Learning biotech, part 4: The nature of biotech is to grope your way through darkness, looking for unseen treasures with mortal danger ever looming. In biotech, confidence = $. The value we create in biotech is not a physical product, but confidence in an approach to solving a human ailment in need of a solution. You build up that value by reducing the ways you might fail in those pivotal Phase 3 trials. For a walkthrough of a typical process, I recommend these two blog posts from @age1vc & @LynettaWang126: On picking targets: librariesforthefuture.bio/p/… And on going from target to medicine: librariesforthefuture.bio/p/… If you are coming from academia, the GOT-IT framework is a nice summary of what you should be thinking about: nature.com/articles/s41573-0… If you're already thinking about drugs, @rplenge of BMS published an excellent overview of important considerations during late stage drug design and into clinical trials. pubmed.ncbi.nlm.nih.gov/2746… (One note is the strong emphasis on human genetics, which gives you much higher confidence in the target, but is what's leading to immense target crowding). Fundamentally, finding the right solutions depends on the number of options you can assess, and the predictive validity of your assessment. Predictive validity is @JackScannell13's term, and it's worth reading his paper: nature.com/articles/s41573-0… One key takeaway (relevant in the AI era), is that testing more options plateaus much faster than improving the precision of your tests. What's missing from the paper is how to determine predictive validity. Here we can list the things you'll want to be confident in: 1) Will doing a thing have the biological effect I predict? 2) Does that biological effect actually prevent disease, or allow for reversing it? 3) When and where does this happen in human beings? 4) Do I have a way to do the thing? 5) What side effects will my way of doing the thing have? Each of these has its own set of experimental models, but in all cases what you want is that the system you can experiment in gives the same answers as the patients you eventually want to treat. There's a lot of lazy thinking on models these days: 'mIcE aReNt HuMaNs!', sure, and neither are embryonic cell blobs. The question is which models have high predictive validity for the specific questions you're asking them. So think about that separately for question 1, 2, 4, 5. For 1 and 2 you can estimate false positive rates by how often drugs that failed trials for efficacy seemed to work in each model. False negatives are harder, since lack of cures is why we work on diseases. As a surrogate/directional indicator, it's wise to consider whether the model develops the disease in the same way as humans (eg get obese on high calorie diets), and whether it manifests as the same cellular and physiological changes. This paper from GSK is a good repreentative approach for assessing models: pmc.ncbi.nlm.nih.gov/article… One thing I think is underrated is replicating conditions of disease whether or not you understand their importance, e.g. using aged animals for age-related diseases. That way the things you don't understand are less likely to fool you. Just as important as establishing your target's role in disease, is knowing whether your drug affects the target (i.e. #4). This is simple(ish) in vitro, but harder to track in living systems. You want to show that your target is bound by the drug, but also how much of your drug ends up in which organs, and what else it binds to. Even seemingly simple in vitro assays can fool you, when you're not measuring what you think you're measuring. Check out this Derek Lowe post on the fate of David Sinclair's first company, Sirtris: science.org/content/blog-pos…, and his other posts on assays. The golden rule here, is to use multiple independent methods to answer key questions, so you avoid assay artifacts. Same answer in different contexts boosts confidence. But really, just hire people who have a lot of experience with drug dev assay development. Things get really tricky when you're in human trials. You'll try to sort out question 3 beforehand, and seeing your target change in human disease is key. But there's only so much you can conclude from postmortem samples. When should you start treatment? How do you know where the drug goes when you can only get a blood sample now and then? The best drug programs thinks about this early, planning around how to measure whether what happens in your preclinical systems holds true in patients. Roche put out a nice framework, focused on neuro but with general lessons: sciencedirect.com/science/ar… Another framework, from Eli Lilly, focuses on how you can use early stage trials to build confidence: assets.ctfassets.net/srys4uk… One factor that junior biotech people undervalue is determining things you CAN measure in blood or other minimally invasive samples, i.e. biomarkers, that tell you about what's happening in the patient. The core theme across all these approaches is that you need to find reliable ways of measuring things, ideally multiple orthogonal ways, and apply those in systems that behave the same as patients for that specific question. Of course, the safest choice of all is to take a target that already has an approved drug, and make a slightly better one. No target risk! No patient risk! Now you understand how we got to that picture of target crowding in part 2. A note on 'key opinion leaders', AKA the most esteemed (clinician-)scientists in your disease area: Some people use KoL conversations to build confidence in a target. I would not do this. First, you'll be chasing the same targets as everyone else they consult for. Second, just like Jim Cramer, if they were actually sure about what to do they'd be billionaires already. I would definitely talk to KOLs, but I'd use them to understand what's going on in field, not to tell me what to do.

Learning biotech part 2: Once you understand process of finding drugs, you should study the incentives that drive decisions. Why does someone choose to launch the 200th EGFR program? Inevitably there's a set of incentives, and often data-driven risk analysis, motivating these seemingly silly decisions. While new biotechs often challenge existing assumptions, understanding Chesterton's fence is the smartest way to decide where and why you deviate. I would start by reading AstraZeneca's retrospective on why their programs won or lost: “Lessons learned from the fate of AstraZeneca’s drug pipeline: a five-dimensional framework.” pubmed.ncbi.nlm.nih.gov/2483… The '5Rs' they identified are a good representation of the risks pharma prioritizes: right target, right tissue, right safety, right patient, and right commercial potential. YC tells their companies to talk to users. But who is the 'user' for biotech? - Ultimately, we make medicines for patients to enjoy healthy lives. - The FDA judges whether it works. - But doctors decide what to prescribe. - And insurance and medicare actually pay for the treatment. This makes the situation more confusing for new biotechs. But your product needs to work for all four of those user. (even if you plan to sell to pharma before approval, they have the same users, so you just have one more user to please) I recommend the book 📚 "Her-2" about Genentech's mega-successful drug Herceptin, which includes decisions about which diseases and trials they tried it for first. If those decisions had been wrong, the drug might have died. I'd also recommend (any and all) episodes of 🎧 "Hard Drugs" from @salonium and @JacobTref, especially for examples where the science incentives are stronger than financial ones. harddrugs.worksinprogress.co… After that, try deconstructing decisions you see in the news, like Genentech choosing to exit cancer immunology two years ago or everyone buying 'radiopharmaceutical' biotechs. What got derisked, or saw added risk? One source of 'irrational behavior' might be classic short-term performance goals among executives at public companies. At any given time, pharma's revenue come from a small number of patent-protected drugs. When patents expire on a top product revenue plummets, so you often see acquisitions of mature companies that might get a new big approval before the cliff. Basics of patent cliffs here: biopharmadive.com/news/pharm… and this episode of @alexkesin and Matthew Pech's 🎧 Approved goes deep on that drug, including at ~90m how Merck sought to protect its profits: podcasts.apple.com/us/podcas… Ultimately, incentives all come back to 1) a desire to improve humanity's conditions by scientists who go into biomedicine, and 2) free market competition. When you see something that seems cool (like PCSK9 gene therapy) but isn't super hot, check the incentive stack to understand why (in this case, strong competition with lower risk). And when you see an opportunity nobody is filling, at least find out why. More good content on this topic includes: @alexkesin’s ✍️ “Pharmacopoeia” broadly covers how biopharma behaves, for example in “Pharma Is Hollywood, Roughly”: alexkesin.com/p/pharma-is-ho… @ElliotHershberg’s ✍️ “The Century of Biology” has a series on platform companies, like BridgeBio, that dissects incentives: centuryofbio.com/p/bridgebio Finally, 📚 “The Great American Drug Deal” by @PeterKolchinsky is relevant as big picture context on how healthcare works and how the medicines it pays for flows back into the next generation of drugs. You can also get the short version on this website: nopatientleftbehind.org/fund…