AI is Not Going to Revolutionize Drug Discovery

It’s just another technology

Atorvastatin (stick figure) bound to its target, HMG CoA-reductase (blob). By A2–33, CC BY-SA 3.0,

The pharma industry is in trouble. Although still immensely profitable, it is on an unsustainable trajectory. Simply put, the cost to develop a new drug rises exponentially, while revenues from drugs rise linearly. Absent a revolution, pharma is doomed. Can AI save it?

Eroom’s law (Moore’s law in reverse) in action. US pharma R&D expenditures per approved drug. Own work, data from Diagnosing the decline in pharmaceutical R&D efficiency
US drug sales. Own work, data from Medicines Use and Spending in the U.S.

George Self recently wrote in TDS about “How Artificial Intelligence Is Accelerating Drug Discovery”. It’s a good article and you should read it. I don’t disagree much with any of the arguments, which boil down to this: AI will enable drug companies to do what they are doing now, only faster, cheaper and more effectively.

What do drug companies do? They find molecules that bind to other molecules. Pharma companies find small molecules that bind to large molecules (see the figure at the top). Biotechs find large molecules that also bind to large molecules. AI will indeed help them do this (pharmas much more than biotechs).

That won’t be nearly enough to save them.

Although finding drug candidates is inefficient and costly, it doesn’t drive the exponential rise in costs for drug discovery and approval. Look at the x-axis of the first graph. It goes back to 1950 in a smooth, unbroken trend. In 1950 we weren’t even sure that DNA was the material of heredity. We couldn’t sequence proteins, and had only begun solving their structures. In fact, it was not even clear then that they had stable structures.

Schematic diagram of the structure of myoglobin. It was the first protein structure to be solved. That happened in 1958, when it cost about $40M to bring a new drug to market. AzaToth [Public domain]

I won’t waste your time with a litany of all the tech advances that have repeatedly revolutionized life science in the last seven decades. Any reader of TDS knows that they are numerous and substantial.

Drug discovery today is orders of magnitude more powerful and efficient than it was one or two or three etc decades ago. Yet drug discovery costs keep going up 9% per year, compounded annually. Search efficiency for new drug candidates cannot possibly drive drug discovery costs; if it did, these costs would decrease exponentially.

OK, so what is driving drug discovery costs?

Drug targets — more specifically, the lack thereof. These are the large molecules that drugs bind to. Nearly all of them are proteins, which are of course encoded by genes. There aren’t that many human drug targets (bacterial and viral targets are a different story). The human genome encodes only 19,000 proteins. About a tenth of these are implicated in disease and are plausible drug targets. About half have already been “drugged”.

The ones that got drugged first are the ones that have the strongest therapeutic effects; they are the easiest to identify. In a sense, they are like the nuggets in a mine. The biggest ones get picked first, with little effort. Over time most nuggets get found, even if they are not found efficiently. All that’s left are specks of dust, and you have to sift through tons of dirt to find them.

Already, the majority of new-drug approvals are for orphan indications: those that affect less than 200K patients in the US. Many are ultra-orphans, affecting less than 10K patients. We’ve reached the endpoint of rare diseases, where a treatment was devised for a single patient.

Indeed, one of the first claims of success for AI (grossly overhyped by Deep Genomics) is a treatment for Wilson’s Disease, which affects maybe 10K people in the US. The AI-developed treatment is suitable for 1 in 50 of these patients (if it works — it has not entered clinical trials). That works out to helping 1 in every 1.5M people in the best-case scenario.

Maybe AI can crack some longstanding conundrums in drug development. A non-toxic gain-of-function ligand for the tumor suppressor protein p53, often described as “undruggable”, would be a big win. There are a few blockbuster opportunities like this, but just a few.

A recent computational approach to drug discovery. From PathFX provides mechanistic insights into drug efficacy and safety for regulatory review and therapeutic development. CC BY license.

I expect AI to mostly impact niche conditions like Wilson’s Disease. That will be genuine creation of value. But not a revolution. Not in public health, and not in the financial health of drug companies.

We’ve had seven decades in which Eroom’s Law has described the trajectory of drug discovery costs. Those seven decades saw more revolutionary advances in bioscience than in all previous history combined. They saw multiple evolutions in regulatory regimes. They saw any number of improvements in management best practices. None of those things nudged Eroom’s, not even a little bit. Maybe AI is the technology that will succeed where all others have failed. But I sure wouldn’t bet on it.

Disclosure: I have held stock in Gilead since they bought out my employer (NeXstar Pharma) in 1999, and in SomaLogic since I helped found it in 2000.

Source link
Back to top button
Thanks !

Thanks for sharing this, you are awesome !

Pin It on Pinterest

Share This

Share this post with your friends!