More Dakka in Medicine

By Sarah Constantin (blog – 1, 2)

The More Dakka story is common in medicine. You do an intervention; the disease doesn’t get better, or gets only marginally better; the research literature concludes it doesn’t work; nobody tries doing MORE of that intervention, but when somebody just raises the dose high enough, it does work.

Examples:

a.) Chemotherapy didn’t work on cancer until doctors made cocktails of drugs, raised the dose so high it would kill you, and then mitigated the side effects with prednisone and intermittent dosing schedules. If they just used a safe daily dose of a single chemotherapeutic agent, they’d have concluded chemo didn’t work.

Prednisone-2D-skeletal

Prednisone

b.) Light therapy barely works for SAD; two internet-famous people have independently found that REALLY BRIGHT light therapy completely fixes SAD.

c.) The example in the post is about allopurinol. Allopurinol prevents gout attacks by lowering uric acid. “In studies, [allopurinol] improved [uric acid] linearly with dosage. Studies observed that sick patients whose [uric acid] reached healthy levels experienced full remission. The treatment was fully safe. No one tried increasing the dose enough to reduce [uric acid] to healthy levels.

d.) The standard treatment for hypothyroidism is thyroid hormone. People with “subclinical hypothyroidism”– people whose thyroid hormone levels are lower than average, but still above the cutoff for hypothyroid, and still suffer from exactly the same symptoms as hypothyroid–, ALSO benefit from thyroid hormone therapy. It’s not standard of care yet, though.

e.) I believe some vitamin deficiencies, don’t remember which exactly, are the same way; there’s an official cutoff for “deficient” but people slightly above that cutoff still have symptoms and still experience symptom relief from supplementation.

f.) Same deal with HIV. Virus has a replication rate & a clearance rate; its replication rate is also its mutation rate; an antiviral drug can raise the clearance rate above the replication rate, which will make the population drop exponentially, but if there’s only one drug the virus will have a chance to evolve to be resistant before the population drops low enough to be undetectable. And this is a simple differential equation that you can calculate years before you know what the drugs even are. One drug: death. Two drugs: death. Three or more drugs: survival.

Luckily David Ho was a physicist and thought about it this way, so when the antiviral drugs came out he was ready to test them in cocktails.

So “single antibiotics don’t work for chronic Lyme but cocktails do and this wasn’t realized for decades” isn’t an unprecedented story. It could turn out that way.

I bet this is something that has a more formal and accurate phrasing, but: if there’s an exponential-growth dynamic (like in a malignant cancer or an infection) where you’re trying to kill the exponentially-growing population, and if there’s a dose-response relationship where higher dose = more killing, then you have a bifurcation point in the outcome as t -> infinity, where a dose below that point means the enemy takes over and the patient dies and a dose above that point means “the enemy is killed faster than it can reproduce and so dies out in the long run.” And in principle you can calculate this cutoff if you know the dose-response relationship, as Ho did.

And separately, there’s a safety threshold; is the minimum effective dose safe or unsafe? With chemotherapy, the minimum effective dose is UNSAFE, which is why they have to get clever with ways to give you doses high enough to kill you while keeping you alive anyway. (Or “find a better drug”, but nobody has found a cytotoxic drug with strictly better tolerability/effectiveness tradeoffs since the 1960’s.)

This is kinda how you get a continuous/analog system to give you discrete outcomes: bifurcation points! Works in gene regulation too. “This regulatory gene turns on that gene’s transcription” – well, what’s actually happening is a continuous scalar, a rate of transcription and a rate of clearance, but because exponential functions are involved you get bifurcations in “steady-state” outcomes over the several-hour timescales needed to get to “this cell has tons of mRNAs for that gene or it’s literally empty of them”.

Systems biology is cool, it explains the math that gets you from a statistical-chemistry model of the cell (as a bag of molecules that bump into each other and have a probability of interaction) to a tinkertoy model that you can treat like a graph. (Gene regulatory networks, protein-protein interaction networks, neuron networks, etc.)

Coffee Saves Lives

[July 18 2019 addendum: This assumes that coffee has a causal- rather than merely correlational- influence on longevity. See comment section for more details.]


The T-shirt in the featured image was probably designed as a joke, but I take it very seriously.

Indeed, I think there is a strong case to be made that subsidizing coffee could be seen as an Effective Altruist priority. You see, you can save a life with coffee for as little as $50k. This makes coffee an intervention that is on par with some of the top charities in the world, and it is an outlier when it comes to the cost-benefit ratio of medical interventions. Consider how, e.g. this article on QALY states that:

“The UK’s recommendations, for example, are about £20,000 to £30,000 ($30,000 to $45,000) for each additional year of good health, once it has been adjusted to take into account the quality of life. So a drug that achieved 0.5 on the QALY measure would only merit £10,000-15,000 ($15,000 to $22,500).”

Assuming a QALY-adjusted average life-span of about 60 years per person, coffee is about 30 to 50 times more cost-effective than the types of medical interventions the UK is willing to subsidize to extend people’s lives. And that’s not even considering what people themselves are willing to pay to extend their own lives, which is, of course, a lot more than what a government would.

Relative to GiveWell‘s top charities this is still not the best intervention out there (with some of the ultra-effective charities saving a life for about 2,000 dollars). I would nonetheless point out that the ultra-effective charities out there are all effective because they address populations where very basic human needs are not typically met. In Malaria-ridden, war-torn areas, a little can go a long way. But what’s different about coffee is that it is as effective everywhere in the world. Sure, you can save a life with $50k in many African countries. But can you do so in Sweden?! With coffee you can!

Anyhow, how did I arrive at these numbers? Well consider that you can get about 380 doses of coffee for as little as 10 dollars.*

So this means you can have a cup of coffee for as little as 2.63 cents(!). In turn, we know from a lot of research that each cup of coffee up to 4 cups a day prevents about 1/2 micromorts (interestingly, it is just as cost-effective to encourage people who don’t drink coffee to drink 1 cup as it would be to encourage people who drink 3 to go ahead and drink 4).

Given those numbers, we have that the cost of a full life-span worth of micromorts is about $52,631.58.

Why are we not funding this?!


*With: Example 48 oz brand. (we could do even better buying in bulk – I reached out to a delivery company to get a quota and will update when I know more).

Triple S Genetic Counseling: Predicting Hedonic-Set Point with Commercial-Grade DNA Testing as an Effective Altruist Project

The term “Transhumanism” has many senses. It is a social movement, a philosophy, a set of technologies, and a conceptual rallying flag. David Pearce pins down the core sentiment behind the term like this:

If we get things right, the future of life in the universe can be wonderful beyond the bounds of human imagination: a “triple S” civilisation of superlongevity, superintelligence and superhappiness.

– David Pearce, in The 3 Supers

The concept of a “triple S” civilization is very widely applicable. For example, one can imagine future smart homes designed with it in mind. Such smart homes would have features to increase your longevity (HEPA filters, humidity control, mold detectors, etc.), increase your intelligence (adaptive noise-canceling, optimal lighting, smart foods), and happiness (mood-congruent lighting, music, aromas, etc.). Since there are trade-offs between these dimensions, one could specify how much one values each of them in advance, and the smart home would be tasked with maximizing a utility function based on a weighted average between the three S’s.

Likewise, one could apply the “triple S” concept to medical care, lifestyle choices, career development, governance, education, etc. In particular, one could argue that a key driver for the realization of a triple S civilization would be what I’d like to call “triple S genetic counseling.” In brief, this is counseling for prospective parents in order to minimize the risks of harming one’s children by being oblivious to the possible genetic risk for having a reduced longevity, intelligence, or happiness. Likewise, in the more forward-looking transhumanist side of the equation, triple S genetic counseling would allow parents to load the genetic dice in their kid’s favor in order to make them as happy, long-lived, and smart as possible.

Genetic counseling, as an industry, is indeed about to explode (cf. Nature’s recent article: Prospective parents should be prepared for a surge in genetic data). Predictably, there will be a significant fraction of society that will question the ethics of e.g. preimplantation genetic diagnosis for psychological traits. In practice, parents who are able to afford it will power ahead, for few prospective parents truly don’t care about the (probabilistic) well-being of their future offspring. My personal worry is not so much that this won’t happen, but that the emphasis will be narrow and misguided. In particular, both predicting health and intelligence based on sequenced genomes are very active areas of research. I worry that happiness will be (relatively) neglected. Hence the importance of emphasizing all three S’s.

In truth, I think that predicting the hedonic set-point of one’s potential future kids (i.e. the average level of genetically-determined happiness) is a relatively more important project than predicting IQ (cf. A genome-wide association study for extremely high intelligenceBGI). In addition, I anticipate that genetic-based models that predict a person’s hedonic set-point will be much more accurate than those that predict IQ. As it turns out, IQ is extremely polygenetic, with predictors diffused across the entire genome, and it is a very evolutionary recent axis of variance across the population. Predictors of hedonic-set point (such as the “pain-knob gene” SCN9A and it’s variants), on the other hand, are ancient and evolutionarily preserved across the phylogenetic tree. This makes baseline happiness a likely candidate for having a straight-forward universal physiological implementation throughout the human population. Hence my prediction that polygenetic scores of hedonic-set point will be much more precise than those for IQ (or even longevity).

Given all of the above, I would posit that a great place to start would be to develop a model that predicts hedonic set-point using all of the relevant SNPs offered by 23andMe*.  Not only would this be “low-hanging fruit” in the field of genetic counseling, it may also be a project that is way up there, close to the top of the “to do” list in Effective Altruism (cf. Cause X; Google Hedonics).

I thought about this because I saw that 23andMe reports on health predispositions based on single SNPs. From a utilitarian point of view, of particular interest are SNPs related to the SCN9A gene. For example, I found that 23andMe has the rs6746030 SNP, which some studies show can account for a percentage of the variance associated with pain in Parkinson’s and other degenerative diseases. The allele combination A/A is bad, making you more prone to experience pain intensely. This is just one SNP, though, and there ought to be a lot of other relevant SNPs, not only of the SCN9A gene but elsewhere too (e.g. involved in MAO enzymes, neuroplasticity, and pleasure centers innervation).

Concretely, the task would involve making two models and then combining them:

The first model uses people’s responses to 23andMe surveys to come up with a good estimate of a person’s hedonic set-point. Looking at some of the questions they ask, I would argue that there are more than enough dimensions to model how people vary in their hedonic set-point. They ask about things such as perception of pain, perception of spiciness, difficulty sleeping, stress levels, whether exercise is pleasant, etc. From a data science point of view, the challenge here is that number of responses provided by each participant is very variable; some power users respond to every question (and there are hundreds and hundreds), while most people respond to a few questions only, and a substantial minority respond to no questions at all. Most likely, the distribution of responses per participant follows a power law. So the model to build here has to be resilient against absent data. This is not an insurmountable problem, though, considering the existence of Bayesian Networks, PGMs, and statistical paradigms like Item Response Theory. For this reason, the model would need to both predict the most likely hedonic set-point of each participant, and provide confidence intervals specific to the participant based on the quality and relevance of the questions answered.

The second model would involve clustering and dimensionality reduction applied to the SNPs that are likely to be relevant for hedonic set-point. For example, one dimension would likely be a cluster of SNPs that are associated with “maximum intensity of pain”, another might be “how quickly pain subsides once it’s stimulated”, another “how much does pleasure counter-balance pain”, and so on. Each of these dimensions is likely to be determined by different neural circuits, and interact in non-linear ways, so they deserve their own separate dimension.

And finally, one would make a third model that combines the two models above, which predicts the hedonic set-point of a person derived from the first model using the genetic dimensions found by the second model. If you are an up-and-coming geneticist, I would like to nudge you in the direction of looking into this. As a side effect, you might as well get filthy rich in the process, as the genetic counseling field explodes in the next decade.


Bonus Content: What About Us?

Admittedly, many people will note that predicting a fraction of the variance of people’s hedonic set point with commercial DNA testing products will only really alleviate suffering in the medium to long term. The people who will benefit from this technology haven’t been born yet. In the meantime, what do we do about the people who currently have low hedonic set-points? Here is a creative, politically incorrect, and enticing idea:

Let’s predict which recreational drugs have the best cost-benefit profile for individuals based on their genetic makeup.

It is no secret that people react differently to drugs. 23andMe, among others, is currently doing research to predict your particular reaction to a drug based on your genetic makeup (cf. 23andMe can now tell you how you’ll react to 50+ common drugs). Unfortunately for people with anxiety, depression, chronic pain, and other hedonic tone illnesses, most psychiatric drugs are rather subtle and relatively ineffective. No wonder, compared to heroin, an SSRI is not likely to make you feel particularly great. As David Pearce argued in his essay Future Opioids, there is substantial evidence that many people who become addicts are driven to take recreational substances due to the fact that their endogenous opioid system is dysfunctional (e.g. they may have bad variants of opioid receptors, too many endorphin-degrading enzymes, etc.). The problem with giving people hard drugs is not that they don’t work in the short term; it is that they tend to backfire in the long-term and have cumulative negative health effects. As an aside, from the pharmaceutical angle, my main interest is the development of Anti-tolerance Drugs, which would allow hard drugs to work as mood-enhancers indefinitely.

This is not to say that there aren’t lucky people for whom the cost-benefit ratio of taking hard drugs is, in fact, rather beneficial. In what admittedly must have been a tongue-in-cheek marketing move, in the year 2010 the genetic interpretation company Knome (now part of Tute Genomics) studied Ozzy Osbourne‘s entire genome in order to determine how on earth he has been able to stay alive despite the gobs and gobs of drugs he’s taken throughout his life. Ozzy himself:

“I was curious, [g]iven the swimming pools of booze I’ve guzzled over the years—not to mention all of the cocaine, morphine, sleeping pills, cough syrup, LSD, Rohypnol…you name it—there’s really no plausible medical reason why I should still be alive. Maybe my DNA could say why.”

Ozzy Osbourne’s Genome (Scientific American, 2010)

Tentatively, Knome scientists said, Ozzy’s capacity to drink entire bottles of Whisky and Gin combined with bowlfuls of cocaine and multiple packs of cigarettes over the course of… breakfast… without ending up in the hospital may be due to novel mutations in his alcohol dehydrogenase gene (ADH4), as well as, potentially, the gene that codes for CLTCL1, a protein responsible for the intake of extra-cellular material into the cell’s inside. These are wild speculations, to be clear, but the general idea is brilliant.

Indeed, not everyone reacts in the same way to recreational drugs. A recent massive study on the health effects of alcohol funded by the Bill and Melinda Gates Foundation (cf. No amount of alcohol is good for your overall health) suggests that alcohol is bad for one’s health at every dosage. This goes against the common wisdom backed up with numerous studies that light-drinkers (~1 alcohol unit a day) live longer and healthier lives than teetotalers. The new study suggests that this is not a causal effect of alcohol. Rather, it so happens that a large fraction of teetotalers are precisely the kind of people who react very badly to alcohol as a matter of poor metabolism. Hence, teetotalers are not unhealthy because they avoid alcohol; they avoid alcohol because they are unhealthy, which explains their shorter life expectancy on average. That said, the study did show that 1 alcohol unit a day is, although damaging, very minimally so:

Anyhow, the world’s cultural fascination with alcohol is bizarre to me, considering the existence of drugs that have a much better hedonic and cost-benefit profile (cf. State-Space of Drug Effects). Perhaps finding out with genetic testing that you are likely to be an above-average alcohol metabolizer might be good to lessen your worry about having a couple of drinks now and then. But the much bigger opportunity here would be to allow you to find drugs that you are particularly compatible with. For example, a genetic test might determine based on a polygenetic score that you might benefit a whole lot from taking small amounts of e.g. Khat  (or some such obscure and relatively benign euphoriant). That is, that your genetic make-up is such that Khat will be motivation enhancing, empathy-increasing, good for your heart and lungs, reduce the rate of dopamine neuron death, etc. while at the same time producing little to no hangovers, no irritability, no sleep issues, or social dysfunction. Even though you may have thought that you are “not an uppers person”, perhaps that’s because, genetically, every other upper you have ever tried is objectively terrible for your health. But Khat wouldn’t be. Wouldn’t this information be useful? Indeed, I would posit, this might be a great step in the right direction in order to achieve the goal of  Wireheading Done Right.


*23andMe is here used as a shorthand for services in general like this (including Ancestry, Counsyl, Natera, etc.)

Featured image credit: source.