12,273 Publications · 71,443 Citations

Question: does mesoscale tracer axon density predict how often thalamic and visual-cortex units show fast directed spike links? Data: Allen Visual Coding Neuropixels (46 mice, 1,049,307 thalamus-cortex pairs) and Visual Behavior Neuropixels (69 mice, 768,943 pairs); LGd and LP units against units in six visual areas (V1, LM, RL, AL, PM, AM). A link is a cross-correlogram peak at +1 to +4 ms (thalamus leads) or -4 to -1 ms (cortex leads) above a hollow-Gaussian baseline. Anatomy: Allen tracer experiments; because many LP and LGd injections label both nuclei, the primary estimate used clean injections only. Result 1 (pre-specified, passed): with nucleus and area fixed effects, spike-count and cell-class controls, the odds of a thalamus-leading link increased 2.59 times (95% CI 1.58-4.24) and 1.92 times (1.47-2.52) per 10-fold density; route permutation p = 0.045 and 0.0013. Unmixed and all-injection tracer estimates, a sharp baseline, strict thresholds and different-probe pairs gave the same sign. The relation is compressed: density spans about 900-fold across the 12 routes, link rates about 10-fold. In Visual Coding a control lag (+21 to +24 ms) also had a slope (1.66), and the slope depended on the LGd-to-V1 route. Exploratory: LP projects more densely to V1 than to higher areas, but LP-to-V1 pairs had fewer links (odds ratio 0.49, 0.34-0.71; Visual Behavior 0.77, 0.53-1.10); LP-to-V1 axons end mostly in layers 1 and 5. LGd links preferred V1 only 2.6-4 times, against a 150-fold density ratio. Result 2 (pre-specified, failed): cortex-leading links did not follow area-to-thalamus density (1.40 and 1.14 per 10-fold, CIs include 1), stayed near control-lag rates, and came more from fast-spiking units, which do not project to the thalamus. Limitations: CCG links are putative; 12 routes give few degrees of freedom; tracer data come from other mice and few clean injections. Allen data are not bundled; code regenerates all derived tables.

We tested cat and rabbit rules for thalamocortical connections in the awake mouse. We used spike cross-correlograms between LGd and V1 (VISp) units in two public Allen Institute Neuropixels data sets: Visual Coding (27 mice, 86,182 LGd-VISp pairs) and Visual Behavior (43 mice, 68,309 pairs). A causal 1-4 ms peak above a hollow-Gaussian baseline identified putative connections in 0.48% and 0.52% of pairs (LP control: 0.04% and 0.12%). Fast-spiking targets had 1.8% detection versus 0.13-0.24% for regular-spiking targets (odds ratio 7.5 and 4.1 after spike-count controls), and layer 4 had more detected input than layers 5 and 6. These effects stayed with strict thresholds, different-probe pairs only, and a sharp-peak detector. RF-centre distance had a graded effect (odds ratio per 10 degrees 0.85-0.88 with the main detector, 0.71-0.77 with the sharp detector), and connections still occurred at RF distances above 30 degrees. Same ON/OFF flash preference gave no consistent advantage (VCN null; VBN odds ratio about 2). Single connections were weak (median 0.5-0.7% excess spikes per LGN spike onto FS cells). Tonic spikes after 10-30 ms intervals had 14-26% lower relative gain than spikes after 30-100 ms intervals in both data sets. A burst-spike advantage appeared only with the broad baseline. Running did not change the relative gain. Limitations: CCG peaks are putative connections and can include LGN synchrony and a ~60 Hz rhythm; FS class, layer and RF estimates are approximate. Allen data are not bundled; code regenerates all derived tables.

Question: does dendrite shape predict orientation tuning in mouse visual cortex? Weiler et al. (2022) reconstructed about 30 L2/3 pyramidal cells. They reported that the apical tree is elongated along the preferred orientation after retinotopic mapping (cc = 0.5), and that sharply tuned cells have less complex apical trees (r = 0.58 with tuning width). We tested both claims in the MICrONS mouse, which has in vivo two-photon tuning and an electron-microscopy reconstruction for each cell (12,829 coregistered neurons, v661 skeletons). We calibrated the image-frame conventions from the digital twin responses. We mapped each apical axis to visual space with a local retinotopic fit and compared it with the preferred bar orientation against shuffle nulls. Result for C1: in 1,524 tuned excitatory V1 L2/3 cells, the alignment was not above the null (mean cos 2 delta 0.050; within-scan null 0.078, p = 0.083; local null 0.055, p = 0.73). All 15 variants were inside or below the null range, and the signed one-axis test went in the opposite direction. Result for C2: in 3,779 V1 L2/3 cells, apical Sholl peak did not predict in vivo gOSI (rho 0.007; r 95% CI -0.015 to 0.049; depth- and scan-adjusted slope -0.021, CI -0.055 to 0.036). Other areas, layers and complexity measures gave the same result. In random 30-cell subsets, |r| >= 0.58 occurred in 0.12% of draws. Limitations: one mouse; most skeletons are automated. A post hoc proofread subset (136 cells, mostly higher visual areas) gave a weak trend in the Weiler direction (rho -0.27; adjusted slope -0.19, p = 0.13) that is not resolved. Unproofread cells with complete-looking apical trees (n = 2,589) were null. The tuning stimulus was Monet noise, not gratings.