Project Proposal: Gears of Aging

Imag­ine build­ing an air­plane the way we do biol­ogy re­search.

Hun­dreds of sep­a­rate re­search groups each pick differ­ent air­plane parts to work on. There’s no par­tic­u­lar co­or­di­na­tion in this; peo­ple choose based first on what they know how to build, sec­ond on what seems fun/​in­ter­est­ing/​flashy/​fash­ion­able, and fi­nally on what they ex­pect to be use­ful (based on limited lo­cal knowl­edge). Funds are al­lo­cated to any­thing which sounds vaguely re­lated to air­planes. There might be fifty differ­ent groups all build­ing pro­pel­lers, one guy toiling away at gy­ros and no­body at all on fuel lines. One group is build­ing an au­topi­lot sys­tem which could be handy but isn’t re­ally nec­es­sary; oth­ers are build­ing things which won’t be use­ful at all but they don’t re­al­ize they won’t be use­ful.

There’s ob­vi­ously room to gen­er­ate value by as­sem­bling all the parts to­gether, but there’s more to it than that. It’s not just that no­body is as­sem­bling the parts, there isn’t even a plan to as­sem­ble them. No­body’s re­ally sure what parts are even be­ing pro­duced, and no­body has a com­pre­hen­sive list of parts needed in or­der to build an air­plane. If things are miss­ing, no­body knows it. If some work is ex­tra­ne­ous, no­body knows that ei­ther; no­body knows what the min­i­mum vi­able path to an air­plane looks like. There is no air­plane blueprint.

This is what the large ma­jor­ity of ag­ing re­search looks like. There’s hun­dreds of differ­ent groups each study­ing spe­cific sub­sys­tems. There’s lit­tle co­or­di­na­tion on what is stud­ied, few-if-any peo­ple as­sem­bling the parts on a large scale, and noth­ing like a blueprint.

The even­tual vi­sion of the Gears of Aging pro­ject is to cre­ate a blueprint.

What Does That Look Like?

A blueprint does not have to in­clude all the in­ter­nals of ev­ery sin­gle sub­sys­tem in com­pre­hen­sive de­tail. The idea is to in­clude enough de­tail that we can calcu­late whether the air­plane will fly un­der var­i­ous con­di­tions.

Like­wise, a blueprint-analogue for ag­ing should in­clude enough de­tail that we can calcu­late whether a given treat­ment/​in­ter­ven­tion will cure var­i­ous age-re­lated dis­eases, and in­de­pen­dently ver­ify each of the model’s as­sump­tions.

Such a calcu­la­tion doesn’t nec­es­sar­ily in­volve a lot of nu­mer­i­cal pre­ci­sion. If we know the root cause of some age-re­lated dis­ease with high con­fi­dence, then we can say that re­vers­ing the root cause would cure the dis­ease, with­out do­ing much math. On the other hand, we prob­a­bly do need at least some quan­ti­ta­tive pre­ci­sion in or­der to be highly con­fi­dent that we’ve iden­ti­fied the root cause, and haven’t missed any­thing im­por­tant.

Like an air­plane blueprint, the goal is to show how all the com­po­nents con­nect—a sys­tem-level point of view. Much re­search has already been pub­lished on in­di­vi­d­ual com­po­nents and their lo­cal con­nec­tions—any­thing from the elastin → wrin­kles con­nec­tion to the thymic in­volu­tion → T-cell ra­tio con­nec­tion to the stress → sir­tu­ins → het­e­rochro­matin → ge­nomic in­sta­bil­ity path­way. A blueprint should sum­ma­rize the key pa­ram­e­ters of each lo­cal com­po­nent and its con­nec­tions to other com­po­nents, in a man­ner suit­able for trac­ing whole chains of cause-and-effect from one end to the other.

Most im­por­tantly, a blueprint needs some de­gree of com­pre­hen­sive­ness. We don’t want to model the en­tirety of hu­man phys­iol­ogy, but we need a com­plete end-to-end model of at least some age-re­lated dis­eases. The more dis­eases we can fully model, from root cause all the way to ob­served pathol­ogy, the more use­ful the blueprint will be.

Sum­mary: a blueprint-analogue for ag­ing would walk through ev­ery causal link, from root cause to pathol­ogy, for one or a few age-re­lated dis­eases, in enough de­tail to calcu­late whether a given in­ter­ven­tion would ac­tu­ally cure the dis­ease. See the Gears vs Aging se­quence so far for some early pieces work­ing in that di­rec­tion.

What’s the Value-Add?

Why would a blueprint be use­ful?

I’d phrase the key fea­ture as “ver­ti­cal com­pre­hen­sive­ness”, in anal­ogy to ver­ti­cal in­te­gra­tion in eco­nomics. It’s map­ping out ev­ery step of the causal chain from root cause to pathol­ogy—the whole “pro­duc­tion chain” of one or a few patholo­gies.

To see why this is use­ful, let’s com­pare it to the dual fea­ture: hori­zon­tal com­pre­hen­sive­ness. A good ex­am­ple here is the SENS pro­ject: a pro­gram to pre­vent ag­ing by cat­a­logu­ing ev­ery po­ten­tial root cause, and reg­u­larly re­pairing each of them. This is a purely-hori­zon­tal ap­proach: it does not re­quire any un­der­stand­ing at all of the causal path­ways from root causes to patholo­gies, but it does re­quire a com­pre­hen­sive cat­a­logue of ev­ery root cause.

SENS re­quires find­ing root causes, while a blueprint re­quires full path­ways. Note that this di­a­gram is a loose anal­ogy; ac­tual biolog­i­cal sys­tems do not look like this.

The rel­a­tive dis­ad­van­tage of a SENS-style hori­zon­tal ap­proach is that there’s no way to check it lo­cally. If it turns out that we missed a root cause, SENS has no built-in way to no­tice that un­til the whole pro­ject is done and we no­tice some pathol­ogy which hasn’t been fixed. Con­versely, if we mis­tak­enly in­cluded a root cause which doesn’t mat­ter, we have no built-in way to no­tice that at all; we waste re­sources fix­ing some ex­tra­ne­ous prob­lem. For ex­am­ple, here’s the origi­nal list of low-level dam­age types for SENS to ad­dress (from the Wikipe­dia page):

  • Ac­cu­mu­la­tion of lyso­so­mal aggregates

  • Ac­cu­mu­la­tion of senes­cent cells

  • Age re­lated tumors

  • Mi­to­chon­drial DNA mutations

  • Im­mune sys­tem damage

  • Buildup of ad­vanced gly­ca­tion end-products

  • Ac­cu­mu­la­tion of ex­tra­cel­lu­lar pro­tein aggregates

  • Cell loss

  • Hor­monal mus­cle damage

Note the in­clu­sion of senes­cent cells. To­day, it is clear that senes­cent cells are not a root cause of ag­ing, since they turn over on a timescale of days to weeks. Se­nes­cent cells are an ex­tra­ne­ous tar­get. Fur­ther­more, since senes­cent cell counts do in­crease with age, there must also be some root cause up­stream of that in­crease—and it seems un­likely to be any of the other items on the origi­nal SENS list. Some root cause is miss­ing. If we at­tempted to ad­dress ag­ing by re­mov­ing senes­cent cells (via senolyt­ics), what­ever root cause in­duces the in­crease in senes­cent cells in the first place would pre­sum­ably con­tinue to ac­cu­mu­late, re­quiring ever-larger doses of senolyt­ics un­til the senolytic dosage it­self ap­proached tox­i­c­ity—along with what­ever other prob­lems the root cause in­duced.

This isn’t to bash the SENS pro­gram; I’m per­son­ally a fan of it. The point is that the SENS pro­gram lacks a built-in way to cheaply ver­ify its plan. It needs to rely on other kinds of re­search in or­der to make sure that its list of tar­gets is com­plete and min­i­mal.

Con­versely, built-in ver­ifi­ca­tion is ex­actly where ver­ti­cal com­pre­hen­sive­ness shines.

When we have a full causal path­way, we can ask at each step:

  • Does this causal re­la­tion­ship ac­tu­ally hold?

  • Do the im­me­di­ate causes ac­tu­ally change with age by the right amount to ex­plain the ob­served effects?

Be­cause we can do this lo­cally, at each step of the chain, we can ver­ify our model as we go. Much like a math­e­mat­i­cal proof, we can check each step of our model along the way; we don’t need to finish the en­tire pro­ject in or­der to check our work.

In par­tic­u­lar, this gives us a nat­u­ral mechanism to no­tice miss­ing or ex­tra­ne­ous pieces. Check­ing whether senes­cent cells are ac­tu­ally a root cause is au­to­mat­i­cally part of the ap­proach. So is figur­ing out what’s up­stream of their age-re­lated in­crease in count. If there’s more than one fac­tor up­stream of some pathol­ogy, we can au­to­mat­i­cally de­tect any we missed by quan­ti­ta­tively check­ing whether the ob­served change in causes ac­counts for the ob­served change in effects.

Sum­mary: the main value-add of a blueprint-style end-to-end model is that we can lo­cally ver­ify each link in the causal chain, usu­ally us­ing already-ex­ist­ing data.

Is It Tractable?

I think that the data re­quired to figure out the gears of most ma­jor hu­man age-re­lated dis­eases is prob­a­bly already available, on­line, to­day. And I don’t mean that in the sense of “a su­per­in­tel­li­gent AI could figure it out”; I mean that hu­mans could prob­a­bly figure it out with­out any more data than we cur­rently have.

That be­lief stems mainly from hav­ing dug into the prob­lem a fair bit already. Every­where I look, there’s plenty of data. Some­one has ex­per­i­men­tally tested, if not the ex­act thing I want to know, at least some­thing close enough to provide ev­i­dence.

The hard part is not lack of data, the hard part is too much data. There’s more than a hu­man could ever hope to work through, for each ma­jor sub­sys­tem. It’s all about figur­ing out which ques­tions to ask, guess­ing which ex­per­i­ments could provide ev­i­dence for those ques­tions and are likely to have already been done, then track­ing down the re­sults from those ex­per­i­ments.

So I think the data is there.

The other piece of tractabil­ity is whether the sys­tem is sim­ple enough, on some level, that a hu­man can hope to un­der­stand all the key pieces. Based on hav­ing seen a fair bit, I definitely ex­pect that it is sim­ple enough—not sim­ple, there are a lot of mov­ing pieces and figur­ing them all out takes a fair bit of work, but still well within hu­man ca­pac­ity. We could also make some out­side view ar­gu­ments sup­port­ing this view—for in­stance, since the vast ma­jor­ity of molecules/​struc­tures/​cells in a hu­man turn over on much faster timescales than ag­ing, there are un­likely to be more than a hand­ful of in­de­pen­dent root causes.

Out­side-View Tractability

If a pro­ject like this is both use­ful and tractable, why hasn’t it already been done?

The usual aca­demic out­let for a blueprint-style ver­ti­cally-com­pre­hen­sive work would be a text­book. And there are text­books on ag­ing, as well as mono­graphs, and of course books on var­i­ous subtopics as well. Un­for­tu­nately, the field is still rel­a­tively young, and text­book-writ­ing tends to be un­der-in­cen­tivized in the sci­ences; most aca­demic hiring and tenure com­mit­tees pre­fer origi­nal re­search. Even those text­books which do ex­ist tend to ei­ther in­volve a broad-but-shal­low sum­mary of ex­ist­ing re­search (for sin­gle-au­thor books) or stan­dalone es­says on par­tic­u­lar com­po­nents (for multi-au­thor mono­graphs). They are part cat­a­logues, not blueprints.

But the biggest short­com­ing of typ­i­cal text­books, com­pared to the blueprint-style pic­ture, is that typ­i­cal text­books do not ac­tu­ally perform the lo­cal ver­ifi­ca­tion of model com­po­nents.

This is ex­actly the sort of prob­lem where we’d ex­pect a ra­tio­nal­ist skil­lset—statis­tics, causal­ity, notic­ing con­fu­sion, mys­te­ri­ous an­swers, etc—to be more of a limit­ing fac­tor than biolog­i­cal know-how. Add that to the lack of in­cen­tive for this sort of work, and it’s not sur­pris­ing that it hasn’t been done.

A hand­ful of ex­am­ples, to illus­trate the sort of rea­son­ing which is lack­ing in most books on ag­ing:

  • Many re­view ar­ti­cles and text­books claim that the in­creased stiff­ness of blood ves­sels in old age re­sults (at least par­tially) from an in­crease in the amount of col­la­gen rel­a­tive to elastin in ves­sel walls. But if we go look for stud­ies which di­rectly mea­sure the col­la­gen:elastin ra­tio in the blood ves­sels, we mostly find no sig­nifi­cant change with age (rat, hu­man, rat).

  • Many re­views and text­books men­tion that the bulk of re­ac­tive oxy­gen species (ROS) are pro­duced by mi­to­chon­dria. At­tempts at di­rect mea­sure­ment in­stead sug­gest that mi­to­chon­dria ac­count for about 15% (PhysAging, table 5.3).

  • In 1991, a small-count ge­netic study sug­gested that amy­loid pro­tein ag­gre­gates in the brain cause Alzheimers. Notably, they “con­firmed di­ag­noses via au­topsy”—which usu­ally means check­ing the brain for amy­loid de­posits. At least as early as 2003, it was known that amy­loid de­posits turn over on a timescale of hours. Yet, ac­cord­ing to Wikipe­dia, over 200 clini­cal tri­als at­tempted to cure Alzheimers by clear­ing plaques be­tween 2002 and 2012; only a sin­gle trial ended in FDA ap­proval, and we still don’t have a full cure.

  • A great deal of effort has gone into imag­ing neu­ro­mus­cu­lar junc­tions in ag­ing or­ganisms. As far as I can tell, there was never any sig­nifi­cant ev­i­dence that the ob­served changes played any sig­nifi­cant causal role in any age-re­lated dis­ease. They did pro­duce re­ally cool pic­tures, though.

Th­ese are the sorts of things which jump out when we ask, for ev­ery link in a hy­poth­e­sized causal chain:

  • Does this causal re­la­tion­ship ac­tu­ally hold?

  • Do the im­me­di­ate causes ac­tu­ally change with age by the right amount to ex­plain the ob­served effects?


I think that the data re­quired to figure out the gears of most ma­jor hu­man age-re­lated dis­eases is prob­a­bly already available, on­line, to­day. The parts to build an air­plane are already on the mar­ket. We lack a blueprint: an end-to-end model of age-re­lated patholo­gies, con­tain­ing enough de­tail for each causal link in the chain to be in­de­pen­dently val­i­dated by ex­per­i­men­tal and ob­ser­va­tional data, and suffi­cient to calcu­late whether a given in­ter­ven­tion will ac­tu­ally cure the dis­ease.