Personalized Medicine For Real

Link post

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. 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.”