those targets and finally to identifying patients who have the right “omic” profile to respond to the drug. Of course as Bruce Booth pointed out in one of his great blogs there is no getting around the hard work of biological validation, testing for toxicity etc. but if this work is guided by the right kind of AI it may be possible to speed things up. For example, if predictions about biology have a track record of being right why not skip cell lines (which usually don’t translate well anyway) and go directly to animal models? Also, if a prediction is precise in terms of mechanism of action and related biomarkers, it might be possible reduce the number of animals that are used. If the right patients are selected and a higher percentage of patients are responding, then smaller trials are needed to show efficacy. This is the case for example with gene edited cell therapies that have shown miraculous effect.

There are a lot of AI companies big and small that are focused on the first task of designing molecules. The goal is to reduce the enormous cost of manufacturing large chemical libraries needed to carry out high throughput screens and look for “hits” i.e. evidence that one of the molecules binds to the target of interest. Insilico or, as it is often called, computer-aided drug design (CADD) has a long history. Well known and successful companies like Nimbus Therapeutics (co-founded by Atlas and the founder of Schrodinger the dominant chem software platform in the industry) have been doing this for ages. Many compounds that were discovered and/or optimized with CADD have gained FDA approval.[3] The current wave of AI has generated new contenders such as Insilico Medicine, Atomwise and Exscentia that are pursuing this approach, as well as many smaller challengers like GTN, Antiverse, ProteinQure, and others.

A limitation of current technologies is that you still generally need a known physical structure as a starting point e.g. a crystal structure (or strong homology model) of the targets and/or ligands. Current technology does not work well, for example, with the many important proteins that do not have such structures, so called intrinsically disordered proteins or IDPs (see my blog). Progress is being made but it is hard problem to solve. To target IDPs it is likely that novel physical screens, rather than AI, will be important for a long time to come. Companies like Kronos Bio and Phoremost are tackling this space in different ways.

But so long as a good structural starting point exists there are many effective insilco approaches, supported by sophisticated software for designing molecules. The problem isn’t solved but there are lots of fantastic companies in the space — old and new — and one wonders if further developments are likely to be more incremental than revolutionary. There may be revolutionary opportunities still in the area of disordered proteins (perhaps quantum computers could help model these) and also in the area of generating greater chemical diversity — — as one frequent criticism is that the vast chemical libraries of pharma are biased towards certain areas of chemical space and that we need new technologies that can generate more chemical diversity. Developing drugs with more diversity including complicated 3-D shapes that are highly specific to their targets and, thereby, minimize off target effects is a big priority.

But what about the discovery of new biological targets and biomarkers? There are lots of companies doing wet biology to find new targets. For example, they might be knocking out genes with Crispr/Cas9, using RNA or working directly at the level of proteins. Companies like Repare Therapeutics (using Crispr/Cas9 screens), Moderna (using RNA) and Phoremost (using protein libraries) are examples of companies doing wet biology to look for new targets.

While insilico approaches to molecular design are well validated and used routinely, insilico approaches to target and biomarker discovery are still emerging. Examples of companies working in this space are Benevolent.ai in London, Row Analytics in Oxford and, of course, Turbine.ai. The company I will introduce in more detail in my next post.

[1] On October 10, 2019. Simon Smith updates this list from time to time so I expect it will continue to grow.

[2] For example, a recent report by Deloitte on return on capital (ROC) in health care/pharma predicted that the highest returns would accrue to businesses that learn to mine data to delivery personalized medical solutions. This may be the single most important ability that pharma needs to acquire to improve efficacy in clinical trials.

[3] See e.g. https://www.ncbi.nlm.nih.gov/pubmed/19929824

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade