Artificial group selection: replies and a way around the problem
A follow-up to David Salamon, founder of the microbiome startup General Biotics, after he had pushed back on the back-of-the-envelope model in the previous email — raising the role of population size, mutation, and rapid environmental shifts in real bacterial populations. Alyssa works through each objection in turn, and then sketches a different scheme that might actually generate the kind of selection pressure needed: growing each subpopulation in a deliberately extreme and distinct environment, so that variants which are otherwise neutral become strongly favored or disfavored within the lab dish.
Taking the analysis one thing at a time: In the model I used, if I’m doing my big-O calculations right, the amount of time you need should be linearly proportional to bacterial subpopulation size. Hence, it takes forever if you have a subpopulation of 10 12 bacteria, but would still be very reasonable with subpopulations of 2 (or even thousands) of bacteria. The tradeoff there, of course, is that the smaller the subpopulations are, the less diverse they’ll be, and (with a lowish mutation rate) they’ll get less diverse with each generation, because rare alleles or species will go extinct. (I’m also not actually sure how you could maintain a bacterial population at a fixed, low level for extended periods — maybe limit the food supply? — but that one seems like something biologists would already have solved.)
Bacteria do mutate, of course, but if a mutation is fitness-neutral within the lab dish environment then (in expectation) it would take literally forever to spread through the subpopulation. Per Eliezer Yudkowsky’s Evolutions Are Stupid (But Work Anyway) , mean time to fixation is 2 · ln( N ) / s generations, where N is the population size and s is the fitness advantage of the mutation. If s = 0, then the equation blows up and you get a result of infinity. If s is very small (say 10 −5 ), you still get a very long time: 2 · ln(10 12 ) / 10 −5 = 5.526 × 10 6 generations ≈ 630 years at one hour per generation. If s isn’t very small (and isn’t fantastically unlikely to arise at all, by say requiring multiple simultaneous base pair flips), then it should already have arisen and reached fixation at some point in the past, since bacteria have been doing their thing for a pretty long while now.
Shifts in bacterial subpopulations can happen rapidly, but that seems like it would have to be induced by some environmental change (as in the examples cited). There’s an initial environment change when the bacteria are first moved from the chicken to the lab, but that seems to be mostly irrelevant in the model I outlined, because it doesn’t accumulate across generations and doesn’t vary between subpopulations. Eg., suppose that, under lab conditions, A does a bit better than B, and the ratio of A/B now becomes 80:20 (or B is eliminated altogether and the new variants are A and A′, or C turns out to be a beneficial mutation and the new variants are A and C, or what have you). That change might be beneficial to the chickens, or it might not; but it isn’t something you could iteratively select upon, because it only happens once and it happens the same everywhere. If C is a mutation that’s beneficial in the lab dish, but harmful to the chickens, you’re just going to be screwed because every subpopulation will have C. My model starts by assuming a population at equilibrium, but if the population doesn’t start at equilibrium it will reach one soon enough, and then the original model still applies from that point forward. In computer science lingo, it’s just a constant and doesn’t affect the big-O runtime, and (if I understand correctly) it’s a constant which is zero in expectation, since there’s no reason mutations that are helpful in the lab dish environment should be beneficial to the chickens rather than harmful.
However, it seems like you might be able to get around the last problem by placing the lab bacteria in lots of environments that are each different from the natural environment (and from each other), in the hope that chicken-positive mutations which used to be deleterious would now be highly advantageous. Eg., in the last model, suppose that type A bacteria do much better in acidic conditions, and type B do much better in alkaline conditions, but in the natural environment they’re equally fit (which explains why one hasn’t already out-competed the other). If you grow one population in a highly acidic environment and another in a highly alkaline environment, then type A vs. type B becomes a large fitness advantage within the subpopulation; if you assume a fitness advantage of 0.05, the fixation equation gives you 1,105 generations to fixation, which is only six and a half weeks. And once the bacteria are either all-A or all-B, it’s easy to tell through experiments which subpopulation is more beneficial. The difficulty there is that the mutations you gave an artificial fitness advantage to must also be plausibly beneficial to chickens; eg. if you induce selection by growing bacteria in a very hot environment at 80 °C, those bacteria will likely just die when placed into chickens at a temperature of ~30–40 °C.
That latter part would become a lot easier if we had data on how different microbiomes affected mammals — not just raw correlations between having certain microbes and doing well, but the causal connections between (say) microbes producing a certain protein and that affecting mammalian health. Has the literature gotten to that point yet?