If you’ve followed this blog for a while, you’ll know I often say drug discovery is more like poker than chess (see here and here). You are also likely familiar with our end-to-end R&D principles at Bristol Myers Squibb – our belief that if we select the right targets based on human biology, match the right kind of medicine to the biological mechanism of disease, and provide early evidence that an idea is working, we can increase the probability of success and accelerate the pace of delivering transformational medicines to more patients.
Recently, I built out my poker analogy more thoroughly in a presentation at the Broad Institute’s Variant to Function (V2F) Symposium. Using specific examples within the BMS portfolio, I demonstrated the connection between poker and our R&D principles, highlighting assets at different stages of play. You can view the full presentation here.
The crux of my thesis is this: strong causal human biology, paired with the right modality for the mechanism of action, gives you a great starting hand – much like being dealt a queen and an ace in a round of Texas Hold ‘em poker. But a great starting hand isn’t always a winning hand.…
In a previous blog post, I argued that pharmaceutical Research and Development (R&D) is more like poker than chess given the probabilistic nature of developing new medicines. In this blog, I build out this argument further and introduce a new concept, the Critical Value Creation Period (CVCP). The CVCP is the moment in R&D where the probability of success (PoS) for an investigational medicine becoming a new medicine jumps substantially. For most investigational medicines, this period is in early clinical development. By analogy, early development is like the “flop” in Texas hold ’em poker. Stay with me on this one…hopefully this will become clear by the end of this blog post!
Here, I first build out the CVCP argument, which is modeled after an investment thesis proposed by Bain Capital Life Sciences (see here for a presentation by Dr. Adam Koppel at BCLS). Then, I introduce the poker analogy – with focus on the flop. Next, I discuss practical implication of the CVCP model on decision making in biopharma. Finally, I provide two brief case studies that reinforce these concepts. For those not interested in poker, you can skip the second section and focus on the first, third and fourth sections.…
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).…
At the beginning of the summer, I had the opportunity to join the team at Royalty Pharma for a great event at MIT (link here). It was an interesting time for me as I was thinking about the new role I was about to take at Bristol Myers Squibb. While I had certainly been “a leader” for several years now, I was pushing myself to think through and question my perspectives on R&D now that I was taking on increasing responsibilities as “the leader” of the research organization.
And so the presentation opportunity with Royalty really pushed me to articulate my views in a way that would hopefully resonate with and inspire others. I titled my presentation “Bullseye.Aim.Fire” and then renamed it “Increasing R&D Productivity to Deliver Transformational Medicines” so the topic would be more obvious. What I’m really sharing in the presentation is my fundamental belief about R&D, linking together several factors that I see as mission critical.
To me, it really all comes down to causal human biology. In order to be successful, we must understand the cause-and-effect relationship between perturbing a particular biological target with a medicine and the outcome that will then impact human physiology.…
For me, the most enjoyable aspect of discovery research is exploring the unknown. It is about having a big idea; believing in that big idea based on a scientific belief framework; coming to a crossroads in the validity of the big idea, which is usually marked by deep uncertainty and skepticism; making a data-driven scientific decision to proceed (or not) to the next inflection point of testing the big idea; and ultimately arriving at a conclusion of whether the big idea is true.
Unfortunately, most of these scientific adventure stories are lost in the way we communicate about science. We tell a story to communicate the final message – we have a new medicine that is effective in treating patients – as that is the cleanest way to communicate to an audience not familiar with the gory details of the discovery. Such retrospective narratives are also the simplest way to communicate the validity of the big idea, not the tortuous and often complicated path to arrive at truth.
But such retrospective narratives don’t capture the immensely personal nature of our research discoveries. Moreover, such retrospective narratives often make the big idea seem preordained or obvious, when the big idea was anything but.…
[ 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.…
[Disclaimers: I am an employee of Bristol Myers Squibb. The views expressed here are my own. Also, I am not a particularly good poker or chess player. It is one reason I am a popular invited guest to poker nights with friends.]
I posted on poll on Twitter to ask the question is drug discovery more like poker or chess. There were over 300 responses, with the results split nearly equally (54% poker, 46% chess).
My answer to the question, “Is drug discovery more like poker or chess?”, derives from the following truths:
Poker is a game of skill and chance, where critical information about how to win is hidden. In poker, one has to make probabilistic decisions with incomplete information.
Chess is a game of pure skill, where all information is available and – for the best players – decisions are deterministic. Unlike poker, chess contains no hidden information and very little luck.
Thus, my “answer” to the question is drug discovery is more like poker than chess – largely because of available information (poker = incomplete, chess = complete) and the importance on probabilistic (poker) vs deterministic (chess) decision-making. Here is more context.
Thinking in bets
I recently read the book “Thinking in Bets” by poker champion Annie Duke (@AnnieDuke).…
[Disclaimer: I am an employee of Bristol Myers Squibb. The views expressed here are my own.]
One of my favorite questions to ask is: “What captures your imagination?” At a recent family dinner, responses were varied but encouraging for the next generation: black swan events, comparative anatomy & human physiology, space exploration & intelligent life beyond our planet, and more. My response was programmable therapeutics, a topic which I have blogged about in the past.
In this blog I define programmable therapeutics and provide a few recent examples (severe combined immune deficiency and mRNA vaccines). As you will see, programmable therapeutics is more than pure imagination – we are seeing this new concept evolve before our very eyes.
What is the concept of programmable therapeutics?
While there are different definitions of the concept of programmable therapeutics (see a16z talk; programmable cells; synthetic biology; CRISPR base editing), my definition of programmable therapeutics relates to a platform with modular components that can shorten the time from new target to drug candidate and ultimately regulatory trials that can lead to an approved medicine.
For most drug development programs, the identification of a drug target represents the start of a long journey that is highly artisanal.…
[I am an employee of Bristol-Myers Squibb. The views expressed here are my own.]
One of my predictions for the next decade – the “clear view” decade – is that we will have the ability to click on any gene in the human genome to generate function-phenotype maps. These maps should enable drug discovery by informing on mechanism, magnitude and markers of target perturbation. In particular, I have championed an “allelic series” model, whereby genes with a series of alleles are used to derived genetic dose-response curves (see here, here).
During a recent presentation to my former colleagues at the Division of Genetics at Brigham & Women’s Hospital (BHW, slides here), I discussed important assumptions underlying this model:
Large-scale sequence data will identify a range of protein-coding variants associated with traits of medical interest that are suitable surrogates for drug discovery (allelic series architecture assumption).
It will be possible to use high-throughput functional assays to interrogate the impact of trait-associated variants on cell physiology for the majority of genes in the genome (functional readout assumption).
Large-scale biobanks will emerge to enable testing of these same trait-associated variants for pleiotropic effects across a wide-variety of clinical phenotypes in the real world (PheWAS assumption).
[I am an employee of Celgene. The views expressed here are my own.]
In the Wizard of Oz, Dorothy clicks her heels and hopes for re-entry from her dream world by repeating, “There’s no place like home…there’s no place like home…” I often feel that many in the genetics community look at their human genetics data with the same youthful optimism as Dorothy – clicking their genetic heels and wishing “my genetic discovery will become a drug…my genetic discovery will become a drug…” But without rigor and discipline, such heel-clicking won’t overcome many of the challenges that face drug hunters along the tortuous journey from a genetic idea to a new medicine.
In this blog, I discuss a recent study on the genetics of multiple sclerosis (MS) published in Science (see here). This is a beautiful study that substantially advances the genetic landscape of patients with a devastating disease. However, the study falls short in terms of the application of human genetics to drug discovery. To chart a course for the future, I introduce the concept of mechanism, magnitude and markers (oh my!), which I refer to as the three M’s. …
[I am an employee of Celgene. All views expressed here are my own.]
At the 2018 Annual Atlas Ventures Retreat (AVR), I participated in a panel on Digital Health (along with David Schenkhein, John Reed, Scott Brun). The panel discussion was led by Michael Ringel, who also provide an excellent introduction to Digital Health (his slides here). While there are many aspects to digital health, we focused on the application to drug discovery and development. In this blog, the main point I want to emphasize is that I believe that the digital health tipping point will occur when products that benefit patients (e.g., therapeutics) facilitate the integration of digital health initiatives that currently reside in silos.
What is digital health in relation to drug discovery & development? There are many different definitions with many different components, and this, in essence, is part of the challenge (see Figure below). In early discovery biology, digital health represents various data types (e.g., human genetics, ‘omics data, cell models) and analytical methods (e.g., simple regression, machine learning, artificial intelligence). In late discovery biology, digital health includes sophisticated analytical methods for in silico drug design and organoid models to recapitulate the human system for pre-clinical testing.…
[Disclaimer: I am an employee of Celgene. The views reported here are my own.]
On a recent family vacation to Cumberland Island, a 9,800-acre barrier island off the coast of Georgia, I was mesmerized by the dense forest of live oak trees covered with Spanish moss. Upon first glance, the branches of these magnificent trees extend chaotically in all directions, and it is difficult to discern where the trees begin and end. But upon closer inspection, the root structure can be identified, moss disentangled, and the overall complexity unraveled.
These craggy oak trees serve as metaphor for our complex human biological ecosystem: a dense forest of molecules with gnarled branches of pathways meandering in all directions, without an obvious root structure of human disease. Extending the metaphor further, the oak trees make the point that I see as one of the most difficult aspect of drug discovery and development: understanding root cause of disease, and matching therapeutic modality and biological mechanism to prevent or cure devastating illness.
In this blog, I highlight two recent publications that underscore the importance of matching modality and mechanism. The first article, published in the New England Journal of Medicine, reported clinical data on 22 patients with beta-thalassemia treated with ex vivo gene therapy (here).…
[Disclaimer: I am an employee of Celgene. The views reported here are my own.]
Drug research and development (R&D) is a slow, arduous process. As readers of this blog know, it takes >10 years and upwards of $2.5 billion dollars to bring new therapies to patients in need. An aspiration of the biopharmaceutical ecosystem is to shorten cycle times and increase probability of success, thereby dramatically improving the efficiency of R&D.
One potential solution is to use human genetics to pick targets, understand molecular mechanism, select pharmacodynamics biomarkers, and identify patients most likely to respond to treatment (see Science Translational Medicine article here). While intuitively appealing and supported by retrospective analyses (here), it is not yet routinely implemented in most R&D organizations (although see Amgen blog here; Regeneron study below). Indeed, human genetics often represents an inconvenient path to a new therapeutic, as it takes substantial effort to understand the molecular mechanism responsible for genetic risk and many such targets are difficult to drug.
But what if…
…it were possible to go from gene variant to therapeutic hypothesis instantly via in silico analysis;
…it were possible to select an “off-the shelf” therapeutic molecule that recapitulates a human genetic mutation, and take this molecule into humans almost immediately, with limited pre-clinical testing;
…it were possible to select pharmacodynamics (PD) biomarkers that capture underlying human physiology, and to measure those PD biomarkers in a small, human proof-of-mechanism clinical trial;
…it were possible to model the magnitude of effect of a therapeutic intervention relative to existing standard-of-care, and thereby to estimate the commercial market of an as-yet-to-be-approved drug?…
As readers of my blog know, I am a strong supporter of a disciplined R&D model that focuses on: picking targets based on causal human biology (e.g., genetics); developing molecules that therapeutically recapitulate causal human biology; deploying pharmacodynamic biomarkers that also recapitulate causal human biology; and conducting small clinical proof-of-concept studies to quickly test therapeutic hypotheses (see Figure below).As such, I am constantly on the look-out for literature or news reports to support / refute this model.Each week, I cryptically tweet these reports, and occasionally – like this week – I have the time and energy to write-up the reports in a coherent framework.
Of course, this model is not so easy to follow in the real-world as has been pointed out nicely by Derek Lowe and others (see here).A nice blog this week by Keith Robison (Warp Drive Bio) highlights why drug R&D is so hard.
Here are the studies or news reports from this week that support this model.
(1) Picking targets based on causal human biology: I am a proponent of an “allelic series” model for target identification. Here are a couple of published reports that fit with this model.…
Like many, I waited with bated breath for results of the anti-PCSK9 (evolocumab) FOURIER cardiovascular outcome study last week. There have been many interesting commentaries written on the findings.A few of my favorites are listed here (Matthew Herper), here (David Grainger), here (Derek Lowe), and here (Larry Husten), amongst others, with summaries provided at the end of this blog.Most of these articles focused on clinical risk reduction vs. what was predicted for cardiovascular outcome, as well as whether payers will cover the cost of the drugs.These are incredibly important topics, and I won’t comment on them further here, other than to say that the debate is now about who should get the drug and how much it should cost.
In this blog, I want to emphasize key points that pertain to human genetics and drug discovery.And make no mistake: the anti-PCSK9 story and FOURIER clinical trial outcome is a triumph for genetics and drug discovery. This message seems to be getting muddled, however, given the current cost of evolocumab and the observation that cardiovascular risk reduction was less than expected, based on predictions from a 2005 study published by Cholesterol Treatment Trialists (CTT) (see Lancet study here).…
Yesterday I participated in the National Academy workshop, “Enabling Precision Medicine: The Role of Genetics in Clinical Drug Development” (link here). There were a number of great talks from leaders across academics, industry and government (agenda here).
I was struck, however, by a consistent theme: most think that “precision medicine” will improve delivery of approved therapies or those that are currently being developed, whether or not the therapies were developed originally with precision medicine explicitly in mind. Many assume that the observation that ~90% medicines are effective in only 30% to 50% is the result of biological differences in people across populations (see recent Forbes blog here). This hypothesis is very appealing, as there are many unique features to each of us.
An alternative explanation is that most medicines developed without precision medicine from the beginning only work in ~30% patients because the medicines don’t target the biological pathways that make each of us unique.
I believe the most likely application is in the discovery and development of new therapies. That is, I believe that the greatest impact will come when precision medicine strategies are incorporated into the very beginning of drug discovery, and will only rarely have an impact on therapies that were not developed with precision medicine in mind from the start.…
A new sickle cell anemia gene therapy study published in the New England Journal of Medicine (see here, here) gives hope to patients and the concept of rapidly programmable therapeutics based on causal human biology. But how close are we really?
It takes approximately 5-7 years to advance from a therapeutic hypothesis to an early stage clinical trial, and an additional 4-7 years of late stage clinical studies to advance to regulatory approval. This is simply too long, too inefficient and too expensive.
But how can timelines be shortened?
In the current regulatory environment, it is difficult to compress late stage development timelines. This leaves the time between target selection (or “discovery”) and early clinical trials (ideally clinical proof-of-concept, or “PoC”) as an important time to gain efficiencies. Further, discovery to PoC is an important juncture for minimizing failure rates in late development and delivering value to patients in the real world (see here).
Here, I argue that rapidly programmable therapeutics based a molecular understanding of the causal disease process is key to compressing the discovery to PoC timeline.
Imagine a world where the molecular basis of disease is completely understood. For common diseases, germline genetics contributes approximately two-thirds of risk; for rare diseases, germline genetics contributes nearly 100% of risk.…
I know it may seem strange that I am commenting on my own commentary. While writing the Science Translational Medicine (STM) article (here, here), I wanted to focus on a clear vision for improving R&D productivity via integrating human biology, therapeutic modulation, pharmacodynamics biomarkers, and proof-of-concept clinical trials (see Figure 1 of the manuscript, also above). Too often, these four key concepts get lost amongst the many steps in drug development. I won’t revisit these concepts here, as I encourage you to read the article itself.
However, there is much more to R&D productivity than just these four concepts, and the purpose of this blog is to broaden the perspective piece a bit.
[Disclaimer: I am a Merck/MSD employee. The opinions I am expressing are my own and do not necessarily represent the position of my employer.]
1. Drug discovery should always start with the patient. Every drug discovery journey must begin with a clear definition of the unmet medical need. As a former practicing rheumatologist, I always try to keep in mind the questions that a patient would ask me: “Why did I develop this disease?” “Will I respond to this medication?” “What is my prognosis?”…
Here are my thoughts on the Discussion Paper by Bernard H. Munos and John J. Orloff, “Disruptive Innovation and Transformation of the Drug Discovery and Development Enterprise” (download pdf here). This blog won’t make much sense if read out-of-context. Thus, I recommend reading the Discussion Paper itself, and using this blog as a companion guide at the completion of each section.
[Disclaimer: I am a Merck/MSD employee. The opinions I am expressing are my own and do not necessarily represent the position of my employer.]
STRENGTHS AND WEAKNESSES OF THE CURRENT INDUSTRY MODEL
In the near-term (10-years), I suspect that the pharma model of late development and commercialization will likely persist, as the cost and complexity of getting a drug approved is difficult by other mechanisms. Over time, however, new ways of performing late-stage trials will likely evolve. Drugs that are today in the early R&D pipeline will drive this evolution. If drugs look like they do today, dominated by small molecules and biologics with high probability of failure in Phase II/III, then the current model will likely continue with incremental improvements in efficiency. However, if new therapeutic modalities emerge (CRISPR, mRNA, microbiome, etc) and/or the probability of success in Phase II/III improves substantially, then the model of late development and commercialization will be forced to evolve, too.…
Inevitably when I post a blog on “human biology” I get a series of comments about the importance of non-human model organisms in drug discovery and development. My position is clear: pick targets based on causal human biology, and then use whatever means necessary to advance a drug discovery program to the clinic.
Very often, non-human model organisms are the “whatever means necessary” to understand mechanism of action. For example, while human genetic studies identified PCSK9 as an important regulator of LDL cholesterol, mouse studies were critical to understand that PCSK9 acts via binding to LDL receptor (LDLR) on the surface of cells (see here). As a consequence, therapeutic antibodies were designed to block circulating PCSK9 from the blood and increase LDLR-mediated removal of circulating LDL (and hopefully to protect from cardiovascular disease).
Moreover, non-human animal models are necessary to understand in vivo pharmacology and safety of therapeutic molecules before advancing into human clinical trials.
Beyond drug discovery, of course, studies from non-human animal models provide fundamental biological insights. Without studies of prokaryotic organisms, for example, we would not have powerful genome-editing tools such as CRISPR-Cas9. Without decades of work on mouse embryonic stem cells, we would not have human induced pluripotent stem cells (iPSCs).…
