Cortical type as a substrate for laminar whole-brain models: a program

A roadmap for building a brain model that knows what kind of cortex each region is, not just how bright it looks on an MRI.

The core problem is that laminar whole-brain models — models that simulate how different cortical layers talk to each other across the whole brain — are under-specified. You need to decide, for each brain region: which neural populations exist there, how they connect to other regions, and what their baseline excitability is. The usual shortcut is to use the T1w/T2w myelin ratio from MRI, a single number per region. That's doubly bad: it's derived from the same imaging you're trying to validate against, and a single number carries no information about laminar structure — which layer talks to which.

The proposed substrate is cyto7, a hand-drawn atlas that classifies every cortical region into one of seven types, from allocortex (the evolutionarily oldest, simplest cortex) to koniocortex (the most elaborately layered, like primary sensory areas). Crucially, it was drawn from histology — actual cell-counting under a microscope — with zero reference to MRI. That independence matters: when you find that cyto7 aligns with fMRI gradients, MEG timescales, and receptor density maps, that's a genuine discovery, not a circular comparison. And because it's categorical, it sets node topology — allocortex literally doesn't have a layer IV, so you can't just plug in a number, you need a different circuit architecture entirely.

The paper lays out four parallel work strands. C1 converts the atlas into directed, laminar connectivity: instead of a scalar "how strongly are regions A and B connected," you get a (source layer, target layer, weight) triple, derived from the Structural Model of cortical connections, which predicts laminar projection patterns from the relative cortical type of the two areas. C2 triangulates excitation/inhibition balance three independent ways — from receptor autoradiography, from the aperiodic slope of the EEG/MEG power spectrum (a validated E/I proxy), and from the model's own simulated spectrum. Agreement among all three would close a circuits-to-dynamics loop. C3 is the most ambitious: it adds slow hormonal and neuromodulatory state as a third timescale sitting above the millisecond circuit dynamics, letting the model represent mood rather than just a snapshot of activity. The worked example is allopregnanolone, a neurosteroid that directly modulates GABA-A conductance — a parameter that literally lives in the model — with an approved drug (brexanolone) acting on it. C4 is the deliverable that ties everything together: a 7-row parameter table, one row per cortical type, with uncertainty estimates in every cell, that the ARCHEON agent model can consume.

The paper is honest about what's hard. The Allen Brain Atlas gene expression data underlying the hormone-receptor maps comes from six donors, mostly male, mostly left hemisphere — a real limitation for sex-hormone receptor maps specifically. Transcript levels are not protein levels, and there are no PET tracers for steroid receptors in humans. The authors flag these as exploratory rather than burying them. The macaque validation for C1 is listed as a prerequisite, not an afterthought: if type difference doesn't predict laminar projection patterns in macaque tract-tracing data where you can directly observe origin and termination layers, the whole human extrapolation is in trouble.

This is a program paper — its job is to write down the plan so four parallel efforts stay aligned and the atlas paper can cite a concrete roadmap rather than vague future intentions. The source is substantive and detailed, not sparse, though several of the four strands are explicitly works-in-progress.