Some limits on interpreting causality in neuroscience experiments

All our fiddling in the brain is supernatural

“Analyzing” by Aneeque Ahmed from the Noun Project

Causality is the killer test for science. And nowhere is causality more elusive than in neuroscience. We want answers to seemingly simple questions. Does the activity of neuron “Alice” cause behaviour “Boris”? The answers are not forthcoming, for the simple reason that brains are phenomenally complex, constructed from neuron upon neuron upon neuron.

If a rat’s motor cortex contains one million neurons, then how many of those are needed to make its whisker twitch? One? Ten? A hundred? A thousand? All of them? The answer is “somewhere in between”. And “somewhere in between” is a vast territory to explore.

And does the whisker twitch because of one spike from each of those neurons? Or the count of spikes? Or the pattern of spikes? The answer is “we don’t know”. And “we don’t know” is an even vaster territory to explore.

So what we need is:
1. A way to turn lots of neurons on and off, to find out if they are causal for the behaviour we are interested in, and
2. A guide to interpreting what it means to the brain when we do turn neurons on and off

Remarkably we have the first. We have an ever-increasing arsenal of tools for selectively turning on and off specific neurons. We can use optogenetics to target specific types of neurons — like the pyramidal cells in motor cortex, or the dopamine-containing neurons of the midbrain — and turn them on (or off) with coloured light. We can even tag neurons that were active during a specific behaviour, like remembering a location in a box, and have just those express the genes for being controlled by light. That way, the experimenter can use the light to just turn on that set of behaviourally-active neurons, whenever they want.

And for the second, the guide to interpretation, we have a beginning. (An outside observer of science would perhaps comment here that surely a rational science would have worked out the interpretation of the technique before using it on absolutely everything that wasn’t already nailed down. That observer is not aware that neuroscience has only a passing acquaintance with rationality — it’s that dot disappearing over the horizon — and is more like the dot.com bubble of the 90s: its full of people doing what ever the hell they want, driven by an influx of obscene amounts of money, and to hell with the consequences). A recent paper in Neuron had a brave stab at starting this conversation, but tripped over its own shoelaces by forgetting that time existed until the very end of its explanation, and then getting into a muddle trying to patch it in as an afterthought. Whereas, in reality, time is everything.

But kudos to the paper’s authors for starting the conversation. Here I’m going to carry on that conversation, and take another (wild) stab at how to interpret causality experiments. With time thrown in for free. And I’m going to tell you that, to the brain, what we do to it is supernatural.


You’re doing an experiment to test causality. You turn on a bunch of neurons in Brain Area X using light, and see what the animal does in response. What you want to know is this: when I turn on those neurons, what does that look like to the brain — does it look natural?

You want to know this because, if you make the neurons do something they never do in the course of the whole of the animal’s life, then how can you make any sensible conclusions?

Step one then is: what are the neurons doing normally?

We can see this by thinking about all their activity as marking out a natural region, where it stays, and an unnatural region, where it cannot go. For two neurons, we get this:

The natural region of neural activity. Each point in the graph is a possible combination of the activity of two neurons. The red shaded region is the “natural” region in which the activity of these two neurons always stays. The rest is “unnatural”. The black line is one possible path the activity of the two neurons might take in the natural region over time.

This shows the activity of two neurons at the same time: they can both have low activity, or both have high activity, or any combination in between. At all moments in time, the possible ways that these two neurons can be active at the same time is limited. There is a natural region where that activity stays.

The natural region could have many shapes — like a cigar:

A different shape for the natural region. The example path this time shows a repeating loop: neuron 1 has high activity when neuron 2 has low activity, and vice-versa.

The natural region could have holes in it:

A donut-shaped natural region. The activity of the two neurons never falls into the centre of the donut, or the outside of it.

We have, of course, more than two neurons. Seeing beyond these two dimensions is kinda tricky, but the same idea applies: no matter how many neurons we are keeping track of, their simultaneous activity can only be in a natural region.

(There’s a word I could use a lot here, but I’m not going to. That word is manifold. Poor old “manifold”. It’s right up there with “functional” for its capacity to be mistreated by a scientific paper. Perhaps only “architecture” and “complexity” outrank it for terms used vaguely to sound clever without any actual content. It simply means: a smaller dimensional space within a larger dimensional space. A plane in a cube, that sort of thing. When talking about dynamics, the manifold of the dynamics is the restricted space within the space of all possible dynamics. The natural region is a manifold).

There can be more than one natural region. Take some neurons sitting in the dark, puttering away. This, their “spontaneous” activity, sits within one natural region. Now flick on the light. A bunch of these neurons become more active, and stay more active. There is now a new natural region in which the activity stays. So “natural” can be context specific. But, still, each of these individual natural regions together makes one large natural region. There is always a defined region in which the activity of all the neurons will naturally stay.


Logically we can also define two different types of unnatural region for our bunch of neurons. The first is the totally unnatural regions neurons can never reach, because of their fundamental physical limits. These are regions of negative activity: neurons can never have less then no output! And regions of extreme output: beyond a certain limit, neurons cannot produce any more spikes.

The second is the supernatural region. This is a region the neurons will never visit in the course of the animal’s life. But in which they can be forced, however temporarily, because it breaks none of the rules of physics or biology. They just don’t go there because the connections between the neurons, and the strength of those connections, prevent them from ever happening.

Take for example a simple negative feedback loop. Two neurons: the first excites the other; the second inhibits the first. So any increase in activity of the first will also increase its own inhibition from the second. In such a loop, the first neuron’s activity is naturally limited by this inhibitory feedback: there is an upper limit of activity beyond which it cannot go. Yet the first neuron in isolation could fire much faster than it ever does in this loop. So the natural and supernatural regions are well-defined here: the natural region in which the feedback loop keeps the activity of both neurons, and the supernatural region where they could both go if they weren’t in this pesky loop.

So putting together the natural, supernatural, and unnatural regions, we have something like this:

The three possible regions of neural activity: natural, unnatural (physically impossible), and supernatural.

Time. You can see time is in everything above. The natural region is not just the set of simultaneous activity of our bunch of neurons. It is also how they get from one state of natural activity to another.

Not all possible ways of being naturally active can be reached from any other. Neurons can’t just jump around at random in the natural region — its physically impossible. Take a bunch of neurons active together in a regular on-then-off pattern: their joint activity goes up and down smoothly, and that smooth up-then-down is the set of natural moves. So in the natural region there are also natural moves and unnatural moves:

Ah, time. Neurons can’t just jump around at random within the natural region: they have to go somewhere close by. The blue arrows show some of the allowed moves, the natural moves. The green arrow is an unnatural move: the activity of these two neurons can never reverse under normal conditions.

Now then, step two. If we now know what neurons are doing normally, how do we interpret the causal effects of switching on a bunch of them with light?

Here’s what happens: We make the brain go supernatural.

In optogenetics experiments, we turn on a bunch of neurons at the same time, and often hold them on for seconds at a time. Or we turn off a bunch of neurons at the same time, and hold them off for seconds at a time. This is very, very far from a natural region for any bunch of neurons we could name

Either we force a bunch of neurons entirely into the supernatural region, whether by turning them on together (yellow arrow):

or turning them off together (yellow arrow):

Or, in our best-case scenario, we break the time rule: we somehow stay inside the natural region, but make a supernatural move:

Blue arrows: natural moves — the physically possible changes of the activity of the neurons over time. Optogenetic stimulation (yellow arrow) creates supernatural moves, even if we’re freakishly lucky and get it to stick inside the natural region of activity.

So we have a fundamental limit to testing causality in the brain: we always push our neurons into the supernatural region, so we can never be sure that what we observe as a behavioural consequence is naturally causal.

(Exactly the same ideas and conclusions apply of course to the older technique of microstimulation: injecting current into a whole bunch of neurons at the same time is just as supernatural).

We do not yet have any technique that keeps us within a natural region of activity across a group of neurons. To get this, we would need to design an optogenetic system that can mimic actual firing patterns. Either it would need to playback an earlier recording of those neurons’ pattern of activity, in their natural region. Or it could be a closed loop feedback design, taking as input the output of the neurons, and stimulating them in return according to some theory of how they are connected to the rest of the brain.


So, optogenetics is totally useless? No. Compared to what we had before, it is still a miracle. It allows tests of crude causality, testing whether or not Brain Area X is involved during a very specific part of a task or behaviour. It allows us to find excellent proof that specific types of neurons in Brain Area X connect to neurons in Brain Area Y: you hit one type of neurons in X with light, and see what neurons in Y respond. And it is fast becoming the best way to identify what type of neurons you are recording from in the brain of an animal: you can’t see the neurons, but if they respond to the light, then they must be the type of neuron tagged by the optogenetic markers.

That optogenetics works at all is actually telling us something fundamental about the brain. It’s one giant, highly resistant, dynamical system.

You see, your typical optogenetic experiment that tests effects on behaviour only turns on the light on alternative trials of the experiment (or randomises which trials it appears on). As we’ve just seen, this optogenetic stimulation is guaranteed to drive neurons away from their natural region. Yet in all these hundreds of reported experiments, the stimulation does not seem to affect behaviour on the very next trial. Which means that after each massive shove into the supernatural, the brain returns rapidly to the natural region. The brain ignores that giant perturbation, and carries on regardless. Robust old thing, isn’t it?

Want more? Follow us at The Spike

Twitter: @markdhumphries