“I always pivot to medicine as an example of all the good [artificial intelligence (AI)] can do because almost everything it’s going to do there is going to be good.” —Geoffrey Hinton (Nobel laureate)
As a prominent advocate for the greater adoption of AI across healthcare, renowned cardiologist Eric Topol, MD, opened Danaher’s conference, “AI-Driven Predictive R&D, From Promise to Practice,” with this quote during his keynote presentation.
Danaher’s full-day summit, held on December 3 in San Francisco, brought together an impressive line-up of experts across Big Tech, academia, drug discovery, pharma, and healthcare to discuss the impact of AI to improve human health, from optimizing drug discovery pipelines to improving patient care. The event also positioned the industry to look ahead toward 2025, as the implementation of AI poses ongoing challenges in building ethical guidelines, driving trust and adoption, and more.
Topol is the author of Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again, and the blog Ground Truths, which has nearly 140,000 subscribers. He presented an array of evidence showcasing the benefits of AI across healthcare, including but not limited to increasing diagnostic accuracy, automation to give clinicians the gift of time, and improving doctor-patient relationships by promoting empathy.
“You hear the term ‘precision medicine’ all the time. Well, it’s not very good precision medicine if you keep making the same mistake over and over again. It’s really very precise but we need accuracy in medicine, which we don’t have,” emphasized Topol.
Approximately 800,000 Americans die or are permanently impacted by diagnostic errors each year. Recent proof points show that AI-driven diagnostics can improve breast cancer detection using mammography screening and adenoma detection rate using colonoscopies.
From regulatory hurdles to limitations in transparency, Topol sees ongoing resistance to change and encourages the medical community to embrace AI and its capabilities. “There is a dark side of AI. We don’t have to go over confabulations, hallucinations, and misinformation. These are the things in medicine that are going to hold us back. We need to build trust so we can really get to implementation,” said Topol.
AI awakening
It is no secret that the AI revolution is moving fast. “The world suddenly woke up and said, ‘My goodness, look at this thing! AI is here suddenly!’ Except that it really wasn’t all that sudden,” explained Daphne Koller, PhD, CEO of insitro, during her keynote presentation. “We have actually been living on an exponential curve of technology transformation that has been going on for multiple decades.”
Prior to founding insitro, Koller was a professor in the department of computer science at Stanford University. She is also a co-founder of Coursera, the global open online education platform launched in 2012, which has partnered with some 300 universities and offers more than 7,000 courses across the sciences and humanities.
In an exponential curve, technological advancements only accelerate over time, a key principle that Koller encouraged the community to keep in mind when judging where AI is today and where it will take the industry in the coming years.
Along the same vein, Topol cited the 2022 launch of ChatGPT as the hallmark of a new AI era that leverages large datasets for generalizable applications. In the closing weeks of 2024, at least ten foundation models were published, including Evo for interpreting and generating genomic sequences at vast scale; the Human Cell Atlas, which provides 3D maps of cells in the body to connect genotype to phenotype; and the open source code release of AlphaFold 3, which expands the protein structure prediction algorithm to incorporate broader molecular interactions with ligands and nucleic acids.
AI in practice
“The key to success is not in the AI itself. It is in the bidirectional integration of the AI and digital world with the physical world,” said Martin Stumpe, PhD, chief AI officer at Danaher.
From large chemical data for small-molecule drug discovery to the virtual cell, the flood of data across multiple biological scales, modalities, and clinical contexts has been one of the main transformational drivers both fueling AI advancements and posing new practical considerations.
“If we think about AlphaFold, [protein structure prediction] was a very well-defined problem. How do we define that problem when we think about the virtual cell? What are the capabilities that we want those cells to do? Once we reach alignment, how can we possibly validate all of these different applications?” posed Emma Lundberg, PhD, associate professor at Stanford University and the co-director of the Human Protein Atlas, which is based at the KTH Royal Institute of Technology in Stockholm, Sweden. The program aims to map all the human proteins across cells, tissues, and organs by integrating various omics technologies.
From the academic perspective, Lundberg encouraged the virtual cell community to be strategic about collaboration. “How do we build from each other’s work and create a global collaborative framework? Ensuring that people run in the same direction and create synergy is a big barrier right now,” she said.
Regina Barzilay, PhD, a distinguished professor at Massachusetts Institute of Technology (MIT), studies the deployment of AI algorithms in clinical settings. She argues that many of the safety mechanisms that exist in other industries do not exist in AI, which is particularly troublesome because we’re moving to the next generation of tools where humans cannot validate the predictions.
“AI should be monitoring AI. These types of technologies still need to be developed to ensure that we can do safe care delivery,” said Barzilay.
Moving to drug discovery, the AI data boom has caused a paradigm shift from building on known targets from the literature to massive searches across biology for novel targets, a movement that is touted to improve R&D efficiency.
“You cannot afford not to adopt [AI]. This technology affords so much speed up, that if you ignore it, competitors can completely change the market on you,” said Usama Fayyad, PhD, executive director of the Institute for Experimental AI at Northeastern University.
Last October, Recursion, one of the more established AI in drug discovery companies founded in 2013, announced the investigational new drug (IND) approval of REC-1245, a new chemical entity for the treatment of biomarker-enriched solid tumors and lymphoma. Recursion, whose combination with U.K.-based Exscientia was approved in November 2024, stated that REC-1245 is the first program to use the company’s end-to-end AI platform and progressed from target identification to preclinical candidate in less than 18 months, nearly twice the speed of the industry average.
While proof points are starting to come in for timeline compression in discovery, success rates in the clinic remain at a low 10%. “I worry that we’re becoming more efficient at making medicines that fail in the clinic. The big challenge is still the translation challenge,” said Steve Crossan, founder of Dayhoff Labs and one of the founding DeepMind contributors of AlphaFold 1.
Crossan highlighted prior studies that showed that 70% of drugs in the clinic were working through off-target effects, indicating an ongoing gap in our goal of targeted therapeutic mechanisms.
Najat Khan, PhD, chief R&D officer and chief commercial officer at Recursion, emphasized the role of big data in understanding disease biology. “The challenge is unlike physician notes or the climate where we have a ton of data. We understand 10–15% of biology at best. There’s so much gap,” said Khan. “Think about it as a map with tons of gaps. You have to generate the data to increase the odds of success.”
To improve drug discovery’s promise, Koller leverages support from human genetics to inform drug targets at insitro, noting that a recent study published in Nature documented the probability of success for drug mechanisms with genetic support to be 2.6 times greater than those without.
While drug discovery pipelines continue to advance, Khan urged the community to maintain momentum while managing expectations. “There’s a lot of hype and things that haven’t worked. A lot of people walk away. Is it going to be right the first time? No. If our success rate is 10%, you can’t expect 100%,” said Khan.
“I love what Eric Topol said, which is ‘embrace it!’ Because if you embrace it, then you say [AI] is a ‘must-have.’ There’s no option not to try,” Khan continued.
“When you’re on the exponential part of a technology curve, it’s very easy to forget that the thing that doesn’t work today will work tomorrow—literally tomorrow,” weighed in Crossan.
Taken together, Hinton’s words ring true as AI’s promise continues to provide increasing benefits across healthcare. As technology accelerates, the industry will watch as strategy, expectations, and acceptance evolve to keep up.
