First AId
Drug and medical device manufacturers are investing billions in AI, but they may be standing in the way of their own success
AI is ballooning in the healthcare industry like a colossal aortic aneurism. Virtually every major pharmaceutical and medical device manufacturer is expanding AI programs and partnerships, an investment that conservatively is expected to grow from $5 billion last year to $20 billion by 2035.
Pharmaceutical giants like Sanofi, Pfizer, and Merck foresee a new age in which generative AI models trained on biomolecular data will crank out blueprints for new medicines to combat cancer, heart disease, and previously “undruggable” illnesses. Device makers, including GE Healthcare, Siemens, and Medtronic, envision AI-enhanced medical imaging with vastly improved diagnostic accuracy, and AI-trained surgical robots that will one day be capable of performing procedures with the same skill as a human doctor.
Of course, the manufacturers aren’t doing this for their health. They expect their massive investment in AI to make them a ton of money. How much? The international finance firm Morgan Stanley is bullish on medical AI, telling investors that even “modest improvements in drug development enabled by the use of artificial intelligence” could result in an additional 50 novel therapies over a 10-year period, representing a greater than $50 billion opportunity. Other analysts project that AI applications have the potential to create 10 times that amount in annual value.
No wonder healthcare is betting big on AI. But how well is their investment paying off?
In a recent review of progress in the field, JAMA AI editors Roy Perlis and Yulin Hswen pointed out several credible achievements. For example, AI is really good at condensing stacks of information about patients’ diagnoses, treatments, and procedures into concise hospital discharge summaries. Occasional errors do occur and human review is still required (as anyone knows who saw Dr. Al’s demonstration on the Pitt last week), but doctors are more than delighted to let AI do the tedious task.
AI has also excelled at screening patients for clinical trials of promising, new treatments; in one study, nearly doubling the rate of eligibility determination and enrollment.
AI-enhanced radiography systems trained on millions of images and videos have outperformed human technicians in picking up lesions on echocardiograms. And an AI-trained CT-scan program developed by the Chinese tech giant Alibaba has proven superior to the human eye at detecting early-stage pancreatic cancer when the deadly disease is still treatable.
What AI is less good at is drug discovery.
A Drug Discovery Today report asserted that 20 of 24 AI-devised therapies were successful in preliminary Phase I clinical trials last year: a high success rate but inconclusive based on the small sample size. In Phase II safety and efficacy studies, four of 10 drugs developed with AI succeeded, but that record is no better than traditional, non-AI compounds.
The AI drug that’s closest to approval is Insilico Medicine’s rentoserib for the treatment of idiopathic pulmonary fibrosis (IPF), a rapidly progressing, inevitably fatal respiratory disease. In clinical trials, rentoserib demonstrated that it can slow progression and improve breathing capacity in patients with IPF but showed no advantage over currently available treatments, none of which are able to halt or reverse the disease.
Doubtless, progress will accelerate in time. But with only modest results so far, the problem of achieving clinical efficacy remains a challenge for AI-designed drugs, and it seems like Big Pharma’s zeal for “Faster, Better” may prove more elusive than industry honchos have hoped.
The problem is partly of their own making. Because to generate new treatments, AI needs to scrub data about past failures as well as successes. However, pharmaceutical manufacturers assiduously guard preclinical and clinical data of drug candidates that don’t succeed, and refuse to share it with rival companies. As a result, unlike other areas of research and development, only a small fraction of drug discovery data is available for training AI machine learning models.
Ironically, this repository of negative data may hold the key to unlocking the AI-driven miracle drugs of the future.
Great column, Jeff! Very interesting. Since no one has any idea what the actual effect AI is going to have on the world, information and research like this is great. Keep up the good work!
This is razor sharp analysis. The irony of pharma companies hoarding their failure data while simultanously investing billions in AI is exactly the kind of self-sabotage nobody talks about. Back when I worked adjacent to biotech, this proprietary mindset was like religion and its wild to see how directly it contradicts the entire premise of machine lerning. The failures probably contain way more signal than the successes anyway.