Five TechBio Myths That Must Go

So, by the way, Eroom's law — the reason Jack Scannell et al. named it that is that it's "Moore's law" spelled backwards. It's a clever pun. You have the technical version, Moore's law, that exponential advancement in technology, and then in pharma you have the exponential deceleration, if you will. So he called it Eroom's law. One of the reasons it was such a resounding hit is the implication in the question you just asked: why is our efficiency decreasing when our technology is increasing? Usually efficiency increases with technology, and costs decrease with efficiency and technology.

People talk about this in the pharma industry a lot. It used to cost something like $2 billion to sequence the first genome. Today — I don't know the exact going rate — you could get it done for a hundred or two hundred bucks. That's Moore's law in action.

Eroom's law is that we have all this genetic information and no freaking idea what to do with it. I allude to this in the article: maybe some of our decrease in efficiency is actually due to our over-reliance on technology, outsourcing the critical thinking to it. Technology also biases us toward a particular way of doing things. Think about it — how did those OG scientists from the '50s generate so many discoveries? They just had a microscope and the power of observation. They were legit scientists, legit thinkers, philosophers in their own way.

Take the molecular revolution that came on the heels of being able to sequence. I think it led us terribly astray in cancer, because there was a compelling narrative we could sell to grant agencies, to investors, to anyone who'd listen: because we can now decode the genome, we can decode cancer. It's as simple as finding out which genes are malfunctioning and then addressing the problem. So we took a very gene-centric view of cancer.

It was like a gene horoscope of cancer. And we've been humbled since — biology is far more complicated. There's what's called emergentism, or holism: the whole is not equal to the sum of the individual parts. A plus B does not equal C. Two plus two doesn't equal four — it's more like two plus two equals five. An example would be consciousness. We know it resides somewhere in the brain, but where? You'll never understand consciousness by studying the functions of a neuron. The neuron enables consciousness, but it tells you nothing about how it works or where it resides. You could dissociate me and put me right back together — and it's like, well, when you dissociated me, where was consciousness? What number of neurons do we need to add for consciousness? No one can answer those questions yet. Consciousness is an emergent property of biology — something much more sophisticated emerges that you can't predict from looking at the individual parts.

Yeah, absolutely. So maybe this is a good point to talk about my contrarian view of cancer. Before we started, you and I were talking about how my very Mexican sense of humor has had a big influence on my life. This is a classic example — how a stupid joke I made in the lab completely ended up transforming my perspective on cancer.

At the time I was being taught how to culture brain cancer stem cells. These are the most aggressive cancer cells you can imagine — resistant to chemotherapy, resistant to radiation. You can nuke them and you won't kill them. They're incredibly robust. And as they're teaching me to culture them, they say: whatever you do, don't add serum media, because they'll stop growing. That's a big deal, because most cells are grown in serum media — serum has all sorts of stuff in it, growth factors and so on, that helps cells grow in culture. But with these cells, whatever you do, do not give them serum media, because they'll stop growing.

So, with my very Mexican sense of humor, I go: oh well, guys, guess what — I just found the cure for brain cancer, it's serum media! The speed of the joke was that, obviously, brain tumors are already exposed to elements of that. My friend gave me a pity laugh, we all scratched our heads, and moved on. But it stuck with me. I thought, wait a minute — there's something here that's not quite making sense. Are you telling me these cells we can't kill with anything, we can make stop with serum media? Actually, a dirty secret in science, if we can call it that, is that it's quite easy to make cancer cells stop growing, or die, in a dish.

And that gets me to the point I was about to make. If I asked you to define cancer — if you asked me to define it according to consensus — I'd say cancer is a disease, or more correctly a collection of diseases, characterized by the uncontrolled proliferation of abnormal cells. And what makes them abnormal? Presumably that they have genetic mutations. Except, if you actually think about it, not one key assumption in that statement is correct. Every part of it is incorrect. And that may sound shocking, almost insulting — wait, that doesn't make sense, we all know that's exactly what it is, so many brilliant people work on it and they all agree. But it's actually very easy to prove.

Take uncontrolled proliferation. Clearly it's controllable. And don't take my word for it — in the clinic we see it all the time. It's called dormancy. Tumors can enter dormancy and then exit it, maybe some months later, maybe ten years later. So clearly cancer cells are capable of controlling their proliferation — which has profound implications for the idea of driver mutations. These tumor cells have these so-called driver mutations, but they can control their proliferation.

So it's not a purely intrinsically driven process. It's clearly a conversation with the extrinsic environment — the microenvironment. There's something promoting or inhibiting the behavior of these cells. From that observation, I started questioning the premise: are genes really drivers? My simplistic, rudimentary answer is that they're not drivers, they're enablers. Cancer is the perfect storm of a bunch of fail-safe systems that failed. There's no one thing.

Apply the scientific method. People born with a susceptibility to breast cancer because they have a BRCA mutation, or to colon cancer — it increases your chances, but it doesn't predetermine you to getting cancer. We see a correlation and think it's the cause. But the opposite test would be: if you have this mutation, do you always get cancer? The answer is a resounding no. So clearly that in itself isn't enough to develop cancer.

Another perfect example — people with Li-Fraumeni syndrome are born with a mutation in p53, classically known as the guardian of the genome, involved in helping cells repair themselves. People with Li-Fraumeni are heavily, heavily predisposed to cancer, and most of them do get it later in life. But they don't get cancer everywhere. There's no random assortment of cancers — it's very specific to certain tissues. Why would that be, when the mutation is present in every single cell in their body? Why aren't they getting cancer in their big toe, or their heart, or their ear? It's very specific to the context. So clearly there's something we're still not identifying that goes beyond the guardian-of-the-genome role. Cancer isn't a disease of uncontrolled proliferation.

Well, what about abnormal cells — mutations? Let me explain why mutations in and of themselves don't explain cancer. And, more importantly, cancer isn't even a disease of cells. Philosophically speaking, there's no such thing as a cancer cell, because cancer isn't a single-cell disease. There's a reason bacteria don't get cancer. Have you ever heard of a bacterium getting a mutation and going, oh no, I can't stop dividing, I'm just proliferating indefinitely?

Of course not. Cancer is a multicellular disease. It can only arise at the tissue level. It does not exist at the genomic level, it does not exist at the cellular level — it exists entirely at the tissue level. That's why we've always relied on pathologists to diagnose cancer. In fact, a pathologist wouldn't describe cancer as a disease of proliferation of abnormal cells. Look at something like Ki-67, a marker for proliferation — a high Ki-67 score is around 20%, meaning one out of five cells within the tumor is dividing at any one time. Twenty percent. That does not seem like uncontrolled proliferation to me. It's clearly not exponential.

So if you think about all these things, it makes you reassess how we're going about cancer. Why are we trying to stop proliferation? The answer is historical. In the lab it's just easier — we know cancer is a thing that grows, and if we do X it stops growing. That was one of the few things we could measure objectively, physically, so it stuck around. But if it's not a disease of proliferation, then why are we trying to stop proliferation? I'm not saying we shouldn't — I'm just saying, let's think about it critically. Is that really the right readout?

My contrarian view is that the cancer models we have today are actually pretty good. Some of them. Not all — some are abysmal. But primary cell lines, freshly created from tumor tissue, have high concordance with clinical responses. That's not the same as durable responses, but they're at least well correlated to how a patient will initially respond. That makes sense, because there's an epigenetic memory in the cells representing the environment they came from. Over time they start losing that memory, or rewriting it.

So the molecular revolution led us astray because it made us gene-centric. It was a sexy pattern to sell — suddenly things seemed feasible and simple, when they were far from it. But it's just as important to be critical about the underlying assumptions. Another example: Alzheimer's. I don't know if you heard the recent news that the data which led the world to pursue the beta-amyloid hypothesis was apparently fabricated. I can't speak for the companies in the space — they may have a perfectly reasonable answer. But I wonder: why didn't all the AI companies in Europe find that? Maybe it's because — and this might well be the answer — they weren't actually pursuing that question. They weren't interested in that avenue and never looked into it. But another reason might be that they did look and just never noticed, because the data we're feeding our algorithms was inherently biased to believe the underlying hypothesis was correct. That's an example of how AI won't get you out of a bind if your underlying assumptions are wrong.

Yeah, you're mostly correct — there are just some clarifications worth bringing up. First, the idea that it takes 17 years to develop a drug is potentially technically true, but also kind of irrelevant, because a lot of those years are already baked in today. If you give me a hundred million dollars, I could — I don't want to say very easily, but quite feasibly — get a drug approval in five years. Why? Because there's already decades of baked-in research I can in-license and then push through clinical trials. If I start a company today, it's not going to take me 17 years.

And that's the cheap trick a lot of us in the techbio field implement: no one is really going after novel drug targets. Now, I'm fascinated, because I do think we're going after novel targets — but it's more that we've rediscovered them. You can read about these phenomena since the early days of pathology. In fact, the father of modern pathology, the German pathologist Rudolf Virchow, was describing these very things in his journal. They're well known, but maybe not fully appreciated.

The point is that the timeline is kind of irrelevant, because a lot of it's already baked into whatever we did today. Take the example I used of Exscientia and their ability to develop a CDK inhibitor. Super awesome chemistry, a really cool platform — I don't want to trivialize it. But the reason it was possible is that there was already an abundance of literature to build off. You can't necessarily apply that same technology to something we don't know about — there's no literature, so you can't feed anything to your algorithms. And even then, you may come up with an awesome CDK inhibitor, but we still don't know if CDK7 inhibition is going to be meaningful in the clinic.

Those are some of the inherent limitations of applying this technology. Most techbio companies today are going after known biology. Some of that is strategic — it lets you benchmark. If you go after novel biology, or a disease indication with no other drug on the market, it's hard to benchmark. It may also be deemed a safer market because there's some validation. So there are mostly strategic reasons. But the idea that it takes 17 years to develop a drug today — I don't really see it as true. Technically, if you trace a drug's history back to its beginnings, then yeah, probably. But that's irrelevant for our present reality. So I think that answers the first part of your question. There was a second part.

Oh, got it. That relates back to Eroom's law. What was so brilliant about Jack Scannell's observation is that technology has not made us more efficient. Drug discovery keeps getting more expensive at the same time operations are getting cheaper. Sequencing genomes is super cheap now compared to before. So that delta — the difference between what the price was before and what it is today — we're filling it up and overshooting, in terms of the number of experiments we now run. We're doing more science than ever. That's one reason things are more expensive: we're doing more science than ever, and yet we're finding less than ever.

It's time to reassess our assumptions, guys. Maybe we've just been working off the wrong ones. We've got to be a little self-critical. If 95% of all trials fail, why do we keep relying on the same pre-clinical readouts? Every clinical trial we run first had to look amazing pre-clinically — it's hard to convince anyone to fund your trial if your data looks like crap. So clearly our pre-clinical readouts have no relevance whatsoever to the clinic. Does that mean our models are wrong, or the things we're measuring are wrong? My stance is that the things we're measuring are wrong.

Viability assays in cancer are a perfect example. What the hell does that even mean? Ask different cancer biologists and you'll get different answers. Some say it's a measure of health — it's not. Some say robustness — what does that mean? Super general. Some say proliferation, which it can be. But you can have two identical viability readouts that mean completely different things. Take a 50% reduction in viability. In one scenario you didn't kill a single cancer cell — they just stopped growing. Since you're comparing to a control that wasn't exposed to the drug, and the control kept growing while this one didn't, the delta is a 50% reduction. But you didn't kill a single cell. In the other scenario, you killed 50% of the cells, and the remaining 50% proliferated to make up for the ones that died — so you end up with the same number of cells, again a 50% reduction versus control. Those are two vastly different scenarios, mechanistically, but our readouts suggest they're identical. That's part of the problem. Our readouts suck, and we don't really think about them.

I get really worked up about this stuff. But going back to why it doesn't increase speed: we're just doing more of the same, faster, and we end up in the same place — failure. As long as we keep failing, we're not going to improve anything. You can make drug discovery a thousand times faster and move a thousand more molecules into the clinic, but if they're going to keep failing, we're not really doing anything. The only way to improve efficiency is to improve the success rate — meaningfully improve the probability of clinical success. The only way to retain market share is to be best in class.

So my personal belief is that only two things matter for AI drug discovery companies: a substantially greater probability of success, and finding the best possible candidate. We should hold ourselves to that bar. You don't need AI to get a drug approval today — we've been doing that without AI. So if AI is such a big difference-maker, the difference should be that we're now finding absolutely the best drugs to put on the market, and doing so at a much higher probability of success.

Yeah, exactly — that's a great analogy. Some people have appropriated tech culture. Tony from Pillar VC made a point I very much agree with: techbio is just biotech. It really is the same thing. The reason I went with "techbio" is that at least you know what type of company we're referring to — the biotech equivalent of tech, younger founders, not the old boys' club of yesteryear. That's the contrast in founder phenotype.

Some people in techbio have also appropriated "fail faster," which in tech makes total sense — you want to iterate as quickly as possible to arrive at the solution. But biology is a vastly different enterprise. You can fail as fast as you want, as many times as you want. Failing doesn't necessarily mean you're going to figure out how to do things — it just means you'll keep failing. I'm not interested in failing faster. I'm interested in being right, as early as possible. I don't want to fail a thousand times before I get to the right answer — I want to get there within ten tries. That's my personal thing. I don't want to fail faster; I want to be right.

Yeah, I definitely have thoughts — as you can tell, I love to pontificate. In a past life I was probably supposed to be a philosopher, just sitting on a rock thinking. But there's one problem that underlies all funding mechanisms: groupthink, which is inevitable. It's just human nature.

The worst thing we can do for innovation is to institutionalize things. We institutionalize public funding through the NIH — and it makes sense, you're trying to disperse that money responsibly, and public bodies want to make sure that happens. But you end up adding bureaucracy and a bunch of gatekeepers. Who reviews the grants? Who determines what gets funded? The people you put in place. And what are they going to fund? The things they like — the things that support their hypothesis.

The same thing happens in VC. I appreciate that VCs are, in some ways, risk-takers, but actually VCs are incredibly risk-averse — just in different ways. The best VCs understand the value of risk. Some VCs — everyone knows them, some people view them as controversial, I'm kind of indifferent, I just love listening to smart people. You can dislike someone, or disagree with them, and it doesn't change the fact that they're super smart and saying something you should pay attention to.

For example, Andreessen Horowitz talks about how the best investors optimize for huge swings, because it's a power law — the law of outliers. If you get that one win, it makes up for every other loss. And what's key is that VCs are remembered not by their losses or the companies that failed, but by the one company that made it big. Every investor in Facebook has been solidified in the hall of fame. They probably invested in a thousand companies that failed before and after Facebook, and nobody cares, because they have that "I invested in Facebook" — instant elite status.

Another investor with a keen insight here — a bit more controversial, but again I'm indifferent — is Peter Thiel. He says what's key to a successful business and market dominance is trade secrets, secret knowledge. I very much agree. In biotech we give too much value to patents. Patents are kind of necessary, but they don't give you the market protection you think. There are a million ways to target the same target, so a patent doesn't prohibit people from coming into your market and taking share. And it doesn't technically stop people from blatantly ripping off your product.

This is hypothetical — I'm not accusing anyone or any government. But say tensions are high between the US and China, and you realize there's a company in China literally selling your drug, having reverse-engineered the formulation, under their own name. You get angry, you call on international bodies, you get your government to call China and say, hey, you've got to shut those guys down. Maybe China isn't that incentivized to help you, because they could let that company provide the drug to their population for a significantly cheaper price, or rush to defend another government's company so it can sell the same drug to their people for much more. My sense is they probably wouldn't be that interested in helping. They might say: yeah, yeah, we'll look into it eventually. And that's a small company — good luck enforcing these policies. Your IP is usually stolen as well. So trade secrets are a big difference-maker.

Going back to funding models — the problem I see with VCs a lot is that groupthink. In techbio I call some of this out: everyone starts writing thought pieces, they all start agreeing with each other, assimilating the same information and ideas, and then collectively decide what to fund. So if someone is outside those schemes, they may be very right, but they're not within the scheme. I'm not necessarily blaming them — it's human nature, it's indispensable. Think of it from the investor's standpoint: they get inundated with deal flow. You don't have the mental bandwidth. After a while they'd all look the same to me. You don't have the energy to think critically about what each one is telling you. So what do you do? You start deferring to social signals. And there's an inherent fear of looking stupid. So you think: what are my buddies investing in?

You might like something, but you don't really know the team — you just think there's something there. Then one of your buddy investors comes in and says: no, no, forget those guys, I've got my champion right here, let's syndicate around this guy, I've known him, we went to Stanford together. And that's exactly how VC syndication works. So at the end of the day you still get VC groupthink.

But what I like more about the VC model than the grant model is the efficiency. You get from a no to a no a lot faster. You can usually get an answer within a few weeks to a couple of months. If it's taking more than two months, either it's a super-large raise that requires much more due diligence — but for the most part you can learn within a week. Whereas the grant model can take literally months before you even know you got it, and many more months for the money to come through. And some of it you have to pay automatically to the university — the overhead fee. If you're wondering why American universities love research, it's because they take a chunk of every grant their scientists get. It's a huge revenue model.

So at the end of the day, the problem is always human. We're the thing that corrupts everything.

For sure. This is probably more of a trait than a habit. I'm a very ADD person, and a very curious person. I love science because it's just endless rabbit holes — the best thing I could ask for. I'll get lost in them.

Usually, when we enter a new field, we try to figure out very quickly where the consensus is — what is everyone saying, what's the state of knowledge — and then build off that. But in my case, because I have this insatiable curiosity, instead of building up from the consensus, I start digging in. I go: okay, this is the consensus, but why are we saying this?

Now, I always joke that my superpower is asking all the stupid questions. People say there's no such thing as a stupid question. Oh, there is — I've asked plenty, trust me. But every now and then, what seems like a stupid question is actually a brilliant one, because the most impactful things are often hidden in plain sight — no one ever really saw them. I gave you the example of cancer dormancy. Everyone knows about it, but no one connects it to the idea that cancer is supposedly a disease of uncontrolled proliferation. Clearly it's controllable — something controls it.

I like to dig in and really understand why we say the things we say. I consider myself a very slow learner, but I think that's part of the ADD. It's not that I learn slowly — it's that I'm processing more things at the same time. "Attention deficit" is such a misnomer; it's the opposite. You're paying attention to too many things at once. It's more of a focus deficit. Normally you have a refined filter — I want to learn this — which is good for efficiency. My filter isn't like that. I'm thinking about a lot of things at once. I'm a slow reader because my wheels are churning as I read. It takes me a while to learn something, but once I do, I have a pretty solid grasp, because I've contemplated all these weird, seemingly unrelated tangents. In my mind I go: oh, that reminds me of this — I create a web of ideas.

I've found it incredibly helpful, as a professional and especially as a scientist, to take that curiosity and apply it to everything — to ask the dumb questions. Rather than working off assumptions, ask why these assumptions are even built in, because most of the time we're working off built-in assumptions we don't even realize we hold. That's what we were alluding to with Peter Thiel — suddenly you learn something that seemed trivial or obvious, hidden in plain sight, and it completely transforms the way you go about things.