Week 3 summary + A note on hypotheses and optimisation

To summarise Week 3:

Tue – intro to phylogenies with Wayne Maddison; discussed some basics of speciation and evolutionary processes as well, in addition to stressing a non-hierarchical view of diversity and evolution (ie. no ‘ladder’ of progression)

Recommended reference for more phylogeny stuff: TR Gregory 2008 Evol Edu Outreach: ‘Understanding evolutionary trees’

For more info on the proper use of the term ‘basal’, see Krell & Cranston 2004 Sys Entomol: ‘Which side of the tree is more basal? — this is for the biologists among us especially! Many of us are guilty of abusing that term…although I’d think it’s ok as long as the other parties all know how phylogenies actually work, as a bit of a dirty illegal shortcut…

Thu – went over some further MURC info, brainstormed ideas for the short presentations, and then discussed that evol psych paper claiming depression is adaptive. Aside from the issues of the paper pertaining to psychology itself, the evolutionary reasoning was rather sketchy. The take-home message was that an adaptationist just-so story can be fairly easily created for just about anything, and just because we can make one up doesn’t mean it’s a useful explanation.

Hypotheses and fitness landscapes

Competing hypotheses and parsimony

Since this topic was brought up in class, I say useful because it’s rather difficult to experimentally reject a hypothesis about something in the past, and adaptationist stories are hard to support either way. We evaluate their likeliness based on understandings of modern organisms (ie. a good hypothesis should have some biological implications we’d be able to trace); but since we’re using ‘functional biology’ (biochem, mol biol, cell biol, physiol, genetics, etc) to explain many of those features anyway, why not just stay with the neutral explanation unless otherwise necessary?

For me, while adaptive explanations may be perfectly valid, they’re more like space fillers until sufficient actual biology is known to explain the phenomenon with real data to support it — if possible, of course.

And of course, proper science requires one to shave one’s hypotheses every once in a while with Occam’s Razor. Often, the only clear-cut algorithm we have to eliminating competing models is parsimony — that is, the simplest explanation (making the least assumptions) is the one we go by until a later experiment proves it insufficient. When things get a bit complicated, we can sometimes have conflicting models in operation in different contexts — ie, the model of physics we tend to use in biology is generally far removed from quantum mechanics and General Relativity — while we tend to ignore those areas, it doesn’t mean they suddenly stop acting in biology. It only means that so far, their effects and implications have not been significant enough for us to worry about.

Thus, competeing hypotheses can coexist, especially when we’re in murky territory. I’d say any currently interesting science is bound to be messy, so it’s important to try to get out of this dichotomous binary framework that seems to be more natural for us. Very often things aren’t either A or B, but rather: maybe A, maybe B, perhaps B is more likely than A; A may be more dominant/significant in situation X, B may become more pronounced in situation Y, etc. Add to this departmentloads of academic egos, and we get conferences and publications…

Analogy from phylogenetic trees

This actually ties into phylogeny, by the way — the number of possible trees grows SO fast for each added taxon, that after about 6 or 7 taxa, we have FAR too many possible trees to test each and every one of them. This is kind of like the set of possible hypotheses — we will never have enough lifetimes to test all possible explanations for pretty much anything. So what do we, and phylogeny software, do? We blindly wander around the space of possibilities, searching for the global optimum. What an computer algorithm does is tweak with the tree, each time testing whether it got better or not (in terms of parsimony). If it got worse, it goes back. If the tree got better, it proceeds further, until eventually it will not get any better — thereby reaching a maximum/optimum. However, that does not mean it’s reached THE (global) maximum, and a dumb (yet cheap) algorithm can easily get stuck on some crappy local watered-down alternative. And it doesn’t know that. In terms of phylogeny, that means we get a crappy tree that may be far from The Tree (‘Truth’) we search for.

So there’s entire fields in mathematics and computer science that study this problem of optimisation. It’s a fascinating topic, and thus far nobody has been able to prove the possibility of a universal optimisation algorithm that always finds the global maximum. Also interestingly, nobody’s been able to prove the impossibility thereof either.

See despite the ever-increasing complexity and ‘intelligence’ of algorithms, we never know whether the tree we have is the global optimum. You can run the algorithms from hundreds and hundreds of starting points, hoping that just by chance you’ve covered the entire terrain and found the highest optimum. But what if you missed the top peak? Keep that in mind when looking at phylogenies — they do not speak the Word of Evolution or anything! They’re just the likely best estimate the researchers got.

Hypothesis Landscapes

Ok, but that was just trees! Unlike dumb computer algorithms, we actually have insight and foresight and real intelligence and all those awesome things! Unlike a computer, we can actually see the landscape! Aren’t we so awesome?

Without going into the philosophical question of ‘what is foresight’, let’s assume, for the sake of argument, we have this magical looking ahead thing that computers don’t — however, it’s still limited! Foresight, not omnisight, and typically the stuff we actually pay attention too is far too complex for our foresight, and we must blindly stumble for an optimum, hoping we’ve found the global peak. Some approaches work better than others (increasing experience generally correlates with better optimisation algorithms, as we’ve tried more of what doesn’t work…), some are hard to compare.

*You can program a degree of foresight into computer algorithms as well, and they do it the same way we do — by sensing/data input, learning/data processing, modeling/memory, and correcting said model when new information conflicts with it. Scientific method, anyone?

So what science should really be presented as is an (more precisely, a set of) optimisation algorithm for finding explanations that best fit the environment, relying on parsimony, and ditching the models that fail to work (explain observations). The current scientific method is most likely the best tool available to us for exploring this landscape for any problems pertaining to understanding the world around us — but we still just blindly search for the best explanation we can get, aiming for the Global Optimum, and we can never know for sure whether we actually attained it.

Now, before I finally shut up, there’s one more entity that engages in optimisation landscapes…

Evolution, fitness landscapes and Design Space

Yes, you got it. Mainly because of selection, evolution appears to strive towards optimisation of fitness. I emphasise appears — evolutionary processes do not have any goals or desires, much like the particles of a chemical reaction lack this capacity. As any who’s suffered through physical chemistry would know, chemical reactions have a thermodynamic landscape to follow — they dislike climbing endothermic slopes, and enjoy sliding down instead, releasing heat (high energy = unstable in the world of chemistry). That was rather crudely put as other stuff kicks in, but we don’t care right now. But we can say chemistry searches for a thermodynamic optimum — the lowest energy state it can attain. This does not mean Chemistry is sitting there, thinking, ‘ah, let’s climb some peaks/crawl into some holes’. It’s just what happens. Same thing to the spawn of satan chemistry that is evolution. Apparently, I have to stress this point,  sorry if you got it ages ago!

So evolution has what is called a ‘fitness landscape’ — a set of possible states at different levels of fitness. Essentially, evolution ends up going upwards along the fitness landscape. For phylogenetic trees we have this 3D warped plane of possibilities, but not all options are physically possible for biological organisms! The fitness landscape is actually constrained by physics, chemistry, developmental biology, molecular genetics, etc. This revised fitness landscape can be said to inhabit the Design Space — a set of all possible designs an organism is ‘allowed’ to take. By far not all of those forms are viable, but they are possible. How much of this Design Space is ‘explored’ (blindly wandered on) by evolution, and how much we know of, is a matter of debate — I’d suspect it may not be very much, considering how each year brings shocking discoveries about how absolutely weird things can really get.

Now unlike our phylogenetic tree searching algorithm, evolution is arguably perfectly fine with sitting still at one fitness level, or  even going down a slight bit (Ohta 1992 Annu Rev Ecol Sys (VPN req’d) ‘The Nearly Neutral Theory of Molecular Evolution’ — highly recommended for those who are interested in neutral evolution as well and pop biol stuff. Math alert, but it’s actually not so bad…), thereby enabling it to explore more space (not that it cares the slightest, mind you). But unlike our hypothesis searching, and the more advanced tree searching algorithms, evolution lacks a capacity for making predictions based on past experiences. While one may still argue it could still have a sort of memory (eg. relic ancestral features that could come in handy if past environments are re-encountered), it definitely lacks any semblance of foresight, more so than computers and us!

Personally, I find it kinda cool to think of evolution as blindly stumbling around and branching out along the caverns and tunnels of design space, searching (but not really searching) for its optimal state. Kind of like some plasmodial slime mould, or fungus. This is quite reminiscent of what we do in our own lives, isn’t it?

PS: Robot sex evolution!  Floreano & Keller 2010 PLoS ONE — Evolution of Adaptive Behaviour in Robots by Means of Darwinian Selection