AI in Drug Repurposing & Design

Much of the traditional process of discovering new drugs is trial and error. Until relatively recently, the primary advances to improve the expected outcome of this costly educated guesswork have been those that allowed biotech and pharma to “guess” more and more (i.e., physically testing more and more compounds) with the hopes of identifying more drug candidates or “hits” worthy of moving on to the next phase of drug development. This methodology is often called “high throughput testing” since numbers win in wars of attrition; the more candidates you start with, the more likely it is that one or more will withstand the high failure rate imposed by the multiple stages of drug development.

This is because, until recently, the effect of an experimental drug remained largely unknown until the drug was tested in vivo or “in the body.” However, a new wave of drug development platforms enabled by artificial intelligence and machine learning ("AI" hereafter) is helping academics, biotech, and big pharma use data from previous experiments to identify compatible therapeutics or even design new ones entirely. Although the earliest clinical trials are still underway to evaluate how these methods stack up to traditional pipelines in cost, time, safety, etc., discussing how the sausage is made ahead of the headlines may be worthwhile. In this case, the best parallels for how AI drug repurposing and discovery work may be drawn from the historical advances in chess.

First, IBM’s Deep Blue relied on human experience, knowledge, and strategy. It “learned” chess by observing previous matches (what worked and what didn’t) and leveraged these to maximize its proportion of wins to losses. Its moves were considered strategically sound, blending the techniques of the greatest grandmasters.

In late 2017, Google's AlphaZero decisively outperformed Deep Blue derivative Stockfish in 1000 games (155 wins, 6 losses, 839 ties). What made it so dominant? Well, unlike Deep Blue, AlphaZero wasn’t trained on any prior games or strategies at all. It was only programmed with the rules of chess and allowed to build its own logic from the ground up about how to win by playing itself over and over again. This resulted in unorthodox moves with a presumable long-term positional pay-off that even chess grandmasters couldn’t have considered in the moment or, in many cases, couldn’t explain afterward.

In many ways, the critical differences between Deep Blue and AlphaZero illustrate the key differences between drug repurposing and drug design. Drug repurposing synthesizes the data from countless clinical and pre-clinical experiments and published human knowledge to draw connections between fields and employ the best drug suited to treat a disease. In a way, these methods are based on previous knowledge, making them lower risk – lower reward because they are grounded yet limited by previous knowledge.

On the other hand, de novo (starting from nothing) drug design AI performs more similarly to ZeroAlpha, programmed with the core tenants of human biology or particular diseases upon which the AI builds its own logic to construct a molecule with a specific desired effect – potentially through mechanisms or cascade of mechanisms we don’t fully understand. While this method may be more revolutionary, the rules of biology are far from wholly characterized, and they are in board games like chess, making it more high risk - high reward.

Put another way, one could learn to play a sport, say basketball or golf, by watching hundreds of games/matches or just being given a list of rules. However, the person brought up on a list of rules may develop a style of play outside the norm, like a Granny Shot or Happy Gilmore swing , which may revolutionize the game or provoke unintended consequences.

Science needs both – incremental and breakthrough progress. While it is likely that both methodologies will serve academics, biotech, and big pharma in bringing new or repurposed drugs to market faster, cheaper, and safer, only time will tell if these technologies live up to the hype. As to be expected, using AI to repurpose existing compounds to target other diseases versus developing entirely new drugs carries very different regulatory and intellectual property implications as well, but that's a topic for another time.