Banner
    Learning Versus Evolution In Microbes.
    By T. Ryan Gregory | June 20th 2008 06:55 AM | 18 comments | Print | E-mail | Track Comments
    About T. Ryan

    I am an evolutionary biologist specializing in genome size evolution at the University of Guelph in Guelph, Ontario, Canada. Be sure to visit

    ...

    View T. Ryan's Profile

    One of my pet peeves is the common description in the media of bacteria "learning" to "outsmart" antibiotics. As anyone with a basic comprehension of evolution knows, learning has nothing to do with it. Learning is what happens during the lifetime of an individual, and it occurs in direct response to some information that the individual encounters. When bacteria become resistant to antibiotics, it is not by learning. The individual bacteria do not sense the antibiotic and change to become resistant. Rather, individual bacteria in a population that happen to be resistant because of some genetic difference (or in whom a mutation conferring resistance arises by chance or through gene transfer from another population) will survive and reproduce more effectively than individuals lacking the genetic characteristic that confers resistance. Over many generations of this process, the gene providing resistance to the antibiotic will be found in the majority of bacteria -- not because it "spreads" and not because individual bacteria develop  resistance, but because the bacteria that are the most abundant in the population after many generations are obviously the descendants of the ancestors that left the most offspring, namely those who survived the antibiotics.

    Unfortunately, the desire to consider evolution "learning" when discussing bacteria is not limited to non-specialists, and some scientists seem only too eager to reinforce this misconception. As a prime example, consider this story posted on PhysOrg:Thinking ahead: Bacteria anticipate coming changes in their environment. In this report, Saeed Tavazoie of Princeton claims "What we have found is the first evidence that bacteria can use sensed cues from their environment to infer future events".

    What they found is that bacteria (E. coli) that move from the outside world to the human gut undergo a switch to anaerobic metabolism before they actually enter the low oxygen surroundings of the digestive tract. They do this by instigating the change in response to a rise in temperature, which they encounter when they enter the mouth. In other words, a rise in temperature is predictably followed by a drop in oxygen in the world of these bacteria, and they respond to the former by making changes appropriate to the latter. Is this "predictive behaviour" that the bacteria "learned"? The authors seem to believe so. In fact, they argue, the bacteria can "learn" to do the opposite, if you establish an experimental environment in which a change in temperature is not followed by a reduction in oxygen availability. As the story says,

    Remarkably, within a few hundred generations the bugs partially adapted to this new regime, and no longer turned off the genes for aerobic respiration when the temperature rose. "This reprogramming clearly indicates that shutting down aerobic respiration following a temperature increase is not essential to E. coli's survival," says Tavazoie. "On the contrary, it appears that the bacterium has "learned" this response by associating specific temperatures with specific oxygen levels over the course of its evolution."

    Lacking a brain or even a primitive nervous system, how is a single-celled bacterium able to pull off this feat? Whereas higher animals can learn new behavior within a single lifetime, bacterial learning takes place over many generations and on an evolutionary time scale, Tavazoie explains.

    No. Individuals that did not exhibit the usual response of turning off aerobic metabolism when they experience a temperature increase -- who under normal conditions would not have done well, but in this experimental situation did better than alternatives -- left more offspring. Over many generations, this became common in the population because these individuals and their progeny consistently reproduced more successfully than alternatives within the population. This is simply microevolution and it has nothing to do with "learning" or "associating" or "inferring" or anything else of the kind.

    Another author interviewed in the report does not do much better with these issues when talking about their simulation of evolution.

    "To predict mealtimes accurately, the microbes would have to solve logic problems," says Tagkopoulos, a fifth-year graduate student in electrical engineering and the principal architect of the Evolution in Variable Environment framework.

    And sure enough, after a few thousand generations, an ecologically fit strain of microbe emerged which did exactly that. This happened for every pattern of cues that the researchers tried. The feeding response of these gastronomically savvy bugs peaked just when food was offered, says Tagkopoulos.

    The bacteria are not solving logic problems. The distribution of gene combinations is changing over time because individuals in the variable population that happen to have certain traits that confer advantages relative to alternatives in a particular environment leave more offspring in every generation. The sole problem of logic is equating evolution over generations with learning in a single lifespan.

    From the actual article:

    Within this in silico ecology, evolving organisms with random initial networks compete against each other in structured environments where signals and resources fluctuate with a distinct temporal correlation structure. In a typical experiment, the combinatorial states of multiple signals convey information about the availability of extractable energy resources in the near future. Cells that can efficiently "learn" such correlations are able to express the energy-extracting metabolic pathway at the appropriate time, giving them sizable fitness advantage over their competitors.

    Replace "Cells that can effectively 'learn' such correlations" with "Genes or combinations thereof that happen to produce cells that respond in a specific way to stimuli correlated with relevant forthcoming environmenal change" and this would be ok. Sure, it's more complicated, but given the major difference in accuracy, I think the page space would have been well invested.

    Bacteria may be the dominant life forms on the planet, but their success is a result of evolution, not learning.

     

     

    Comments

    This is simply microevolution and it has nothing to do with "learning" or "associating" or "inferring" or anything else of the kind.

    To be fair to the authors, evolution by selection is the basis of genetic algorithms, and you'll find chapters on genetic algorithms described in "Machine learning" textbooks. If a system comes up with an answer to a problem it is "learning". It hardly matters if the methods used are the same as in human learning (not that we actually understand those anyway). If you have a problem with bacteria "learning", what do you think of the whole field of "machine learning"?

    T Ryan Gregory
    This wasn't in a machine learning chapter, it was a paper about bacterial adaptation in Science.
    Yes, but my point was that evolution by selection has precedence for being called "learning", and if computers can "learn" that way, so can bacteria. I don't really see the objection.

    T Ryan Gregory
    Out of curiosity, have you *ever* agreed with anything I have posted?
    T Ryan Gregory
    By the way, do you agree with this characterization?
    Cells that can efficiently "learn" such correlations are able to express the energy-extracting metabolic pathway at the appropriate time, giving them sizable fitness advantage over their competitors.
    Even under the (highly questionable) analogy, it would be the *population* that "learned" -- you would have to take a top-down view -- and not the cells. As written, it is Lamarckian and misinterpreted.
    Excellent observation. Populations evolve, while "learning," as we ordinarily use it, is something individuals do. I would suggest that populations can also learn, for example by cultural transmission (which is Lamarckian), but this example is straight natural selection. Calling it learning muddies the water.

    In machine learning, a single computer stores many different algorithms, so one machine is a population. The "learning" (by the computer) represents a selection among the algorithms.

    Please continue to call attention to these cases where the urge to make science more accessible does damage to the underlying concepts.

    It is a cool paper, though.

    T Ryan Gregory
    In machine learning, a single computer stores many different algorithms, so one machine is a population
    Absolutely, which is what I meant by the population "learning" (an unnecessary and misleading analogy though it is), not cells.
    It is a cool paper, though.
    Certainly. But they oversell it substantially by equating it with learning when it is really standard evolution.
    I've agreed with plenty of things you've posted; your annoyance with "Dog's Ass Plots" for example, which you rightly argue imply a progressive nature to evolution that isn't there. In these cases you aren't just arguing semantics; there is a genuine misconception about evolution that people like John Mattick seem to have, and it is valuable to point this out.

    But what mystifies me is when you take the absolutely least charitable interpretation of an author's words (Jonathan Eisen does this from time to time as well, but not nearly as often). In this case, it is obvious that the authors *know* that humans learn in a single generation and this is different than in the bacterial case -- they even say so in the part you quote. I'm not seeing any misconception about this on the authors' part.

    By the way, do you agree with this characterization?
    "Cells that can efficiently "learn" such correlations are able to express the energy-extracting metabolic pathway at the appropriate time, giving them sizable fitness advantage over their competitors."


    If this had been said without context, I'd agree that it sounded Lamarckian. But given the context I don't think that's what the authors meant.

    Perhaps I am being more generous in this particular case because I think it is valuable to stop seeing bacteria as single celled organisms living in isolation of each other. Thinking of a biofilm composed of many bacterial cells from different species as a single organism is very valuable in thinking about why individual cells can have incomplete pathways. Thinking about bacteria distributed in time as well as in space as a single organism seems to be an interesting insight as well.

    T Ryan Gregory
    I'm not seeing any misconception about this on the authors' part.
    I think at the very least the authors are being quite sloppy, and in fact I am not sure they actually do grasp how the process works -- not many biologists I know would phrase the report the way they have. Doesn't matter, really, as scientists also have a responsibility to get it right when talking about their work, and in this case they undoubtedly contributed to confusion.
    adaptivecomplexity
    ...not many biologists I know would phrase the report the way they have.

    I think that's true. When you look at the big question - how a single-celled microbe, without a brain or CPU, can take advantage of informational cues in the environment, drawing an analogy with learning (in the cognitive sense) is tempting, but it is easy to be sloppy with that analogy.

    Mike

    Mike
    Great article. I am horribly unqualified to comment, but I have not learned/evolved to not ask questions. My question is, could the debate also touch on the the timeline, and maybe a parallel to the population/individual comment. Learning to me implies asimilating new information to reinforce or change behavior within a lifespan. My understanding is that evolution only can reinforce a specific characteristic that increases the success rate (survival) of that organism leading to future populations sharing that characteristic.

    I don't see why it's invalid to think of evolution as a process of learning from natural selection. It's a different perspective, sure, but it's just looking at the same process from a different angle.

    The paper's authors are also not the first ones to think of organisms as self-replicating computing machines that learn from their environments by modifying it's software, the genome. Like Jonathan Badger pointed out, there's an entire subfield of machine learning built on this idea. Here's a fun paper from David Mackay that looks at evolution from an information theory perspective http://www.inference.phy.cam.ac.uk/mackay/Evolution.html I guess my point is, bacteria do solve logical problems and learn from their environment. The mechanism by which they do so is evolution.

    You can make a case that the population "learned." We could have semantic arguments about mechanisms, but if learning means that the entity behaves differently afterward, it seems to apply to the population of cells.

    What is dreadfully misleading is applying that word to individual cells, as in:

    Cells that can efficiently "learn" such correlations....

    The cells didn't learn. They just died. Or not.

    jeisen
    Regarding Jonathan Badger's
    But what mystifies me is when you take the absolutely least charitable interpretation of an author's words (Jonathan Eisen does this from time to time as well, but not nearly as often)
    I cannot believe you are accusing me of being Satan here. I resent that and think that you really need to calm down. I am giving you the bad Jonathan award. Jonathan Eisen, UC Davis, Tree of Life Blog
    Jonathan Eisen, UC Davis, Tree of Life Blog NOTE - THIS IS A CROSS POSTING FROM MY "TREE OF LIFE" BLOG .
    Hank
    He's not accusing you of being Satan, he said Ryan was worse. So you are more like Malacoda.

    T Ryan Gregory
    Yeah. In any case, what I actually do is use a sloppy, misleading, or incorrect statement by someone who ought to know better as a launch point for explaining a concept. I provide quotes in their original context and links to the original source. Malacoda there does the same.
    jeisen
    I was kidding ... just trying to live up to Jonathan B's (who is a friend of mine and used to work for me) notion that I take overly pessimistic views of what people write. So I did it with what he wrote there ... Jonathan Eisen, UC Davis, Tree of Life Blog
    Jonathan Eisen, UC Davis, Tree of Life Blog NOTE - THIS IS A CROSS POSTING FROM MY "TREE OF LIFE" BLOG .
    T Ryan Gregory
    We know. *Wink*.