No sooner do I return from my own intelligence conference, about which more later, than I note, courtesy of another scholar, a fascinating new paper showing that 40% of the variance in IQ can be accounted for by a new measure of brain networks. This is strong stuff, so with a spinning head I tried to make sense of the new work.
Morphometric Similarity Networks Detect Microscale Cortical Organisation and Predict Inter-Individual Cognitive Variation, Jakob Seidlitz et al
firstname.lastname@example.org or email@example.com
doi: bioRxiv preprint first posted online May. 9, 2017;
Macroscopic cortical networks are important for cognitive function, but it remains challenging to construct anatomically plausible individual structural connectomes from human neuroimaging. We introduce a new technique for cortical network mapping, based on inter-regional similarity of multiple morphometric parameters measured using multimodal MRI. In three cohorts (two human, one macaque), we find that the resulting morphometric similarity networks (MSNs) have a complex topological organisation comprising modules and high-degree hubs. Human MSN modules recapitulate known cortical cytoarchitectonic divisions, and greater inter-regional morphometric similarity was associated with stronger inter-regional co-expression of genes enriched for neuronal terms. Comparing macaque MSNs to tract-tracing data confirmed that morphometric similarity was related to axonal connectivity. Finally, variation in the degree of human MSN nodes accounted for about 40% of between-subject variability in IQ. Morphometric similarity mapping provides a novel, robust and biologically plausible approach to understanding how human cortical networks underpin individual differences in psychological functions.
Unpicking this abstract takes some time. Morphometric similarity networks need to be mapped out, and then the resultant nodes calculated and correlated with the IQ measures.
How does one get the “feel” of a paper? I am reassured by the first paragraphs being cautionary in tone. Those of us whose understanding of scanning is restricted to being in the prone position, fighting claustrophobia, are prone to accepting the pictures we are shown at the end of the process as the cartographic truth. For many clinical purposes these are good enough, and two orders of magnitude better than anything available 20 years ago. We live in good times. However, when going deeper into the matter of which signal goes from which part of the brain to the other, forming any picture involves using approximations. Signals cannot be detected for long distances, and although even better scanners are in the pipeline, the bottleneck is interpretative power, not just detection. Even the prettiest pictures contain assumptions.
The mapping technique they used is innovative and fascinating. The full details are in their paper, but they integrated three approaches:
First, there is histological evidence from non-human primates that axo-synaptic connectivity is stronger between micro-structurally similar cortical regions than between cyto-architectonically distinct areas.
Second, there is encouraging evidence that conventional MRI sequences can serve as proxy markers of cortical microstructure. Cortical MRI metrics –such as magnetization transfer (MT), a marker of myelination -show spatial gradients in humans which align closely with known histological gradients in non-human primates
Third, there is emerging evidence that structural properties of the human cortex are more precisely estimated by the combined analysis of more than one MRI morphometric index at each region e.g. cortical thickness and sulcal depth, cortical thickness and myelination or cortical thickness and grey matter volume. On this basis, we predicted that morphometric similarity mapping with multiple MRI morphometric indices could provide a new way of estimating the linked patterns of inter-regional histological similarity and anatomical connectivity within an individual human brain.
Although the techniques and the results are exciting, further cautionary words are required. The mapping was done individually on the MRIs of 296 healthy young people, so this gives us the predictive measure to be tested. This involved measuring 10 shape variables in 308 cortical regions. They moved from the sample of discovery to the sample of testing, which was 124 other people. They also tested the accuracy of their mapping by looking at genes closely related to brain architecture, and found a mild but positive correlation between such genes and their mapping measure. They confirmed that the genes most involved in brain shape and signalling were far more likely to be involved than control genes taken at random, and that the random deletion of the most important genes had a disproportionate effect on the correlation. It seems likely that the mapping measure is somewhat related to genetic measures. The authors also applied their measures to the MRIs of 31 juvenile macaque monkeys and found:
Taken together, these findings indicate that the morphometric similarity of two cortical regions is directly related to the strength of monosynaptic axonal connectivity between them.
Then they were in a position to compare their mapping measure with IQ.
We predicted that IQ should be positively associated with integrative topological features that promote efficient information transfer across the whole network. High degree hub nodes are crucial to the global efficiency of the connectome and preferentially impacted by clinical brain disorders associated with cognitive impairment
They went back to their original sample, given here as 292 individuals, in order to test their hypothesis. That is, they went back to their original sample of discovery, which of course increases the possibility that the prediction fits only that sample, or that it fits that sample much better than it would fit any other sample. This is a limitation.
The IQ measures are good, though brief. MRI tests are expensive, and intelligence testing is far cheaper. The Wechsler Vocabulary and Matrix Reasoning test are a very good choice if you are in a hurry, but having done all this work it would have been much better to have a few more intellectual measures. For example, Coding takes 2 minutes, is a proper ratio scale measure, and is likely to depend on speed of connection between brain areas. An opportunity missed. Perhaps it was just part of the collaborative data set, but why not spend a little more time assessing intelligence, the best predictor of human outcomes?
We assessed the relationship between individual differences in IQ and individual differences in nodal degree of each of 308 regions in each of 292 individual MSNs using the multivariate method of partial least squares (PLS) regression, as in Whitaker and Vértes et al. (2016) and Vértes et al. (2016). This dimensionality reduction technique seeks to find the latent variables or PLS components which maximise the correlation between a set of collinear predictor variables and a set of response variables.
By this maximization technique they were able to account for 40% of the variance in IQ in a sample from which they had already derived the brain connectivity measures. As discussed, this may be inflating the strength of the observed correlation, particularly if the maximization technique is working powerfully (that is, fits itself as well as possible to this unique data set).
Breaking with tradition somewhat, the paper does not discuss limitations. I think that the paper is exciting and innovative, and may well be right. However, I have got used to modern genetics papers, which commonly find an excellent match between genetic and behavioural measures in sample sizes of 100,000+ only to find that the match crashes down when tested on a new sample of 25,000+. It may seem churlish to apply those standards to a neuro-psychology paper which depends on expensive MRI measures, but all is not lost. Many researchers have been looking at the link between brain scans and intelligence.
The authors could test their findings against the databases of Rich Haier and Rex Jung and their P-FIT hypothesis. They could also test their approach on the Human Connectome project, which has 1206 MRIs and a fuller battery of intellectual tests. As you might expect, I would like this finding to be proved right. Testing it might be able to be done very quickly, and a prompt confirmation would have a big impact on the well-established field of brain scanning and intelligence.
Disclaimer: My attention was drawn to this paper by Charles Murray. On that basis, you may wish to discount everything I have said, and consign me to the utter depths of perdition.