Sign In

Remember Me

AIs, Superflies, and the Path to Immortality


Summary: Why do we suffer poor health, disease, and limited healthspan? A wide variety of mechanisms are at play, but interpretation of genetic data from fruit flies bred for longevity suggests the major culprit may not be accumulated damage, or any particular mechanism, but rather the intersection of biological complexity with evolutionary adaptation. Long-lived flies demonstrate a huge breadth and depth of genetic differences from ordinary flies, rather than a few key alterations. Adaptations specialized for different stages of life interfere with each other antagonistically, causing many of the phenomena we perceive as aging, and leaving us in late life with a high but constant death rate. The networks underlying longevity are the networks underlying the overall operation of the organism. Today’s “narrow AI” tools, applied to the data from long-lived flies and other experimentally evolved organisms, will likely allow us to discover powerful new drugs for combating disease and increasing healthspan. But to thoroughly solve the “limited healthspan” problem will probably require generally intelligent Artificial Biologists, capable of deeply comprehending the structure and dynamics of biological networks in a way the human mind cannot.

Death and disease are such basic aspects of current and historical human life, that to envision a world without them requires considerable effort.

And yet, as technology advances, it becomes increasingly clear that they’re solvable problems.

As I write these words, I’m still reeling a bit from the tragic death last month of my 14 year old nephew Lev Goertzel Mann. Lev did not die from anything congenital or predictable — rather, from a sudden brain infection whose nature is still unclear, conjectured to perhaps be a rapidly incubating strain of meningitis, one that kills within 2-10 days. One evening he was happily enjoying a family vacation. The next morning, after a few hours of struggle, he was a corpse rather than a person. The death of a child is especially disturbing – but from a transhumanist view, it’s just a special case of a broader problem.

Touching death in one’s own life reminds one how fragile human life as a whole is — and how important it is to solve the various problems at the core of the “human condition,” the problems which traditional wisdom tells us are necessary and unavoidable, but science and transhumanism suggest may be solvable with intelligence and effort.

As Giovanni Santostasi wrote on the Singularity email list recently: “Death is not a metaphysical problem anymore. It is a scientific problem, and it should be solved as soon as possible.” Such views were previously found only in science fiction, but now they’re becoming increasingly mainstream. Futurists taking radical life extension seriously congregate at gatherings like the Immortality Institute conferences, the annual SENS conferences, and the upcoming H+ Summit with its Spockish “Live Long and Prosper” theme (put on by Humanity+, the organization that brings you this fine magazine: it’s in Los Angeles, this November, please come!).

But How Can We Get Rid of Death and Disease?

One approach is to liberate the mind from the body — by uploading human minds into robot bodies or simulated bodies in virtual worlds. An upcoming California workshop focuses on “Substrate-independent minds,” which is a felicitous way of putting it.

But it’s not clear how rapidly such “uploading” technologies will develop. And some people are emotionally or philosophically attached to their biological substrates, but would prefer them death and disease free. This leads to the question: what are the prospects for ridding the human body of the frustrating, sometimes terrifying shortcomings that cause it to die when it gets old – and sometimes, when it’s still young.

Some researchers believe we can massively reduce death and disease via “patching up” the various problems that arise in the body, without fully understanding the mechanisms underlying these problems. Others believe that the key is going to be a full understanding of the biological organism and — once we know how the body works — we’ll be able to systematically figure out how to improve its health and extend its healthspan. Both approaches are being avidly pursued by serious scientists, and, in my view, eventual success is almost certain. But the big question is when. There are many obstacles between here and there, including funding for research and limitations of current experimental technology. However, I’m increasingly convinced the most severe limitation constraining the quest for improved health and extended healthspan is the human mind itself.

Can the Human Mind Comprehend the Human Body?

One of the main lessons we’ve learned in the last few decades, as our exponentially improving instrumentation has enabled us to gather more and more data about humans and other organisms, is that human bodies are damn complex. This is why, over a decade past the completion of the Human Genome Project, our knowledge of the human genome hasn’t yet led to a parade of new blockbuster drugs. The complex self-organizing networks via which genes help produce organisms are something we’re just now beginning to understand. And the human mind, powerful as it is, has a limited capability to comprehend complexity.

This gives rise to the question: Can the human mind, aided only by tools that have dramatically less general intelligence than itself, comprehend the human body? Or will it need help — perhaps from advanced AI programs, specialized for the task?

You’ve probably guessed my answer. Way back in 2002 I created a spinoff of my AI company, Novamente LLC, called Biomind LLC, focused on the application of AI to bioinformatics. A central motivation for doing this was the feeling that, as the 21st century unfolded, the community of human biologists would need a powerful AI helping hand. And in the years since founding Biomind, I’ve spent a fair bit of time working on the application of AI technology to the analysis of genetics data. So I write here, not simply as someone interested in seeing himself and other humans live longer and healthier, but also as someone actively engaged in the quest to make this so.

Some of Biomind’s prior work on applying AI to age-associated diseases is reviewed in my 2009 essay called AI Against Aging. And the work has progressed since then. Among Biomind’s more interesting current project is a collaboration with the genomics firm Genescient Corp., aimed at using current AI technology to help analyze the data Genescient is gathering from the very-long-lived fruit flies they have developed.

The goal of Biomind’s work with Genescient is to learn lessons that may help combat human disease and increase human healthspan. The work is going quite well. But in addition to a host of genetic and pharmaceutical insights, another result of our work on this project has been to increase my conviction that to fully understand biological networks, the human mind will need serious assistance.

What Biomind is doing now is applying specialized “narrow AI” tools to Genescient’s data and using these in a judicious way based on the intuition and knowledge of myself and my collaborators. Although this approach is powerful, it’s also limited. It would be so much better if we had an “artificial biologist” capable of comprehending all online biological knowledge in one big gulp and analyzing the patterns natively, in the same way that we immediately analyze everyday visual scenes or the sound of the human voice. This is where my interest in life extension and biology comes together with my primary research thrust of Artificial General Intelligence.

When Will Human Biology Be Solved?

In 2005, I was invited to Australia to address a conference of geneticists, and over dinner I asked a group of them the following question: When do you think human biology will be solved?

Few of them had thought about the question before, but nearly all agreed it was a sensible one. If our current scientific paradigms hold reasonably true, then it seems likely that, in the same sense that we now understand basically everything we need to know about the physics of an automobile engine or a dishwasher, one day we will have pragmatically near-complete knowledge about human bodies.

Once the question had sunk in, answers I received ranged from “50 years” to “thousands of years.” As a Singularitarian, my own intuition gravitates toward the optimistic end of the spectrum. But as a practicing bioinformatician, I’m also well aware of the difficulties involved in cracking the life code. I think 30-50 years is probably the right answer because I believe we’re likely to create AI biologists that will romp and play where the human mind struggles.

In Search of Simple Solutions

Some may ask this question: even if understanding the human body and solving every problem that kills people is extremely complex and will require powerful AIs, couldn’t massively increasing human healthspan be a more specialized problem that can be done without comprehending all the complexity? If a weaker understanding could be used to solve 90% of medical problems — say, we could cure multiple forms of cancer, cardiovascular and neurodegenerative disease and increase the human healthspan dramatically — that would be reasonably nice as an interim measure. But the more I’ve learned about the healthspan problem, the more I’ve become convinced that it’s intimately tangled up with everything else about life and health — solving more than a moderate fraction of it in a separable way may not be tractable. I don’t want to overstate the case. This doesn’t mean we can’t make real incremental progress, perhaps resulting in powerful drugs that combat disease and extend healthspan. Clearly we can make some progress on our own with the experimental and analytic tools at hand and their modest improvements. But I suspect there are strict limits. (Although, as a human being, I really hope I’m wrong about this. I’d absolutely love to be given a simple pill, acting on a single gene or a handful of genes, that would extend my healthy life dramatically and radically decrease my odds of getting nasty diseases. This just doesn’t seem to be the direction the science is pointing toward, right now.)

Some researchers have sought simple, elegant theories that would tell us why we die as we get old. For example, there’s the Hayflick limit, discovered way back in 1961 — the observation that there’s a fairly strict limit on the number of times a normal cell population will divide before it stops. This is presumably because a cell population’s telomeres progressively shorten, and eventually reach a critical length below which cell division isn’t possible anymore. This is an elegant idea. But currently, as geneticist Joao Pedro de Magalhaes puts it, “there is little evidence to suggest that cells running out of divisions… are a major factor in aging.”

In a somewhat similar conceptual vein, Robert Bradbury (former CEO of the longevity firm Aeiveos) has put forth a theory of aging that appeals to me very much as a mathematician and computer scientist. He notes the evidence that there is increasing DNA damage with age, and notes that this damage is normally fixed by DNA repair mechanisms that are encoded in the genome. But then he asks (paraphrasing): What happens when the DNA that encodes the DNA repair mechanisms stops working? When that happens, he suggests, we’ll have a rapid downhill spiral. This theory has an elegance and beauty not common in biology. Similar to the quest for a single unifying equation in physics, it posits a single unifying dynamic underlying an apparently richly complex phenomenon.

I suspect the Hayflick limit and Bradbury’s recursive dynamic do play roles in limiting human healthspan, along with many other important specialized mechanisms. But a fuller study of the matter suggests that far more complex processes are going on, transcending not only those clever suggestions but any specific mechanism or process.

Repairing the Damage

Aubrey de Grey has put great energy into creatively promoting the notion of “strategies for engineering negligible senescence”, symbolized by the acronym SENS. He couples this with the idea that the aging process is largely a matter of accumulated damage of various sorts, on various levels. He then advocates that — rather than trying to cure the complex genetic and organismic problems underlying the damage — we should simply try to fix the damage.

For example, to minimize accumulated damage to mitochondrial DNA, Aubrey suggests that we move mitochondrial DNA into the nucleus. To prevent the build-up of “junk” composed of fragments of broken protein molecules, Aubrey suggests finding naturally occurring decomposing bacteria (e.g. in soil) and testing which ones may be able to decompose the junk without harming the human body. He lists seven key categories of damage, and proposes a set of possible remedies for each of them.

In a sense, Aubrey’s SENS approach accounts for the complexity of biological systems. Rather than identifying a single phenomenon as the crux of aging, he recognizes it as a phenomenon spanning multiple levels and aspects of biological systems. But in another sense, as some biologists suggest, he may be underestimating the complexity involved. If you move the mitochondrial DNA into the nucleus, what happens to the complex network of signaling relationships involving the DNA when it’s in its usual position outside the nucleus? Can you really find soil decomposer bacteria that won’t cause too many toxic side-effects? I think it’s definitely worth trying. But sometimes I wonder whether the SENS project adequate accounts for the subtly interlocking nature of the underlying biological systems.

Complexly Interlocking Adaptations

Genescient’s founder and Chief Scientist (and professor at the University of California at Irvine), the evolutionary geneticist Michael Rose, advocates a perspective that takes biological complexity even more seriously.  Like de Grey, he views health at a systems level, rather than looking for a single core mechanism underlying aging.  But he doubts whether accumulated damage is the critical factor we should be examining.

Rose likes to highlight results showing that, after a certain age has passed, an organism’s death rate (its odds of dying during a given year) stop increasing.  During this “late life” period, in a fundamental sense, the organism’s health may not be wonderful, but it’s not getting any worse.  Of course, some body systems like human teeth might keep degenerating even during late life, but the point is that the apparent existence of a “late life” period with a constant (or near constant) death rate argues against the notion that accumulating damage gradually kills an old organism off.

In Rose’s view, as organisms evolve different genetic mutations arise to adapt the organism’s functionality at different stages of its life.  But genes (considered as biological actors, not just sequences of amino acids) are complicated things. Each gene may carry out multiple functions, and in each of these functions, it’s subtly interlinked with other genes.  So various adaptations, focused on different stages of life, may interfere with each other in complex ways — an instance of a phenomenon known in genetics lingo as “antagonistic pleiotropy.” And antagonistic pleiotropy causes the organism all sorts of problems.  The more different life-stages worth of adaptations get piled on top of each other, the more confusion occurs in the body’s various interlocking systems, causing the problems we all experience as we age.  When an organism gets old enough, it reaches an age for which there hasn’t yet been much evolutionary adaptation in its history.  Few organisms in its species have lived that long, so the genes of the species haven’t adapted much to the requirements of life at that age. ‘

For instance, not many people have lived to age 100, so the human genome has not adapted much to the particular requirements of life at age 100.  Because of this, the human body at age 100 doesn’t have a lot of new problems due to conflicting adaptations, beyond the problems already present in the body at age 90.  On the other hand, many people have lived to age 50, and 40, and 30, etc.  So a 50 year old human’s body is full of adaptations specialized to improve life at age 50, along with some specialized to improve life at age 40, and some specialized to improve life at age 30, etc.  These different adaptations, specialized to improve life at different ages, often conflict with each other (e.g. because the same gene often serves multiple functions and plays roles in multiple networks), creating problems for the 50 year old, including many of the phenomena we describe as “aging.” In other words, at a certain age (somewhere between 50 and 100 years for humans), the problems experienced earlier in life stop compounding, and the death rate levels off.

If this is correct, then accumulated damage of the sort SENS seeks to address isn’t exactly irrelevant, but it no longer assumes a central role.  Some accumulated damage may continue into late life, but it’s not a star in the drama of mortality. Rather, it is a bit player at best.

Rose’s perspective comes out of an integrative analysis of all the available information, but particularly from his own work applying “experimental evolution” to breed populations of fruit flies (Drosophila melanogaster) according to specific fitness criteria.   For instance — getting back to the Genescient work I alluded to above — in a long-term project beginning in his lab at UC Irvine in the 1970s, Rose supervised the selective breeding of several lines of fruit flies that live four times as long as ordinary flies.  These fly experiments were then spun off from the university to form the basis of Genescient Corp.  The longest-lived lines of these “Methuselah flies” now approach five times the normal lifespan.

And these flies aren’t just longer-lived. They’re all-around superflies.  They have more sex, solve problems better, recover from infections better, have better hearts, etc.  For instance the following graph shows results from an experiment involving running electrical current through a fly’s body to accelerate heart rate, often to the point of failure.  As the figure shows, after a 2-minute recovery time, the Methuselah (O) populations had significantly lower percentage heart failure than controls (B).  Another similar experiment showed similar — although slightly weaker — results.














So we have superflies.  Where are the comparable superhumans?

It seems quite plausible that similar experiments could be done with humans -— however, unlike the 30+ years for Rose’s fly experiments, comparable experiments on humans would take at least tens of thousands of years. But if these experiments were done on humans, the result would be humans who live hundreds of healthy years — not due to any one magic mutation, or any small set of mutations, but due to a huge complex of coordinated, crosslinked changes in various genes influencing various body systems.

Recently published results from another lab about flies living a mere 1.7 times as long as normal flies bolster these ideas. These researchers show that their long-lived flies, even when chronologically “old”, display the gene expression patterns characteristic of younger flies. In other words, they have adapted, evolution-wise, to stay young longer.

But Rose has his doubts about the viability of puzzling out the relevant complex of genetic mutations underlying radically increased healthspan via human intuition. Humans, he points out, are wired to interpret situations in terms of relatively simple narratives with not too many actors or plot twists. But biological networks don’t work that much like the social networks of a primordial human tribe. Our brains seem poorly suited to limning their particular flavors of nonlinear dynamics and self-organizing emergence.

Rose’s views on aging are somewhat different from the mainstream of biologists, but there are others pushing in the same direction. Joao Magalhaes takes a view closely related to Rose’s on the role of developmental genes and processes in the aging process, but with a different emphasis. In this view, there are some developmental processes that cause harm to the organism when they stop (because evolution only deemed it necessary to figure out how to make them work in certain life-stages and not later ones), and others that cause harm when they continue beyond a certain point (because evolution didn’t deem it necessary to figure out how to turn them off at a certain point).

Magalhaes’ perspective does not seem to contradict Rose’s perspective in any fundamental way. At bottom, it just points at some particularly important cases of antagonistic pleiotropy. But it places the emphasis differently. It’s a more specific hypothesis. Rose suggests the culprit that increases our death rate as we age (before late life is reached) is the combination of adaptations for different life-stages; Magalhaes suggests that a lot of the problem has to do specifically with adaptations involving developmental genes.

A Virtuous Cycle

In Rose’s lab at the University of California at Irvine, he has also created a host of other fly populations, displaying other characteristics, such as flies with extra-short rather than extra-long lifespan. For future research, Rose envisions a “virtuous cycle” wherein advanced analysis of data from various specially-evolved fly populations is used to drive the creation of fitness criteria used in the evolution of new fly populations. In this way, experimental evolution and AI data analysis can proceed hand in hand. In principle. this whole iterative process could be roboticized, with the experimental evolution of various fly populations supervised by machines without any human intervention. But at the present time, this isn’t viable because creating appropriate protocols for selecting flies that meet various fitness criteria requires considerable insight and expertise.

Of course, flies are not the only organisms that can be evolved in this way. There is also a great deal to be learned from more rapidly-evolving organisms like yeast, and from more slowly-evolving organisms like mice. Performance Genomics, a Canadian company, has a population of mice that are somewhat similar to the Methuselah flies. They have been experimentally evolved, over a period of thirty years, for increased lifespan. These mice now live nearly twice as long as ordinary mice, and their genomics have not yet been thoroughly studied.

Part of the beauty of experimental evolution, as a research methodology, is that it embraces biological complexity in an automatic and implicit way. Evolution automatically makes use of the complex networks binding the various parts and levels of biological systems. To the extent that AI technology can infer things about these networks, it can guide evolution experiments in intelligent directions. This seems a highly promising avenue for rapidly increasing understanding of health, healthspan and other biological issues.

Lessons from the Superflies

The results of our work on the genetics of the Methuselah flies have not yet been published — but I can say, for now, that they provide general conceptual support for the views of Rose and Magalhaes. And the story should get even more interesting this fall. So far, we have been working with fly gene expression data, which measures (roughly speaking) how much messenger RNA each gene is producing at a certain point in time. But we are now starting to analyze gene sequence data that tells us which genes are mutated where in the Methuselah flies. Putting this together with the expression data should yield even more insights.

So far, one interesting observation from our work is that many of the genes emerging as significant in the longevity of the Methuselah flies show rich connections with mutations (SNPs, Single Nucleotide Polymorphisms) that arise in human studies on disease genetics. The following table (based on work UCI professor Anthony Long did for Genescient) shows some of the diseases that come out most relevant to the Methuselah fly gene expression data we’ve analyzed so far:


Among the more practical things we’ve done is to correlate the results of our genetic analysis with online databases that connect drugs and nutritional supplements to genes and proteins. This doesn’t necessarily tell you what substances to take to increase your lifespan but it’s still interesting. It tells you which of the available substances are richly related to the genetic networks that make the Methuselah flies live a long time. The list of relevant pharmaceutical substances is being kept proprietary for now, but part of the list of nutritional substances follows:

It’s striking that every one of these substances, which came out at the top of our list based on cross-correlating the genes our AI judged most important for distinguishing Methuselah from ordinary flies is suspected of already having some value for increasing human healthspan. At the very least, this suggests our research is on the right track.

Our work so far, applying a limited palette of AI-based data-analysis technologies to the Methuselah flies, has provided some interesting results. On the other hand, it has also validated Rose’s skepticism regarding the human brain as a comprehender of biological networks. Current AI algorithms have found multiple complex patterns that neither the human eye and brain or traditional statistics can find in these complex datasets with billions of data points. We have done our best to deploy these algorithms in a sensible way, using them not only to scan the fly genetic data for patterns, but also to integrate this data with some of the many existing online biological databases. Still, there’s a limit to what even the smartest human minds can do in this regard. As we work through the Methuselah fly genetics, we are acutely aware of how much more could be accomplished by an intelligent system combining humanlike vision and creativity with a computer’s precision, scope and flexibility in executing mathematical algorithms.

A Holistic and Elegant Mess

One of the qualitative findings of our Methuselah fly data analysis is that there are very many genes, proteins and biological networks operating differently in the Methuselah flies versus normal flies. It’s not just the fact that one or two or fifty genes have been tweaked. A vast number of biological subsystems have been gradually adapted to achieve increased healthspan, all of it in an intimately coordinated way. AI algorithms have been used to isolate smaller sets of genes and gene-combinations that play particularly pivotal roles, and to find genes that serve as hubs in the genetic interaction network. One can also study comparative “hubness differential” — i.e. genes showing more “hubness” in the Methuselah genetic networks than the ordinary fly genetic networks may be identified as potential “control genes” for processes highly linked with increased healthspan. But even these putative control genes are fairly numerous and their interactions complex.

There are various analytic gymnastics one can perform in order to isolate a small number of critical genes — for example, looking at genes with a high difference in hubness between Methuselah and normal flies, using specially constructed genetic interaction networks that consider only direct and not indirect interactions. Such sophisticated trickery may help one to isolate medically and economically viable drugs for combating age-associated diseases, and this is very important work. But finding a few genes with particularly critical roles doesn’t obviate the importance of comprehending the whole complex network if we are really going to solve the problem of increasing healthspan in a thoroughgoing way.

To use a fairly loose metaphor, what “control genes” do exist are probably more like the key politicians and businessmen guiding a democratic republic, than like the dictators directing a totalitarian state. If you want to radically modify the nature of a totalitarian state, it may suffice to identify the dictators and then figure out how to influence them. If you want to radically modify the nature of a democratic republic, then identifying the key politicians and businessmen is worthwhile (and can certainly make one a lot of money!), but influencing them isn’t necessarily going to transform the society in any simple or obvious way. There are too many other powerful factors at work, too many other subtle dynamics. One can get some important changes done this way – but to transform the society thoroughly, one really needs to understand the society as a whole, since it doesn’t operate in a top-down way, with a few key actors simply controlling everything else. There are multiple nested, interlocking feedback loops, resulting in complex self-organizing emergent phenomena — a holistic and elegant, iteratively self-modifying and self-creating mess.

Should Biologists Give Up Trying to Cure Diseases, and Retrain as AI Programmers?

Heh. Of course not. It makes sense to push forward on multiple fronts.

In my work with Genescient, we’re trying hard to find clues for new drugs, and new uses for old drugs, by appropriately analyzing data from the Methuselah flies. Notwithstanding all the underlying complexity, it’s perfectly viable to discover substances that reduce disease and increase healthspan by acting on pertinent genetic, proteomic and organismic subnetworks.

Even if such a substance only gives me another 5 or 15 years of life, I’ll take it! Maybe within that extra decade, human biology will be cracked, enabling me to live forever. This is Aubrey de Grey’s notion of the “Methuselarity”. The Methuselarity is the point at which cascading advances in healthspan extension will give you a reasonable chance of avoiding death via internal problems altogether.

The Methuselah flies seem to have better immune systems than regular flies. If analyzing their data leads us to a substance that strongly improves human immune function, it could drastically reduce the number of people dying from bizarre sudden infections like the one that killed my nephew Lev last month.

There’s plenty of good to be done by combining human intelligence, modern biological experimental machinery, narrow-AI and statistical data analytics. Over the next couple of decades, this approach seems reasonably likely to yield a few Nobel Prizes and billion dollar companies.

Still, the problem of radically increasing human healthspan is bigger than the garden-variety Nobel prize or billion dollar company. As my understanding of the problem increases, I become more confident that if we want to solve the big problem of disease and limited healthspan once and for all, we’re going to need an Artificial Biologist – something that can ingest all the available biological data as readily as we take in a beautiful landscape, and limn biological networks as intuitively as we do social networks. There are a lot of ways this Artificial Biologist could be engineered. It could start as a general human-level AI that gets taught biology; or as a collection of narrow-AI data-analysis algorithms that get commonsense knowledge; or as an integration of a bioinformatics toolkit with a commonsense and creativity focused AI. I have my own ideas about the specifics, centered on application of the OpenCog artificial general intelligence framework, but will save those for another article.

Involuntary Death is a Largely Avoidable Tragedy

I’ll return to Santostasi’s quote again: “Death is not a metaphysical problem anymore. It is a scientific problem, and it should be solved as soon as possible.”


Repairing damage may be part of the answer. Understanding biology much better is almost surely part of the answer. Experimental evolution of the sort that Rose’s lab does is probably part of the answer. And I believe that advanced AI (and probably AI with human-level general intelligence, but superhuman number-crunching capability) will be a large part of the answer as well. We must advance toward the goal on multiple fronts.

Some feel death is a valuable and necessary part of life. My philosophy is otherwise. Certainly death can have its beauty, as can torture and AIDS and all sorts of other unpleasant things. But, just because murder can be beautiful, doesn’t mean I want to live in a Tarantino film. To me, every involuntary death is a tragedy. With ongoing effort and intelligence, these are tragedies we can mostly avoid.