Why the Experiment Still Matters More Than the Model
Previously, in Part 1 (skip ahead if you’ve just read it)
Any piece of causal human biology sits in one of three tiers: known (established in the published literature), hidden (latent in data we’ve already generated but not yet extracted), or unknown (not yet revealed by any data, because the experiments haven’t been run).
Which tier applies depends on the question. For drug discovery, there are three: the disease-level question (is this gene causal, and in which direction?), the mechanistic-driver question (through what pathway, in which tissue?), and the therapeutic-intervention question (can it be safely modulated for clinical benefit?).
Two lessons followed. Lesson #1: the answer depends on the application. Judged by clinical trial outcomes, we know roughly 10% of the causal biology relevant to drug discovery. Lesson #2: it varies sharply by disease area, from lipid disorders, where much is known, to psychiatric disease, where most is not.
One gene, FNIP1, stood out in a million-person study of lipid metabolism, with an effect size well beyond the rest. Part 2 starts there.
FNIP1, revisited: one gene, three questions, three different tiers
FNIP1 was one of 59 genes identified in the Regeneron Genetics Center’s exome-wide association study looking at fat-storage capacity and it is a useful test case because its three questions land in three different tiers.…
Read full article...
Part 1: What human genetics of lipids and psychiatric disease teach us about the answer to: What fraction of causal human biology do we actually know?
I have been asking a version of this question for months now – at lunch, over dinner, on walks with friends and colleagues. For fun, I have gotten into the habit of politely pressuring people for a single number. Come on, just give me your best guess!
Almost without exception, as we worked through the answers together, we did not land on a single number. Instead, we debated a structure. This blog is an attempt to write that structure down.
Every one of those conversations eventually ran into the same unexamined assumption behind most discussions of AI and drug discovery: that we already understand causal human biology well enough for a sufficiently capable model to find the cures hiding inside it.
If that assumption is right, as many with computational or technology backgrounds seem to believe, the job left for AI is largely one of connection and interpretation. Compress the literature, integrate the omics, surface the pattern a human missed.
If it is wrong, as many with biological or medical backgrounds seem to believe, if what we are actually missing is not connective tissue between known facts but the causal facts themselves, then AI’s job is a different and much harder one.…
Read full article...
When I think about our vision for AI and automation at Bristol Myers Squibb, I think about the music video by the band OK Go for their song “This Too Shall Pass.”
Stick with me – I promise this will make sense.
The video begins with the band’s bassist donning a pair of protective goggles and then rolling a toy truck into a line of upright dominos.
What follows is one of the most complex – and spectacular – Rube Goldberg machines I’ve ever seen. The dominoes cascade into a tethered string that propels a Hot Wheels car down a ramp, which knocks a billiard ball down another ramp, which knocks over a book connected to a string, and by then, the system is off to the races.
Over the course of about four minutes, tires roll, fans blow, and balloons and umbrellas fall from the ceiling. It ends with all four band members being sprayed with paintballs (hence the protective eyewear at the start), an outcome that began when that first domino fell.
What strikes me most about this video is that each isolated event – the dominoes falling, the tire rolling – is connected with every other step of this Rube Goldberg machine.…
Read full article...
When I last wrote about AI on this blog three years ago, I spoke of it being a tool with the potential to transform scientific discovery, but the application I described was primarily theoretical. For AI to be a meaningful tool in R&D, I argued, we needed better sources of “truth” – better data sets that AI tools could query and learn from over time – and technology capable of integrating multiple steps into a semi-automated system. My message was that AI-enabled drug discovery was coming…someday.
Fast forward to 2025, and that someday is now.
We’ve seen an explosion in the availability and capability of AI tools. Just 10 months after I wrote about the theoretical possibilities of AI in biopharma, OpenAI debuted ChatGPT. Shortly after that, we saw the rollout of Microsoft Copilot and Meta AI. We now have immense computational power at our fingertips, with programs specifically designed to query biological problems. Combined with the ingenuity of skilled scientists, who can define the research problem and generate curated datasets that will enable solutions, AI has become an important and practical tool that is helping researchers accelerate discovery (link to Google DeepMind podcast on this topic here; link to a start-up’s pragmatic journey of AI in drug discovery here).…
Read full article...
[ I am an employee of Bristol Myers Squibb. The views expressed here are my own, assuming I am real and not a humanoid. ]
In the original Blade Runner (1982), Harrison Ford’s character, Deckard, implements a fictitious Voight-Kampff test to measure bodily functions such as heart rate and pupillary dilation in response to emotionally provocative questions. The purpose: to establish “truth”, i.e., determine whether an individual is a human or a bioengineered humanoid known as a replicant.
While the Voight-Kampff test was used to establish truth for humans vs replicants, the concept of “truth” is central to neural networks used in machine learning and artificial intelligence (AI). And for AI to be effective in drug discovery and development, it is critical to ask a fundamental question: what is “truth” in drug discovery and development?
INTRODUCTION
I recently read the book Genius Makers by Cade Metz and was reminded of the long history of machine learning, neural networks, and artificial intelligence (AI). This is a field more than 60 years in the making, with slow growth for the first 50 years – AI was founded as an academic discipline in 1956 – and exponential growth in the last 10. The original mathematical framework of neural networks was created in the 50’s (perceptron), 60’s and 70’s (backpropagation), but went largely unappreciated outside of academics, as the practical applications were few and far between.…
Read full article...
