I have a sad but illustrative (I hope) story in the space of oncology which I hope sketches out why biology is such a vast and complex domain. There's an ongoing simmer on scientific social media, sometimes becoming a very frothy boil, on what the impact will be for the next few years of AI on biological research and drug development. AI boosters and AI doomers have often made some very bold claims - with enough compute experimentation and clinical trials will be obsolete or that most major diseases will be cured in 10 years - and there's been plenty of pushback. I count myself with the pushers, though I (like most of them) aren't saying AI won't have an large impact or that we'd like the AI fanatics and their money to just go somewhere else. Heck, I'm in the business of selling services and machinery to them! Yet still I'm accused of being a Luddite ("why are you against AI"). Also in this vein, I want to make clear that the biomedical story below is not meant to illustrate a set of hardcore litmus tests which AI must solve or be declared useless; it is a story which highlights just how complex biomedicine can be and how great our ignorance is.
One recent flashpoint was mathematician Terence Tao attempting to show to biologists what his field is feeling in terms of AI encroachment and how AI might be problematic. The example given was of an AI designing a drug that looked really good in trials and seemed to cure cancer, but had some insidious slow and negative effect not found in the trials. To which we in drug development tended to react with: what else you got? Some proponents of this view tried to elaborate further, but it all looks a bit familiar. I mean, we'd probably goggle a bit if an AI really could deliberately do this because of misaligned incentives, but this is hardly new ground. Not only do most major drugs have side effects in general, but it's not uncommon for them to have very hard to detect ones because only a subset of patients show the effect. And unfortunately there have been more than a few attempts of drug sponsors trying to hack clinical trial design to put lipstick on pharmacologic pigs.
Someone I knew died recently after treatment for lymphoma. The initial cytotoxic regimen of CHOP (cycloheximide, hydroxydaunorubicin, oncovin, prednisone) failed to regress the tumor (and for this patient Rituxan wasn't an option), so a very interesting alternative was tried. This drug's story is full of unusual aspects - it's not a prodrug but a double prodrug, an interesting but someone enigmatic mechanism-of-action, genetically defined patient population, a further genetically defined subpopulation where it is contraindicated, and an unusual development history which included being shelved and then revived. The biology aspects are all both fair game topics for AI to act and also illustrations of how complex drugs can be.
The drug, once undergoing two metabolic transformations, serves as a guanine analog. That's an interesting space for bit-oriented chemists to explore -can you identify compounds that are more potent after metabolic processing than before? Can you turn that into a useful search for prodrugs? And can you factor in whether these conversions will lead to significantly greater variance in efficacy, due to genetic variation in the transformation steps? Or in problematic food or drug interactions, due to the appropriate metabolic enzymes being up- or down-regulated or being inhibited?
A very interesting question to throw AI at - why are lymphocytes more sensitive to guanine analog poisons than most cells? Can you actually compute a therapeutic index for this strategy - or any other one?
The genetic targeting, anti-targeting and development history all converge on a small fact I left out: genetically defined patient population is "Canis familiaris" - the deceased was a dog. Rabacfosaine (brand name Tanovea) was originally developed by Gilead Sciences for human use, but a side effect showed up in early trials. Some patients developed lung fibrosis. This was viewed as too great a liability for development as a human therapeutic, but a small company (VetDC) negotiated the rights for veterinary use. Given the shorter natural lifespan of dogs and different attitudes toward risk in a veterinary indication, VetDC thought they could gain approval - and did. But it turns out that West Highlands White Terriers are particularly prone to lung fibrosis, so the drug is contraindicated for that entire breed.
Suppose you wanted to use AI to solve this riddle of why some patients, both human and canine, develop lung fibrosis and why Westies are particularly prone to fibrosis, what datasets would you expect would crack the puzzle? Fibrosis is a particularly complex condition, involving interactions of fibroblasts with multiple immune system components. And why lung rather than another site, such as liver?
As an aside, as I noted Rituxan wasn't an option - it's a monoclonal antibody and these are notoriously specific for a species given evolutionary changes in cell surface proteins which such therapeutic monoclonals target. Given the power of AI for protein engineering, which seems to be advancing daily, could one redesign Rituxan to bind the homologous canine epitope on CD20? Is there an economically viable path to such a veterinary therapeutic, given that the biological risk is very low? Lymphomas are one of the more prevalent malignancies in dogs, and many dog keepers will go to great lengths for their best friends - for reference there was the story of a wealthy Australian techie developing a patient-specific cancer vaccine for their pup.
In closing, I want to emphasize that I believe AI will have a strong and largely positive impact on drug discovery. AI tools will accelerate many areas of research and development. What I object to are the "AI will solve everything quickly because we can model everything" takes; just the other day I saw a "AI will make clinical trials unnecessary" tweet (I drafted that sentence a month ago & it's true again as I finish this piece today!). New drugs are one of the most powerful means at our disposal to discover completely unknown biology, which means clinical trials will continue to blow up unexpectedly.
No comments:
Post a Comment