Thread by @jmacshine

Very excited to announce the publication of our new paper in @NatureNeuro: rdcu.be/bhO9H with @russpoldrack, @DrBreaky, @spornslab, @kaylena88, @p_t_brain, Sanmi Koyejo and Rick Shine!

The work was inspired by a skype conversation with @spornslab – we were both interested in finding ways to track the dynamic trajectory of the whole network of the human brain over time, but didn’t love the standard idea of chopping up time series into discrete windows.

Side-bar: you can check out psyarxiv.com/xtzre/ for a nice discussion of the state-of-the-art for time-varying connectivity if you're interested.

I recalled a brilliant paper that I had read by @neuro_theory and co. (ncbi.nlm.nih.gov/pubmed/26478179), in which the authors applied a spatial Principal Component Analysis (or PCA) to calcium imaging data in freely moving C. elegans worms.

The PCA allowed them to summarize the ~200 neurons into a set of 3-5 lower dimensions. Based on the mechanics of PCA, the components that it identifies are orthogonal to one-another, so you can treat them the same way that we use cartesian (i.e., X,Y,Z) co-ordinates.

If you analyze your data this way, you can track the ‘loading’ of each component over time: the dynamic pattern traces out a ‘trajectory’ of the whole system in relatively low-dimensional co-ordinates.

The fascinating result from their study was that different co-ordinates in the low-dimensional space coincided with specific worm behaviours (like turning, moving ahead, stopping, etc.)

To cut a long story short, they could describe what the worm was doing using a small handful of neural ‘components’. That is, the neural activity in the C. elegans evolved on a low-dimensional #manifold. ncbi.nlm.nih.gov/pubmed/26478179

We wondered: could we use the same approach on humans? The 1st challenge was to find data across the spectrum of human cognition. Most studies typically isolate a specific cognitive function (e.g., recognizing faces) and compare it with something similar (e.g., houses).

Fortunately, we were able to use data from the @HumanConnectomeProject, in which a large group of subjects performed a range of different cognitive tasks while undergoing 3T fMRI.

We analyzed the data in a similar way to C. elegans paper: we pooled imaging data together across subjects and ran a PCA. Lo and behold, our analysis revealed a dominant, low-dimensional neural signal: the first five PCs accounted for 67.9% of the variance.

In short, the brain was embedded on a low-dimensional #manifold! #another_dimension

We could also track the trajectory of each of the PCs. To do so, we weighted the original BOLD time series from a 100-subject Replication dataset with the PC loading for each component from a separate 100-subject Discovery dataset at each time point of the experiment.

The trajectory of the 1st PC was strongly correlated with the overall task block structure across all 7 tasks. The next 3 PCs time series were associated with specific tasks and not others, whereas the 5th PC tended to be most active at the beginning and end of task blocks.

Our next question: are the low-dimensional dynamics related to distinct underlying cognitive processes? We used @neurosynth to map the principal component embedding space onto meta-analysis data. Movement on the manifold directed the brain into distinct cognitive states.

So, in some sense, the brains of humans and worms follow a similar logic. The basic building-blocks of functional activity are low-dimensional and relate to the kinds of things that our brains let us do.

This is perhaps unsurprising – wouldn’t it be kind of weird if every time we performed a similar task, our brain used a completely different set of neural regions to complete the task!?!

This would be like leaving the highway to take a completely different route to work each day

Side-bar #2: If you're interested in manifolds, check out this beautiful theoretical work by Vik Jirsa and @ar0mcintosh: osf.io/preprints/pszg…. The basic idea is that low-dimensional constraints ‘enslave’ high-dimensional neural activity.

Next, we showed that PC1 was associated with an integrated network topology, with connections that linked across specialist modules. Evidence for a core brain system that fluctuates over time, bringing new brain systems on-line in accordance with changing demands.

This links the low-dimensional flow with some earlier work, in which we showed that cognitive task performance is associated with network integration: macshine.github.io/publications/2…. @russpoldrack @ChrisFiloG @Dr_Moodie @p_t_brain @JHBalsters

What controls the flow on the manifold? We used neurotransmitter patterns from the Allen Brain Atlas to show that PC1 +vely overlapped with the spatial expression of a range of metabotropic neurotransmitter receptors (D1, α2A, M1 and 5HT2A) that are known to promote cognition.

This suggests that activity from within the brainstem might help to drag the system across state space. That is, activity in e.g., the locus coeruleus would project more noradrenaline to the very regions of cortex that are required for cognitive function, thus recruiting PC1!

In other recent work (macshine.github.io/publications/2…), we have demonstrated a biologically plausible a mechanism for this function: the modulation of ‘neural gain’ -- i.e., the amount of metabotropic NTs in a neuron determines how likely it is to propagate an incoming action potential.

If you think about it, this isn’t the worst way to wire up a brain. Whenever you can’t handle a particular challenge, send a signal down to the brainstem to release more modulatory neurotransmitters in the cortex, integrating the brain and recruiting more neural resources.

There’s a bunch of other cool stuff in the paper, along with links to code and references to all the amazing work that inspired the study. Oh, and the data is all openly available! Shout out to @HumanConnectome! PDF here: rdcu.be/bhO9H. Thanks for reading!

Sorry @ViktorJirsa! Forgot that I followed you on Twitter...