The use of artificial intelligence in the drug discovery process is not new, but it is advancing in significant ways. Several weeks ago the BBC announced the first AI developed drug to be taken to human trials. Now they are announcing the discovery of a new antibiotic using AI. Let’s talk about drug development to see how advances in AI are impacting this process.

Finding a drug that is useful medically is tricky, because it has to have a lot of properties simultaneously, and any one property can be a deal-breaker. A useful drug needs to get into the body, get to the target tissue, survive long enough to have the desired effect, it needs to have a desired effect at a dose that is lower than doses that cause significant side effects, and it needs to lack significant toxicity, such as liver or kidney damage. Will the compound be stable on the shelf? The same needs to be true, at least in lack of side effects and toxicity, for all the metabolites of the drug that may be created before everything is eliminated. On top of that we have to worry about drug-drug interactions, and even interactions with certain foods.

For this reason there is no perfect drug. Every pharmaceutical is a trade-off. Being “natural” is also not a magic wand that bypasses all these concerns. Substances that occur in nature did not evolve for our benefit. They generally evolved to be poisons to creatures that might eat them, including us. Drugs derived from plants are basically poisons that we have purified, usually altered, and then discovered a dose range that can be safely exploited.

The “old” method of drug discovery was lots of trial and error. Candidate drugs were tested in petri dishes to see their effects on cells in culture. They were then tested in animals, and if everything looked good eventually in humans. Chemists understand how chemical structure is likely to affect a drug’s properties, so they were not guessing in the dark. There is sometimes also a history of traditional use to give some anecdotal information about potential effects.

Computer aided drug-design (CADD) has long been helping in this process. They use a large database of known drugs to help anticipate how a certain chemical structure is likely to look three-dimensionally, and the effects this will have on its properties. Will it bind to the target of interest (called Ligand Based Drug Design)? Will it be eliminated by the kidneys, etc.?

But now we are moving into a new phase, going beyond existing CADD. Drug development is starting to benefit from the newest developments in AI, including deep learning. In the first example above:

The molecule – known as DSP-1181 – was created by using algorithms that sifted through potential compounds, checking them against a huge database of parameters.

“There are billions of decisions needed to find the right molecules and it is a huge decision to precisely engineer a drug,” said Prof Hopkins.

The company reports that a process that normally takes 5 years was reduced to 1 year using computer learning algorithms. The drug still needs to be tested in humans, which will take years, and will still need to be the case for the foreseeable future. However, AI-based CADD can be used to get drugs to human trials much quicker and cheaper, and maximize their probability of success. One of the reasons drugs can be so expensive is because a company can spend $100 million developing a drug, only to have it fail in the final stages of human trials.

This first example represents, according to reports, a combination of structure-based drug design and ligand based drug design – modeling potential molecules and predicting using deep learning what structure and properties the molecule will have, in addition to what affinity it will have for the target. In this case the process was used to find a drug that might be useful in treating obsessive compulsive disorder (OCD).

In the second example a different process was used. Deep learning AI was not used to design a chemical, but to screen existing molecule. The BBC reports:

A powerful algorithm was used to analyse more than one hundred million chemical compounds in a matter of days.

This included 2500 known drugs.  The researchers then selected the 100 drugs that the AI predicted would have the best properties for physical testing. The resulting winner, the researchers claim, can kill 35 different types of infectious bacteria. Again, this drug still needs to go through clinical trials. In both cases it remains to be seen if the process will result in an approved effective drug. But either way the potential of this process, using powerful AI in the pre-clinical phase of drug discovery, is very powerful.

By one estimate there are already 203 startups using AI in drug development. It will take 10 years or so to really see if AI will result in a new age of drug development, but it looks extremely promising. This seems like more than an incremental advance, and rather is a serious game-changer. The potential benefits are – dramatically reducing the costs of drug development, reducing the time of drug development (by years), and optimizing the chance of success of clinical trials. Hopefully this will reduce the cost of new drugs, and also provide more and better drugs faster, which has the potential of reducing the costs of health care.

AI also has the potential to radically improve other aspects of medicine as well, such as interpreting diagnostic imaging. This has the potential to reduce medical errors, and also chart more optimal treatment pathways, further reducing the cost of health care. Right now the cost of health care is rising, mainly driven by technological advances. This is becoming unsustainable. We need to leverage technology not only to make health care better but most cost effective. AI is one tool that can help us do that.