When biology met the machine: A (complicated) love story

Do we understand enough about biology for sci-fi seeming tech to truly make a difference in healthcare? Experts have divergent, and nuanced, thoughts.

Note: A sincere apology to readers for Phase 5’s (even-longer-than-expected, and certainly longer-than-welcome) hiatus. I have been out sick, but the personal medical issues, thankfully, are receding into the rearview, and I’m glad to say we’re back to regularly scheduled business. Plenty to catch up on—starting now. I’m grateful for your patience these past few weeks and your readership always.

The roar over artificial intelligence and machine learning’s role in medicine during last month’s JPMorgan Healthcare Conference (aka JPM2024, the yearly health industry extravaganza in San Francisco) rang in two distinct timbres: The hype of the optimists with visions of algorithms that can unlock the secrets of our most foundational biology to spur cures for all sorts of diseases—an outlook buttressed by a few multi-billion dollar deal announcements during JPM—and the responding grunts from those urging the industry to curb their enthusiasm.

What complicates that tug-of-war is the reality that the battle line isn’t etched in red. In fact, it’s not even so much an adversarial rivalry as it is a co-dependent relationship—a push and pull between two interlinked epistemologies that spans both the building blocks of proteins and the intricacies of the evolving AI- and ML-driven codes we use to parse them. Put another way: As with so many love stories… It’s complicated. It can also lead to some tough conversations on how we ought to think about the underlying dynamics of the relationship.

“I think the crowd saying AI and these technologies are going to change the game for drug discovery any time soon are in for a rude awakening,” said Bryan Roberts, partner at VC firm Venrock, during an on-the-record dinner on January 8, the first full day of JPM week. “We just don’t know enough about biology for it to have much more than a marginal effect at those stages right now, maybe 3% to 4% new hits on certain biological targets.” (This was a view echoed by several other veteran VCs in the field who I spoke with during JPM who preferred anonymity.)

Roberts was referring to a foundational element of how we make new drugs, and the infatuation over emerging technologies is understandable given that so, so many experimental medicines ultimately fail.

Source: Duxin Sun et. al, Acta Pharmaceutica Sinica B

90% is a commonly cited failure rate in clinical drug development, and the failures typically mount as that new hope for tackling Alzheimer’s, or a deadly cancer, or heart disease progresses through the clinical pipeline from initial promise in discovery to the more hard-nosed stuff of actual impact on the clinical trial participants who suffer from those diseases. The science and business of drug-making is an inherently fraught endeavor even with the best intentions.

So if AI, ML, or related technologies can uncover a key biological marker at those early discovery stages to maximize the chance of success as the going gets tough (and expensive) in later-stage trials on real humans—well, that’s where the promise lies, the thinking goes.

One day before that Venrock dinner in early January, Google parent Alphabet’s AI venture DeepMind announced a deal that could be worth as much as $3 billion between its Isomorphic drug discovery arm, which has an AI-driven platform called AlphaFold, and drug giants Eli Lilly and Novartis.

Simply put, AlphaFold is trying to map the nuances of protein folding through its tech platform to fully understand these proteins’ structures and, critically, the types of molecules that could fit into their nooks and crannies. It’s the essence of creating a new treatment: Figure out what fits into a bit of biological material to induce a certain effect and, hopefully, staunch or cure the suffering that comes with illnesses wrought by these biological kinks down the line.

But even that staggering technology, which blends molecular screening and chemistry and biology, is a nascent step. Fine, you’ve crafted a theoretical compound that can hit a very specific target—now, how do you get that compound to that space? And what’s to say there won’t be unintended consequences down the line no matter how sophisticated and impressive the technological methods?

This is where the tug-of-war in that complex biology-technology relationship comes into play. “There's a quote from Vinod Khosla from like 10 years ago, something to the effect that data science is going to transform medicine more than all of biology knowledge combined, or all of science combined. And I thought that was such an absolutely 180 degree sort of view of what the reality is,” said Serge Saxonov, CEO and co-founder of 10x Genomics, during an interview at JPM a few days later.

Saxonov, as a digital health and genomics entrepreneur and visionary, is no more a pessimist about emerging technologies’ promise in drug development and the future of healthcare than is Roberts. 10X Genomics, after all, is a company predicated on building our understanding of biology and the fundamentals of what happens in our cells through the use of increasingly sophisticated technologies.

But he returned time and time again to the necessity of understanding the relationships between and limitations of these intertwined fields—the work of molding the thrill of discovery and invention into a form that delivers real human impact.

“No matter how much you you can put together all the EMR data flows and all the rest of that together, and try to find correlations, if you're not measuring the fundamental systems that are actually driving disease and driving health, you’re just chasing statistical correlations,” he said.

“So I think there's a there's an under-appreciation of how much more biology we need to understand. And to some extent you can argue that probably has been a consistent story, where we have overestimated how much we really knew.”