I was part of the founding team at MetaMed, a personalized medicine startup.  We went out of business back in 2015.  We made a lot of mistakes due to inexperience, some of which I deeply regret.

I’m reflecting on that now, because Perlara just went out of business, and they got a lot farther on our original dream than we ever did. Q-State Biosciences, which is still around, is using a similar model.

The phenomenon that inspired MetaMed is that we knew of stories of heroic, scientifically literate patients and families of patients with incurable diseases, who came up with cures for their own conditions.  Physicist Leo Szilard, the “father of the atom bomb”, designed a course of radiation therapy to cure his own bladder cancer.  Computer scientist Matt Might analyzed his son’s genome to find a cure for his rare disorder.  Cognitive scientist Joshua Tenenbaum found a personalized treatment for his father’s cancer.

So, we thought, could we try to scale up this process to help more people?

In Lois McMaster Bujold’s science fiction novels, the hero suffers an accident that leaves him with a seizure disorder. He goes to a medical research center and clinic, the Durona Group, and they design a neural prosthetic for him that prevents the seizures.

This sounds like it ought to be a thing that exists. Patient-led, bench-to-bedside drug discovery or medical device engineering.  You get an incurable disease, you fund scientists/doctors/engineers to discover a cure, and now others with the disease can also be cured.

There’s actually a growing community of organizations trying to do things sort of in this vein.  Recursion Pharmaceuticals, where I used to work, does drug discovery for rare diseases. Sv.ai organizes hackathons for analyzing genetic data to help patients with rare diseases find the root cause.  Perlara and Q-state use animal models and in-vitro models respectively to simulate patients’ disorders, and then look for drugs or gene therapies that reverse those disease phenotypes in the animals or cells.

Back at MetaMed, I think we were groping towards something like this, but never really found our way there.

One reason is that we didn’t narrow our focus enough.  We were trying to solve too many problems at once, all called “personalized medicine.”

Personalized Lifestyle Optimization

Some “personalized medicine” is about health optimization for basically healthy people. A lot of it amounts to superficial personalization on top of generic lifestyle advice. Harmless, but more of a marketing thing than a science thing, and not very interesting from a humanitarian perspective.  Sometimes, we tried to get clients from this market.  I pretty much always thought this was a bad idea.

Personalized Medicine For All

Some “personalized medicine” is about the claim that the best way to treat even common diseases often depends on individual factors, such as genes.

This was part of our pitch, but as I learned more, I came to believe that this kind of “personalization” has very little applicability.  In most cases, we don’t know enough about how genes affect response to treatment to be able to improve outcomes by stratifying treatments based on genes.  In the few cases where we know people with different genes need different treatments, it’s often already standard medical practice to run those tests.  I now think there’s not a clear opportunity for a startup to improve the baseline through this kind of personalized medicine.

Preventing Medical Error

Some of our founding inspirations were the work of Gerd Gigerenzer and Atul Gawande, who showed that medical errors were the cause of many deaths, that doctors tend to be statistically illiterate, and that systematizing tools like checklists and statistical prediction rules save lives.  We wanted to be part of the “evidence-based medicine” movement by helping patients whose doctors had failed them.

I now think that we weren’t really in a position to do that as a company that sold consultations to individual patients. Many of the improvements in systematization that were clearly “good buys” have, in fact, been implemented in hospitals since Gawande and Gigerenzer first wrote about them.  We never saw a clear-cut case of a patient whose doctors had “dropped the ball” by giving them an obviously wrong treatment, except where the patient was facing financial hardship and had to transfer to substandard medical care.  I think doctors don’t make true unforced errors in diagnosis or treatment plan that often; and medical errors like “operating on the wrong leg” that happen in fast-paced decisionmaking environments were necessarily outside our scope.  I think there might be an opportunity to do a lot better than baseline by building a “smart hospital” that runs on checklists, statistical prediction rules, outcomes monitoring, and other evidence-based practices — Intermountain is the closest thing I know about, and they do get great outcomes — but that’s an epically hard problem, it’s political as much as medical and technological, and we weren’t in a position to make any headway on it.

AI Diagnosis

We were also hoping to automate diagnosis and treatment planning in a personalized manner.  “Given your symptoms, demographics, and genetic & lab test data, and given published research on epidemiology and clinical experiments, what are the most likely candidate diagnoses for you, and what are the treatments most likely to be effective for you?”

I used to be a big believer in the potential of this approach, but in the process of actually trying to build the AI, I ran into obstacles which were fundamentally philosophical. (No, it’s not “machines don’t have empathy” or anything like that.  It’s about the irreducible dependence on how you frame the problem, which makes “expert systems” dependent on an impractical, expensive amount of human labor up front.)

Connecting Patients with Experimental Therapies

Yet another “personalized medicine” problem we were trying to solve is the fact that patients with incurable diseases have a hard time learning about and getting access to experimental therapies, and could use a consultant who would guide them through the process and help get them into studies of new treatments.

I still think this is a real and serious problem for patients, and potentially an opportunity for entrepreneurs.  (Either on the consulting model, or more on the software side, via creating tools for matching patients with clinical trials — since clinical trials also struggle to recruit patients.)  In order to focus on this model, though, we’d have had to invest a lot more than we did into high-touch relationships with patients and building a network of clinician-researchers we could connect them with.

When Standard Practice Doesn’t Match Scientific Evidence

One kind of “medical error” we did see on occasion was when the patient’s doctors are dutifully doing the treatment that’s “standard-of-care”, but the medical literature actually shows that the standard-of-care is wrong.

There are cases where large, well-conducted studies clearly show that treatment A and treatment B have the same efficacy but B has worse side effects, and yet, “first-line treatment” is B for some reason.

There are cases where there’s a lot of evidence that “standard” cut-offs are in the wrong place. “Subclinical hypothyroidism” still benefits from supplemental thyroid hormone; higher-than-standard doses of allopurinol control gout better; “standard” light therapy for seasonal affective disorder doesn’t work as well as ultra-bright lights; etc.  More Dakka.

There are also cases where a scientist found an intervention effective, and published a striking result, and maybe it was even publicized widely in places like the New Yorker or Wired, but somehow clinicians never picked it up.  The classic example is Ramachandran’s mirror box experiment — it’s a famous experiment that showed that phantom limb pain can be reversed by creating an illusion with mirrors that allows the patient to fix their “body map.” There have since been quite a few randomized trials confirming that the mirror trick works. But, maybe because it’s not a typical kind of “treatment” like a drug, it’s not standard of care for phantom limb pain.

I think we were pretty successful at finding these kinds of mismatches between medical science and medical practice.  By their nature, though, these kinds of solutions are hard to scale to reach lots of people.

N=1 Translational Medicine for Rare Diseases

This is the use case of “personalized medicine” that I think can really shine.  It harnesses the incredible motivation of patients with rare incurable diseases and their family members; it’s one of the few cases where genetic data really does make a huge difference; and the path to scale is (relatively) obvious if you discover a new drug or treatment.  I think we should have focused much more tightly on this angle, and that a company based on bench-to-bedside discovery for rare diseases could still become the real-world “Durona Group”.

I think doing it right at MetaMed would have meant getting a lot more in-house expertise in biology and medicine than we ever had, more like Perlara and Q-State, which have their own experimental research programs, something we never got off the ground.

Speaking only about myself and not my teammates, while I was at MetaMed I was deeply embarrassed to be a layman in the biomedical field, and I felt like “why would an expert ever want to work with a layman like me?” So I was far too reluctant to reach out to prominent biologists and doctors. I now know that experts work with laymen all the time, especially when that layman brings strategic vision, funding, and logistical/operational manpower, and listens to the expert with genuine curiosity.  Laymen are valuable — just ask Mary Lasker!  I really wish I’d understood this at the time.

People overestimate progress in the short run and underestimate it in the long run.  “Biohackers” and “citizen science” and “N=1 experimentation” have been around for a while, but they haven’t, I think, gotten very far along towards the ultimate impact they’re likely to have in the future.  Naively, that can look a lot like “a few people tried that and it didn’t seem to go anywhere” when the situation is actually “the big break is still ahead of us.”

New to LessWrong?

New Comment
16 comments, sorted by Click to highlight new comments since: Today at 5:44 AM

I believe that Prediction-based Medicine where a treatment provider has to give personalized predictions about the likely outcomes of a treatment is the most straightforward way to get effective personalized medicine.

When you allow patients to enter a lot of information about their issue on an online questionaire, those treatment providers who are actually confident that they can solve the medical issue, can be matched with the patient and they can charge good money for helping the patient because the patient knows the effectiveness of what they are being and think about what a given result would be worth for them.

I'd second this, but, to be fair, I think predictions are basically the answer to everything, so this may not be a big update.

I have to say I do really wish there were some kind of reliable, N=1 medical service out there for when something is wrong and it's not easy to diagnose let alone solve. I have a lot of personal experience in this area on the patient side, where a person close to me was (and still is!) suffering from some kind of medical problem and they keep getting bounced around because whatever is wrong is rare enough that it doesn't show up on anyone's flowchart. The experience is incredible frustrating, because I can see that there's something pretty specific wrong, but every time I or the patient talked to a doctor we'd go through the diagnostic process and, at best, end with "yep, idk what's wrong, let's just treat some symptoms then". I'd think that we'd be able to do better than this, but in the end most doctors just seem to throw up their hands and say "well, too hard for me, good luck". I get why it happens: it's a lot of work, they're not going to get paid extra for doing it, and no one is going to sue them as long as they made a best effort. But it doesn't make it any less frustrating, and any less interesting (to me) a problem to try to solve both for N=1 and for all the N=1s.

Curious if you've tried saying "I will pay extra for extra effort?" (My guess is that wouldn't fly at doctors-in-an-institution but might possibly work for private practices?)

No, because there's generally not an option for that via insurance since doing that would effectively be bribery under the way payment is handled. Have not tried doesn't-take-insurance private practice.

I'm used to the German medical system, but this surprises me. In Germany, doctors can offer patients on insurance additional services for which the insurance doesn't pay for direct payment without any problems.

I'm very glad to read disambiguations like this one.

(It has tentatively prompted me to write up one for all the different things that "rationality" can mean when one is doing "rationality development". We'll see if I get around to actually writing it up anytime soon, though.)

How does the bench-to-bedside model intersect with regulatory oversight?

It seems to me the central problem in the Durona Group vision is that you are largely prohibited from offering such a device to treat the rare condition, even if it gets built and works. I suspect that the optimizations a bench-to-bedside model would make to deliver the product quickly enough and cheaply enough to be useful to that patient would also leave the company drastically short of the amount of data usually demanded to establish safety/efficacy.

I feel like this idea would be a good match for the META Program concept, the goal of which was to speed up delivery of cyber-physical systems for the military by a factor of 5. The regulatory challenges of the FDA are considerable, but defense procurement is worse.

Curated. There were a few reasons I like this post:

1) I'm generally excited to have people who have spent years pursuing a complex goal or developing a complex skill, writing up their insights about it.

2) The object level of "how to think about personalized medicine, and which aspects of it are actually tractable" seems quite valuable. While there aren't that many people working on medical startups, it seems like the field in general is a mixture of "lots of obvious broken things" but "lots of non-obvious reasons why the things are broken that way." This post seems useful for focusing people's efforts in directions that are more likely to work.

3) On the meta level, this seemed like a good lens into Inadequate Equilibria. Even though I'm not working on a medical startup, the framework here feels helpful for looking at other messy-areas-full-of-broken-things that seem like you should be able to fix them, but with non-obvious reasons why fixing them in some ways is harder than other ways.

Through that last lens – one thing I might have appreciated here is more detail on how you came to believe the things you did about the various problem-areas. I'd be interested in a followup post that's something like "the general advice you wish you had at the beginning of MetaMed, that would have enabled you to more quickly figure out which areas to focus on."

> I think we were pretty successful at finding these kinds of mismatches between medical science and medical practice.  By their nature, though, these kinds of solutions are hard to scale to reach lots of people.

I'm curious about the cause of this. It seems like this is relatively straight forward to scale: Simply use marketing to get them to be more standard practice at hospitals and doctors offices.

There's two potential reasons I can imagine off the top of my head, but would really like to hear from you why these were so hard to scale.

1. They weren't hard to scale, but they were hard to make money with. If this is the case, maybe a non-profit could do it?

2. The fact that they weren't already standard practice meant that most of them had some other reason not to scale (the treatment was weird, it added extra liability, etc).

Was there some other reason these types of interventions wouldn't scale?

Simply use marketing to get them to be more standard practice at hospitals and doctors offices.

A lot of money is spent on pharmaceutical and medical device marketing, and it’s a crowded field. Occasionally someone who’s already very high profile like Atul Gawande can successfully promote things like the idea of having checklists at all, but in a substantial share of cases these quickly become distorted by hospital internal politics.

Giving a person a hypnotic induction before an operation means that you need to use less painkillers to sedate people and wounds heal a bit better.

Yet, it's not a standard procedure. I know two anesthetists who actually have the necessary hypnosis skills but who still don't use them when they sedate a patient for an operation.

I talked with one of them in more detail about it. According to him he has 15 minutes with a patient and in those 15 minutes burocratic documents have to be filled. There are additional pressures from his employer to be even faster.

In the end he hated his job and quit being a doctor. It wasn't possible for him in the enviroment in which he was to take 30 minutes to sedate a patient with added hypnosis even when he was confident that this would improve clinical outcomes.

It's even harder in cases where the relevant skills aren't there and they would need to hire additional expertise.

I spoke with a person who did new business development at a Big Pharma company a few times. According to him doctors generally only adopt new ways of treatment when there's something in it for the doctor.

Have you looked at any startups for mass market healthcare? Do any of them seem especially promising?

Do you have a citation for that thing about Tenenbaum?


Your link to the Matt Might article says:

A pharmacogenetic panel that will tell you your response essentially to every drug on the market costs on the order of $300.

Fuck, maybe I should do one for fun of myself! Do you know what this panel is? Any other things I should try in this vein? (I had some rare AF lung/heart problems as a kid, used to be depressed and am generally healthy now, but I'd love to learn more).

There are a handful of SNPs (mainly in Cytochrome P450) that predict how fast you metabolize drugs. This is a shortcut to dosing drugs. But the big problem is that people don't adjust their doses at all. The much more basic personalized medicine is to experiment with doses. If you aren't doing that, using the SNPs as a shortcut to predict where to start isn't much of an improvement.