[Epistemic status: very uncertain. Not to be taken as medical advice. Talk to your doctor before deciding whether or not to get any tests.]

I.

There are many antidepressants in common use. With a few exceptions, none are globally better than any others. The conventional wisdom says patients should keep trying antidepressants until they find one that works for them. If we knew beforehand which antidepressants would work for which patients, it would save everyone a lot of time, money, and misery. This is the allure of pharmacogenomics, the new field of genetically-guided medication prescription.

Everybody has various different types of cytochrome enzymes which metabolize medication. Some of them play major roles in metabolizing antidepressants; usually it’s really complicated and several different enzymes can affect the same antidepressant at different stages. But sometimes one or another dominates; for example, Prozac is mostly metabolized by one enzyme called CYP2D6, and Zoloft is mostly metabolized by a different enzyme called CYP2C19.

Suppose (say the pharmacogenomicists) that my individual genetics code for a normal CYP2D6, but a hyperactive CYP2C19 that works ten times faster than usual. Then maybe Prozac would work normally for me, but every drop of Zoloft would get shredded by my enzymes before it can even get to my brain. A genetic test could tell my psychiatrist this, and then she would know to give me Prozac and not Zoloft. Some tests like this are already commercially available. Preliminary results look encouraging. As always, the key words are “preliminary” and “look”, and did I mention that these results were mostly produced by pharma companies pushing their products?

But let me dream for a just a second. There’s been this uneasy tension in psychopharmacology. Clinical psychiatrists give their patients antidepressants and see them get better. Then research psychiatrists do studies and show that antidepressant effect sizes are so small as to be practically unnoticeable. The clinicians say “Something must be wrong with your studies, we see our patients on antidepressants get much better all the time”. The researchers counter with “The plural of anecdote isn’t ‘data’, your intuitions deceive you, antidepressant effects are almost imperceptibly weak.” At this point we prescribe antidepressants anyway, because – what else are you going to do when someone comes into your office in tears and begs for help? – but we feel kind of bad about it.

Pharmacogenomics offers a way out of this conundrum. Suppose half of the time patients get antidepressants, their enzymes shred the medicine before it can even get to the brain, and there’s no effect. In the other half, the patients have normal enzymes, the medications reach the brain, and the patient gets better. Researchers would average together all these patients and conclude “Antidepressants have an effect, but on average it’s very small”. Clinicians would keep the patients who get good effects, keep switching drugs for the patients who get bad effects until they find something that works, and say “Eventually, most of my patients seem to have good effects from antidepressants”.

There’s a little bit of support for this in studies. STAR*D found that only 33% of patients improved on their first antidepressant, but that if you kept changing antidepressants, about 66% of patients would eventually find one that helped them improve. Gueorguieva & Mallinckrodt (2011) find something similar by modelling “growth trajectories” of antidepressants in previous studies. If it were true, it would be a big relief for everybody.

It might also mean that pharmacogenomic testing would solve the whole problem forever and lets everyone be on an antidepressant that works well for them. Such is the dream.

But pharmacogenomics still very young. And due to a complicated series of legal loopholes, it isn’t regulated by the FDA. I’m mostly in favor of more things avoiding FDA regulation, but it means the rest of us have to be much more vigilant.

A few days ago I got to talk to a representative of the company that makes GeneSight, the biggest name in pharmacogenomic testing. They sell a $2000 test which analyzes seven genes, then produces a report on which psychotropic medications you might do best or worst on. It’s exactly the sort of thing that would be great if it worked – so let’s look at it in more depth.

II.

GeneSight tests seven genes. Five are cytochrome enzymes like the ones discussed above. The other two are HTR2A, a serotonin receptor, and SLC6A4, a serotonin transporter. These are obvious and reasonable targets if you’re worried about serotonergic drugs. But is there evidence that they predict medication response?

GeneSight looks at the rs6313 SNP in HTR2A, which they say determines “side effects”. I think they’re thinking of Murphy et al (2003), who found that patients with the (C,C) genotype had worse side effects on Paxil. The study followed 122 patients on Paxil, of whom 41 were (C,C) and 81 were something else. 46% of the (C,C) patients hated Paxil so much they stopped taking it, compared to only 16% of the others (p = 0.001). There was no similar effect on a nonserotonergic drug, Remeron. This study is interesting, but it’s small and it’s never been replicated. The closest thing to replication is this study which focused on nausea, the most common Paxil side effect; it found the gene had no effect. This study looked at Prozac and found that the gene didn’t affect Prozac response, but it didn’t look at side effects and didn’t explain how it handled dropouts from the study. I am really surprised they’re including a gene here based on a small study from fifteen years ago that was never replicated.

They also look at SLC6A4, specifically the difference between the “long” versus “short” allele. This has been studied ad nauseum – which isn’t to say anyone has come to any conclusions. According to Fabbri, Di Girolamo, & Serretti, there are 25 studies saying the long allele of the gene is better, 9 studies saying the short allele is better, and 20 studies showing no difference. Two meta-analyses (1 n = 1435, 2 n = 5479) come out in favor of the long allele; two others (1 n = 4309, 2, n = 1914) fail to find any effect. But even the people who find the effect admit it’s pretty small – the Italian group estimates 3.2%. This would both explain why so many people miss it, and relieve us of the burden of caring about it at all.

The Carlat Report has a conspiracy theory that GeneSight really only uses the liver enzyme genes, but they add in a few serotonin-related genes so they can look cool; presumably there’s more of a “wow” factor in directly understanding the target receptors in the brain than in mucking around with liver enzymes. I like this theory. Certainly the results on both these genes are small enough and weak enough that it would be weird to make a commercial test out of them. The liver enzymes seem to be where it’s at. Let’s move on to those.

The Italian group that did the pharmacogenomics review mentioned above are not sanguine about liver enzymes. They write (as of 2012, presumably based on Genetic Polymorphisms Of Cytochrome P450 Enzymes And Antidepressant Metabolism“>this previous review):

Available data do not support a correlation between antidepressant plasma levels and response for most antidepressants (with the exception of TCAs) and this is probably linked to the lack of association between response and CYP450 genetic polymorphisms found by the most part of previous studies. In all facts, the first CYP2D6 and CYP2C19 genotyping test (AmpliChip) approved by the Food and Drug Administration has not been recommended by guidelines because of lack of evidence linking this test to clinical outcomes and cost-effectiveness studies.

What does it even mean to say that there’s no relationship between SSRI plasma level and therapeutic effect? Doesn’t the drug only work when it’s in your body? And shouldn’t the amount in your body determine the effective dose? The only people I’ve found who even begin to answer this question are Papakostas & Fava, who say that there are complicated individual factors determining how much SSRI makes it from the plasma to the CNS, and how much of it binds to the serotonin transporter versus other stuff. This would be a lot more reassuring if amount of SSRI bound to the serotonin transporter correlated with clinical effects, which studies seem very uncertain about. I’m not really sure how to fit this together with SSRIs having a dose-dependent effect, and I worry that somebody must be very confused. But taking all of this at face value, it doesn’t really look good for using cytochrome enzymes predicting response.

I talked to the GeneSight rep about this, and he agreed; their internal tests don’t show strong effects for any of the candidate genes alone, because they all interact with each other in complicated ways. It’s only when you look at all of them together, using the proprietary algorithm based off of their proprietary panel, that everything starts to come together.

This is possible, but given the poor results of everyone else in the field I think we should take it with a grain of salt.

III.

We might also want to zoom out and take a broader picture: should we expect these genes to matter?

It’s much easier to find the total effect of genetics than it is to find the effect of any individual gene; this is the principle behind twin studies and GCTAs. Tansey et al do a GCTA on antidepressant response and find that all the genetic variants tested, combined, explain 42% of individual differences in antidepressant response. Their methodology allowed them to break it down chromosome-by-chromosome, and they found that genetic effects were pretty evenly distributed across chromosomes, with longer chromosomes counting more. This is consistent with massively polygenic structure where there are hundreds of thousands of genes, each of small effects – much like height or IQ. But typically even the strongest IQ or height genes only explain about 1% of the variance. So an antidepressant response test containing only seven genes isn’t likely to do very much even if those genes are correctly chosen and well-understood.

SLC6A4 is a great example of this. It’s on chromosome 17. According to Tansey, chromosome 17 explains less than 1% of variance in antidepressant effect. So unless Tansey is very wrong, SLC6A4 must also explain less than 1% of the variance, which means it’s clinically useless. The other six genes on the test aren’t looking great either.

Does this mean that the GeneSight panel must be useless? I’m not sure. For one thing, the genetic structure of which antidepressant you respond to might be different from the structure of antidepressant response generally (though the study found similar structures to any-antidepressant response and SSRI-only response). For another, for complicated reasons sometimes exploiting variance is easier than predicting variance; I don’t understand this enough to be sure that this isn’t one of these cases, though it doesn’t look that way to me.

I don’t think this is a knock-down argument against anything. But I think it means we should take any claims that a seven (or ten, or fifty) gene panel can predict very much with another grain of salt.

IV.

But assuming that there are relatively few genes, and we figure out what they are, then we’re basically good, right? Wrong.

Warfarin is a drug used to prevent blood clots. It’s notorious among doctors for being finicky, confusing, difficult to dose, and making people to bleed to death if you get it wrong. This made it a very promising candidate for pharmacogenomics: what if we could predict everyone’s individualized optimal warfarin dose and take out the guesswork?

Early efforts showed promise. Much of the variability was traced to two genes, VKORC1 and CYP2C9. Companies created pharmacogenomic panels that could predict warfarin levels pretty well based off of those genes. Doctors were urged to set warfarin doses based on the results. Some initial studies looked positive. Caraco et al and Primohamed et al both found in randomized controlled trials with decent sample sizes that warfarin patients did better on the genetically-guided algorithm, p < 0.001. A 2014 meta-analysis looked at nine studies of the algorithm, over 2812 patients, and found that it didn’t work. Whether you used the genetic test or not didn’t affect number of blood clots, percent chance of having your blood within normal clotting parameters, or likelihood of major bleeding. There wasn’t even a marginally significant trend. Another 2015 meta-analysis found the same thing. Confusingly, a Chinese group did a third meta-analysis that did find advantages in some areas, but Chinese studies tend to use shady research practices, and besides, it’s two to one.

UpToDate, the canonical medical evidence aggregation site for doctors, concludes:

We suggest not using pharmacogenomic testing (ie, genotyping for polymorphisms that affect metabolism of warfarin and vitamin K-dependent coagulation factors) to guide initial dosing of the vitamin K antagonists (VKAs). Two meta-analyses of randomized trials (both involving approximately 3000 patients) found that dosing incorporating hepatic cytochrome P-450 2C9 (CYP2C9) or vitamin K epoxide reductase complex (VKORC1) genotype did not reduce rates of bleeding or thromboembolism.

I mention this to add another grain of salt. Warfarin is the perfect candidate for pharmacogenomics. It’s got a lot of really complicated interpersonal variation that often leads to disaster. We know this is due to only a few genes, and we know exactly which genes they are. We understand pretty much every aspect of its chemistry perfectly. Preliminary studies showed amazing effects.

And yet pharmacogenomic testing for warfarin basically doesn’t work. There are a few special cases where it can be helpful, and I think the guidelines say something like “if you have your patient’s genotype already for some reason, you might as well use it”. But overall the promise has failed to pan out.

Antidepressants are in a worse place than warfarin. We have only a vague idea how they work, only a vague idea what genes are involved, and plasma levels don’t even consistently correlate with function. It would be very strange if antidepressant testing worked where warfarin testing failed. But, of course, it’s not impossible, so let’s keep our grains of salt and keep going.

V.

Why didn’t the warfarin pharmacogenomics work? They had the genes right, didn’t they?

I’m not too sure what’s going on, but maybe it just didn’t work better than doctors titrating the dose the old-fashioned way. Warfarin is a blood thinner. You can take blood and check how thin it is, usually measured with a number called INR. Most warfarin users are aiming for an INR between 2 and 3. So suppose (to oversimplify) you give your patient a dose of 3 mg, and find that the INR is 1.7. It seems like maybe the patient needs a little more warfarin, so you increase the dose to 4 mg. You take the INR later and it’s 2.3, so you declare victory and move on.

Maybe if you had a high-tech genetic test you could read the microscopic letters of the code of life itself, run the results through a supercomputer, and determine from the outset that 4 mg was the optimal dose. But all it would do is save you a little time.

There’s something similar going on with depression. Starting dose of Prozac is supposedly 20 mg, but I sometimes start it as low as 10 to make sure people won’t have side effects. And maximum dose is 80 mg. So there’s almost an order of magnitude between the highest and lowest Prozac doses. Most people stay on 20 to 40, and that dose seems to work pretty well.

Suppose I have a patient with a mutation that slows down their metabolism of Prozac; they effectively get three times the dose I would expect. I start them on 10 mg, which to them is 30 mg, and they seem to be doing well. I increase to 20, which to them is 60, and they get a lot of side effects, so I back down to 10 mg. Now they’re on their equivalent of the optimal dose. How is this worse than a genetic test which warns me against using Prozac because they have mutant Prozac metabolism?

Or suppose I have a patient with a mutation that dectuples Prozac levels; now there’s no safe dose. I start them on 10 mg, and they immediately report terrible side effects. I say “Yikes”, stop the Prozac, and put them on Zoloft, which works fine. How is this worse than a genetic test which says Prozac is bad for this patient but Zoloft is good?

Or suppose I have a patient with a mutation that makes them an ultrarapid metabolizer; no matter how much Prozac I give them, zero percent ever reaches their brain. I start them on Prozac 10 mg, nothing happens, go up to 20, then 40, then 60, then 80, nothing happens, finally I say “Screw this” and switch them to Zoloft. Once again, how is this worse than the genetic test?

(again, all of this is pretending that dose correlates with plasma levels correlates with efficacy in a way that’s hard to prove, but presumably necessary for any of this to be meaningful at all)

I expect the last two situations to be very rare; few people have orders-of-magnitude differences in metabolism compared to the general population. Mostly it’s going to be people who I would expect to need 20 of Prozac actually needing 40, or vice versa. But nobody has the slightest idea how to dose SSRIs anyway and we usually just try every possible dose and stick with the one that works. So I’m confused how genetic testing is supposed to make people do better or worse, as opposed to just needing a little more or less of a medication whose dosing is so mysterious that nobody ever knows how much anyone needs anyway.

As far as I can tell, this is why they need those pharmacodynamic genes like HTR2A and SLC6A4. Those represent real differences between antidepressants and not just changes in dose which we would get to anyway. I mean, you could still just switch antidepressants if your first one doesn’t work. But this would admittedly be hard and some people might not do it. Everyone titrates doses!

This is a fourth grain of salt and another reason why I’m wary about this idea.

VI.

Despite my skepticism, there are several studies showing impressive effects from pharmacogenomic antidepressant tests. Now that we’ve established some reasons to be doubtful, let’s look at them more closely.

GeneSight lists eight studies on its website here. Of note, all eight were conducted by GeneSight; as far as I know no external group has ever independently replicated any of their claims. The GeneSight rep I talked to said they’re trying to get other scientists to look at it but haven’t been able to so far. That’s fair, but it’s also fair for me to point out that studies by pharma companies are far more likely to find their products effective than studies by anyone else (OR = 4.05). I’m not going to start a whole other section for this, but let’s call it a fifth grain of salt.

First is the LaCrosse Clinical Study. 114 depressed patients being treated at a clinic in Wisconsin received the GeneSight test, and the results were given to their psychiatrists, who presumably changed medications in accordance with the tests. Another 113 depressed patients got normal treatment without any genetic testing. The results were:

Taken from here, where you’ll find much more along the same lines.

All of the combinations of letters and numbers are different depression tests. The blue bars are the people who got genotyped. The grey bars are the people who didn’t. So we see that on every test, the people who got genotyped saw much greater improvement than the people who didn’t. The difference in remission was similarly impressive; by 8 weeks, 26% of the genotyped group were depression-free as per QIDS-C16 compared to only 13% of the control group (p = 0.03)

How can we nitpick these results? A couple of things come to mind.

Number one, the study wasn’t blinded. Everyone who was genotyped knew they were genotyped. Everyone who wasn’t genotyped knew they weren’t genotyped. I’m still not sure whether there’s a significant placebo effect in depression (Hróbjartsson and Gøtzsche say no!), but it’s at least worth worrying about.

Number two, the groups weren’t randomized. I have no idea why they didn’t randomize the groups, but they didn’t. The first hundred-odd people to come in got put in the control group. The second hundred-off people got put in the genotype group. In accordance with the prophecy, there are various confusing and inexplicable differences between the two groups. The control group had more previous medication trials (4.7 vs. 3.6, p = 0.02). The intervention group had higher QIDS scores at baseline (16 vs. 17.5, p = 0.003). They even had different CYP2D6 phenotypes (p = 0.03). On their own these differences don’t seem so bad, but they raise the question of why these groups were different at all and what other differences might be lurking.

Number three, the groups had very different numbers of dropouts. 42 people dropped out of the genotyped group, compared to 20 people from the control group. Dropouts made up about a quarter of the entire study population. The authors theorize that people were more likely to drop out of the genotype group than the control group because they’d promised to give the control group their genotypes at the end of the study, so they were sticking around to get their reward. But this means that people who were failing treatment were likely to drop out of the genotype group (making them look better) but stay in the control group (making them look worse). The authors do an analysis and say that this didn’t affect things, but it’s another crack in the study.

All of these are bad, but intuitively I don’t feel like any of them should have been able to produce as dramatic an effect as they actually found. But I do have one theory about how this might have happened. Remember, these are all people who are on antidepressants already but aren’t getting better. The intervention group’s doctors get genetic testing results saying what antidepressant is best for them; the control group’s doctors get nothing. So the intervention group’s doctors will probably switch their patients’ medication to the one the test says will be best, and the control group’s doctors might just leave them on the antidepressant that’s already not working. Indeed, we find that 77% of intervention group patients switched medications, compared to 44% of control group patients. So imagine if the genetic test didn’t work at all. 77% of intervention group patients at least switch off their antidepressant that definitely doesn’t work and onto one that might work; meanwhile, the control group mostly stays on the same old failed drugs.

Someone (maybe Carlat again?) mentioned how they should have controlled this study: give everyone a genetic test. Give the intervention group their own test results, and give the control group someone else’s test results. If people do better on their own results than on random results, then we’re getting somewhere.

Second is the Hamm Study, which is so similar to the above I’m not going to treat it separately.

Third is the Pine Rest Study. This one is, at least, randomized and single-blind. Single-blind means that the patients don’t know which group they’re in, but their doctors do; this is considered worse than double-blind (where neither patients nor doctors know) because the doctors’ subtle expectations could unconsciously influence the patients. But at least it’s something.

Unfortunately, the sample size was only 51 people, and the p-value for the main outcome was 0.28. They tried to salvage this with some subgroup analyses, but f**k that.

Fourth and fifth are two different meta-analyses of the above three studies, which is the lowest study-to-meta-analysis ratio I’ve ever seen. They find big effects, but “garbage in, garbage out”.

Sixth, there’s the Medco Study by Winner et al; I assume his name is a Big Pharma plot to make us associate positive feelings with him. This study is an attempt to prove cost-effectiveness. The GeneSight test costs $2000, but it might be worth it to insurers/governments if it makes people so much healthier that they spend less money on health care later. And indeed, it finds that GeneSight users spend $1036 less per year on medication than matched controls.

The details: they search health insurance databases for patients who were taking an psychiatric medication and then got GeneSight tests. Then they search the same databases for control patients for each; the control patients take the same psych med, have the same gender, are similar in age, and have the same primary psychiatric diagnosis. They end up with 2000 GeneSight patients and 10000 matched controls, whom they prove are definitely similar (even as a group) on the traits mentioned above. Then they follow all these people for a year and see how their medication spending changes.

The year of the study, the GeneSight patients spent on average $689 more on medications than they did the year before – unfortunate, but not entirely unexpected since apparently they’re pretty sick. The control patients spent on average $1725 more. So their medication costs increased much more than the GeneSight patients. That presumably suggests GeneSight was doing a good job treating their depression, thus keeping costs down.

The problem is, this study wasn’t randomized and so I see no reason to expect these groups to be comparable in any way. The groups were matched for sex, age, diagnosis, and one drug, but not on any other basis. And we have reason to think that they’re not the same – after all, one group consists of people who ordered a little-known $2000 genetic test. To me, that means they’re probably 1) rich, and 2) have psychiatrists who are really cutting-edge and into this kind of stuff. To be fair, I would expect both of those to drive up their costs, whereas in fact their costs were lower. But consider the possibility that rich people with good psychiatrists probably have less severe disease and are more likely to recover.

Here’s some more evidence for this: of the ~$1000 cost savings, $300 was in psychiatric drugs and $700 was in non-psychiatric drugs. The article mentions that there’s a mind-body connection and so maybe treating depression effectively will make people’s non-psychiatric diseases get better too. This is true, but I think seeing that the effect of a psychiatric intervention is stronger on non-psychiatric than psychiatric conditions should at least raise our suspicion that we’re actually seeing some confounder.

I cannot find anywhere in the study a comparison of how much money each group spent the year before the study started. This is a very strange omission. If these numbers were very different, that would clinch this argument.

Seventh is the Union Health Service study. They genotype people at a health insurance company who have already been taking a psychotropic medication. The genetic test either says that their existing medication is good for them (“green bin”), okay for them (“yellow bin”) or bad for them (“red bin”). Then they compare how the green vs. yellow vs. red patients have been doing over the past year on their medications. They find green and yellow patients mostly doing the same, but red patients doing very badly; for example, green patients have about five sick days from work a year, but red patients have about twenty.

I don’t really see any obvious flaws in this study, but there are only nine red patients, which means their entire results depend on an n = 9 experimental group.

Eighth is a study that just seems to be a simulation of how QALYs might change if you enter some parameters; it doesn’t contain any new empirical data.

Overall these studies show very impressive effects. While it’s possible to nitpick all of them, we have to remind ourselves that we can nitpick anything, even the best of studies, and do we really want to be that much of a jerk when these people have tested their revolutionary new product in five different ways, and every time it’s passed with flying colors aside from a few minor quibbles?

And the answer is: yes, I want to be exactly that much of a jerk. The history of modern medicine is one of pharmaceutical companies having amazing studies supporting their product, and maybe if you squint you can just barely find one or two little flaws but it hardly seems worth worrying about, and then a few years later it comes out that the product had no benefits whatsoever and caused everyone who took it to bleed to death. The reason for all those grains of salt above was to suppress our natural instincts toward mercy and cultivate the proper instincts to use when faced with pharmaceutical company studies, ie Cartesian doubt mixed with smoldering hatred.

VII.

I am totally not above introducing arguments from authority, and I’ve seen two people with much more credibility than myself look into this. The first is Daniel Carlat, Tufts professor and editor of The Carlat Report, a well-respected newsletter/magazine for psychiatrists. He writes a skeptical review of their studies, and finishes:

If we were to hold the GeneSight test to the usual standards we require for making medication decisions, we’d conclude that there’s very little reliable evidence that it works.

The second is John Ioannidis, professor of health research at Stanford and universally recognized expert on clinical evidence. He doesn’t look at GeneSight in particular, but he writes of the whole pharmacogenomic project:

For at least 3 years now, the expectation has been that newer platforms using exome or full-genome sequencing may improve the genome coverage and identify far more variants that regulate phenotypes of interest, including pharmacogenomic ones. Despite an intensive research investment, these promises have not yet materialized as of early 2013. A PubMed search on May 12, 2013, with (pharmacogenomics* OR pharmacogenetc*) AND sequencing yielded an impressive number of 604 items. I scrutinized the 80 most recently indexed ones. The majority were either reviews/commentary articles with highly promising (if not zealot) titles or irrelevant articles. There was not a single paper that had shown robust statistical association between a newly discovered gene and some pharmacogenomics outcome, detected by sequencing. If anything, the few articles with real data, rather than promises, show that the task of detecting and validating statistically rigorous associations for rare variants is likely to be formidable. One comprehensive study sequencing 202 genes encoding drug targets in 14,002 individuals found an abundance of rare variants, with 1 rare variant appearing every 17 bases, and there was also geographic localization and heterogeneity. Although this is an embarrassment of riches, eventually finding which of these thousands of rare variants are most relevant to treatment response and treatment-related harm will be a tough puzzle to solve even with large sample sizes.

Despite these disappointing results, the prospect of applying pharmacogenomics in clinical care has not abided. If anything, it is pursued with continued enthusiasm among believers. But how much of that information is valid and is making any impact? […]

Before investing into expensive clinical trials for testing the new crop of mostly weak pharmacogenomic markers, a more radical decision is whether we should find some means to improve the yield of pharmacogenomics or just call it a day and largely abandon the field. The latter option sounds like a painfully radical solution, but on the other hand, we have already spent many thousands of papers and enormous funding, and the yield is so minimal. The utility yield seems to be even diminishing, if anything, as we develop more sophisticated genetic measurement techniques. Perhaps we should acknowledge that pharmacogenomics was a brilliant idea, we have learned some interesting facts to date, and we also found a handful of potentially useful markers, but industrial-level application of research funds may need to shift elsewhere.

I think the warning from respected authorities like these should add a sixth grain of salt to our rapidly-growing pile and make us feel a little bit better about rejecting the evidence above and deciding to wait.

VIII.

There’s a thing I always used to hate about the skeptic community. Some otherwise-responsible scientist would decide to study homeopathy for some reason, and to everyone’s surprise they would get positive results. And we would be uneasy, and turn to the skeptic community for advice. And they would say “Yeah, but homeopathy is stupid, so forget about this.” And they would be right, but – what’s the point of having evidence if you ignore it when it goes the wrong way? And what’s the point in having experts if all they can do is say “this evidence went the wrong way, so let’s ignore it”? Shouldn’t we demand experts so confident in their understanding that they can explain to us why the new “evidence” is wrong? And as a corollary, shouldn’t we demand experts who – if the world really was topsy-turvy and some crazy alternative medicine scheme did work – would be able to recognize that and tell us when to suspend our usual skepticism?

But at this point I’m starting to feel a deep kinship with skeptic bloggers. Sometimes we can figure out possible cracks in studies, and I think Part VI above did okay with that. But there will be cracks in even the best studies, and there will especially be cracks in studies done by small pharmaceutical companies who don’t have the resources to do a major multicenter trial, and it’s never clear when to use them as an excuse to reject the whole edifice versus when to let them pass as an unavoidable part of life. And because of how tough pharmacogenomics has proven so far, this is a case where I – after reading the warnings from Carlat and Ioannidis and the Italian team and everyone else – tentatively reject the edifice.

I hope later I kick myself over this. This might be the start of a revolutionary exciting new era in psychiatry. But I don’t think I can believe it until independent groups have evaluated the tests, until other independent groups have replicated the work of the first independent groups, until everyone involved has publicly released their data (GeneSight didn’t release any of the raw data for any of these studies!), and until our priors have been raised by equivalent success in other areas of pharmacogenomics.

Until then, I think it is a neat toy. I am glad some people are studying it. But I would not recommend spending your money on it if you don’t have $2000 to burn (though I understand most people find ways to make their insurance or the government pay).

But if you just want to have fun with this, you can get a cheap approximation from 23andMe. Use the procedure outlined here to get your raw data, then look up rs6313 for the HTR2A polymorphism; (G,G) supposedly means more Paxil side effects (and maybe SSRI side effects in general). 23andMe completely dropped the ball on SLC6A4 and I would not recommend trying to look that one up. The cytochromes are much more complicated, but you might be able to piece some of it together from this page’s links links to lists of alleles and related SNPs for each individual enzyme; also Promethease will do some of it for you automatically. Right now I think this process would produce pretty much 100% noise and be completely useless. But I’m not sure it would be more useless than the $2000 test. And if any of this pharmacogenomic stuff turns out to work, I hope some hobbyist automates the 23andMe-checking process and sells it as shareware for $5.