PaperlayerView exact revision on arXiv

Source-linked AI summary

A predictive structural model of the primate connectome

Sarah F. Beul, Helen Barbas, Claus C. Hilgetag

arXiv:1511.07222v2q-bio.NC

TL;DR

The paper examines which anatomical factors help explain the organization of primate cortical connections. It compares architectonic similarity, spatial proximity, and cortical thickness, finding that architectonic similarity is most consistently related to connection features and that combining it with distance predicts projections with more than 90% accuracy.

  • Problem

    The principles organizing cortical connections remain unsettled, despite their strong constraints on functional interactions and the nonrandom organization of brain networks.

    Source: Introduction · p. 2

  • Method

    The study compares architectonic similarity, spatial proximity, and cortical thickness as predictors of primate cortical connections and related features.

    Source: Introduction · p. 3

  • Results

    Integrating cytoarchitectonic similarity and spatial proximity predicted whether two areas were connected with more than 90% accuracy, while cytoarchitecture was associated with connection presence, number, and laminar pattern.

    Source: Discussion · p. 9, Conclusions · p. 13

  • Takeaways & Limitations

    The findings suggest that cortical architecture provides a general organizing principle for connections in the primate brain.

    Source: General principles of cortical… · p. 13

    Cortical thickness is an imperfect surrogate and does not capture fundamental network aspects delineated by cytoarchitectonic differentiation.

    Source: Discussion · p. 12

Abstract

from arXiv · show

Anatomical connectivity imposes strong constraints on brain function, but there is no general agreement about principles that govern its organization. Based on extensive quantitative data we tested the power of three models to predict connections of the primate cerebral cortex: architectonic similarity (structural model), spatial proximity (distance model) and thickness similarity (thickness model). Architectonic similarity showed the strongest and most consistent influence on connection features. This parameter was strongly associated with the presence or absence of inter-areal connections and when integrated with spatial distance, the model allowed predicting the existence of projections with very high accuracy. Moreover, architectonic similarity was strongly related to the laminar pattern of projections origins, and the absolute number of cortical connections of an area. By contrast, cortical thickness similarity and distance were not systematically related to connection features. These findings suggest that cortical architecture provides a general organizing principle for connections in the primate brain.

Introduction

The paper asks which anatomical factors organize cortical connections, motivated by the nonrandom structure of brain networks and their functional constraints. It compares architectonic similarity, spatial proximity, and cortical thickness as predictors of primate connectome features.

  • Structural connections impose strong constraints on functional interactions among brain areas, while cortical networks show nonrandom organization and regular laminar patterns.Networks also contain modules and highly connected hubs.Source: Introduction · p. 2
  • The study tests whether inherent cortical properties explain prominent connectome characteristics by comparing three structural factors.The factors are cortical architecture, spatial proximity, and cortical thickness.Source: Introduction · pp. 2–3
  • The structural model uses relative architectonic similarity between linked areas to predict the laminar distribution of their interconnections.Earlier work applied this approach to projection patterns and other connection properties in macaque prefrontal and cat cortex [8,9,18,19,20].Source: Introduction · p. 2
  • The distance model proposes that remote areas have fewer and sparser connections than nearby areas, and may explain connection existence, strength, and laminar patterns [21–26].Source: Introduction · p. 3
  • Cortical thickness is assessed non-invasively by MRI and has been considered a possible anatomical covariate or surrogate of connectivity [27–37].The paper tests thickness similarity as a predictor alongside architecture and distance.Source: Introduction · p. 3

Results

Neuron-density similarity, especially combined with distance, best predicted whether cortical areas are connected. It was also associated with laminar projection patterns and connectivity topology, while thickness and distance showed less consistent relationships with connection features.

Results

  • Connected areas were more similar and closer than unconnected areas across the tested variables, with presence frequency declining as distance or structural dissimilarity increased.Relative-frequency rank correlations were ρ = -1.00 for density ratio, -0.98 for distance, and -0.93 for thickness ratio.Source: Relations among anatomical variables · pp. 4–5
  • Density similarity combined with Euclidean distance had the highest mean Youden-index among tested combinations; adding thickness did not improve performance, and higher thresholds exceeded 90% accuracy with fewer observations classified.At lower thresholds, accuracy quickly exceeded 80%; higher thresholds reduced false positives while maintaining a high true-positive rate.Source: Relations among anatomical variables · pp. 5–6
  • The trained classifier was also used to predict the status of projections marked unknown in the dataset, using a probability threshold of 0.85.Predicted present and absent projections are listed in Supplementary Table S1.Source: Relations among anatomical variables · p. 7

Laminar patterns of projection neurons

  • Supragranular projection fraction correlated with density similarity, and this relationship remained significant after controlling for thickness; thickness and distance were not independently associated with laminar patterns.The partial correlation for density similarity controlling for thickness was r = 0.43, p < .001.Source: Laminar patterns of projection neurons · p. 7

Relation of cytoarchitecture with connection topology

  • Neuron density was lower in structural-core areas and negatively correlated with areal degree, whereas thickness results were inconsistent across core status and degree.Density-degree correlation was r = -0.60; thickness-degree correlation was r = 0.38, but core and non-core areas did not differ significantly in thickness.Source: Relation of cytoarchitecture with… · p. 8
  • Network modules differed in neuron density but not thickness; among pairwise density comparisons, the frontal module exceeded the occipito-temporal module.The module-level density difference was H = 13.7, p < .01; the thickness difference was not significant.Source: Relation of cytoarchitecture with… · p. 8

Relations among anatomical variables

Discussion

Cytoarchitectonic similarity, especially neuron-density differentiation, was the strongest predictor of cortical connection features; combined with spatial proximity, it predicted projections with more than 90% accuracy. The discussion links cytoarchitecture to laminar origins, connection counts, and network organization, while describing thickness as an imperfect surrogate.

Discussion

  • More than 90% accuracy was achieved by integrating cytoarchitectonic similarity with spatial proximity to predict whether cortical areas were connected.The classification procedure also performed above chance, and the authors used it to generate predictions for unsampled projections.Source: Discussion · pp. 9–10
  • Thickness similarity correlated with the frequency of present projections but did not improve classification; the authors describe thickness as an imperfect surrogate for cytoarchitectonic differentiation.Source: Discussion · pp. 10, 12
  • Connection likelihood declined with greater cytoarchitectonic differences or distance, but similarity or proximity mitigated those effects, including for costly long-range projections.The authors identify cytoarchitectonic differentiation as a factor for predicting costly connections between remote areas.Source: Discussion · p. 10
  • Cytoarchitectonic similarity explained laminar projection-origin patterns across the macaque cortex, whereas distance and thickness showed no systematic relationship after accounting for it.Source: Discussion · pp. 10–11
  • Across the entire cortex, areas with lower density had more connections and higher area degree, while cortical thickness showed a weaker, inconsistent relationship to structural-core membership and area degree.Areas in the structural network core had lower neuronal density than peripheral areas.Source: Discussion · p. 11
  • Network modules differed in cytoarchitecture, suggesting it may be another factor in structural-module formation.Source: Discussion · pp. 11–12
  • Across macaque and cat, cytoarchitectonic differentiation was associated with network organization, including an inverse relationship between differentiation and area connection degree.The authors suggest this association may reflect general principles of mammalian cortical connectivity, while noting that its basis is unclear.Source: Discussion · p. 11, General principles of cortical… · p. 12
  • The authors propose that cytoarchitectonic differentiation may shape connections and constrain structural and functional aspects of the connectome.They also note that thickness is accessible non-invasively by MRI, but its relationship to structural connectivity remains unclear.Source: Discussion · p. 12, General principles of cortical… · p. 13

Methods

The methods combine macaque tracing data with measures of cortical architecture, distance, thickness, and laminar projection origins, using normalized frequency analyses and cross-validated classification to evaluate projection patterns. They also describe neuron density as an objective measure of cytoarchitectonic differentiation and examine its role in the association between projection-origin distribution and relative thickness.

Connectivity data: Presence of projections

  • Tracing data covered 2,610 sampled projections from 29 injected areas to 91 atlas areas, including an edge-complete 29×29 subgraph of injected areas.Projection strength was measured as the fraction of labeled neurons; the edge-complete subgraph enabled analyses without uncertainty from unsampled connections among those areas.Source: Connectivity data: Presence of… · p. 14

Connectivity data: Laminar origin of projection neurons

  • Laminar-origin analyses used NSG% for 625 projections from 11 injected areas and focused on 429 projections with more than 20 neurons.The undirected inequality measure |NSG%| mapped values near 0% or 100% to larger values; results were qualitatively unchanged using a 10-neuron threshold.Source: Connectivity data: Laminar origin of… · p. 15

Structural model: Neuron density

  • Neuron density quantified cortical cytoarchitecture; for each pair of connected areas, the log-ratio of their densities measured relative cytoarchitectonic differentiation, and profiles were determined for 1,128 sampled projections.Source: Structural model: Neuron density · pp. 16–17

Distance model: Spatial proximity

  • Spatial proximity was defined as Euclidean distance between mass centers for all 91 cortical areas, and relative thickness as pairwise log-ratios, included for all 2,610 sampled projections.Across 33 areas, histological and MR measurements correlated at r = 0.62 (left, p < 0.001), r = 0.48 (right, p < 0.01), and r = 0.56 for mean thickness values (p < 0.001).Source: Distance model: Spatial proximity · p. 17, Thickness model: cortical thickness · p. 18

Relative projection frequencies

  • Relative projection frequencies were computed by binning anatomical variables and dividing present projections by all studied projections in each bin to account for sampling disparities.Results were checked for robustness to bin-size changes.Source: Relative projection frequencies · p. 18

Classification of projection existence

  • A binary SVM used distance, absolute neuron-density log-ratio, and absolute thickness log-ratio in different combinations to classify projection presence or absence.The model used a linear kernel, standardized features, and uniform class priors, with five-fold cross-validation.Source: Relative projection frequencies · p. 19
  • Classifier performance was averaged across 100 repeated five-fold cross-validation runs, with mean Youden-index J computed over thresholds from p(present) = 0.85 to 1.00.A permutation analysis randomized labels and averaged performance across 1,000 five-fold cross-validation rounds.Source: Relative projection frequencies · p. 20

Statistical analyses

  • Group comparisons used two-tailed independent-samples t-tests or Kruskal–Wallis tests, with Pearson or Spearman correlations and a two-tailed significance level of α = 0.05.Welch’s t-test did not change results qualitatively.Source: Statistical tests · p. 21
  • Neuron density can be measured objectively and is an established characteristic measure of cytoarchitectonic differentiation in sensory and high-order association cortices.It closely corresponds to expert ratings that take multiple dimensions into account.Source: Methodological considerations · pp. 21–22
  • MRI thickness measurements covered all cortical areas, agreed with histological measurements, and were considered reliable despite the small sample.Reliability was further strengthened by averaging corresponding left- and right-hemisphere regions.Source: Methodological considerations · p. 22
  • The correlation between projection-origin distribution and relative cortical thickness was mainly driven by differences in neuron density.Source: Methodological considerations · p. 22

Figures

The figures map macaque cortical neuron density and visualize projection existence, classification, laminar origins, and connectivity topology. They also compare projection-classification performance across anatomical predictors.

Figures

  • Figure 1 depicts neuron densities in the macaque cortex on the M132 parcellation; cortical areas are positioned by density, and present projections are color-coded by absolute neuron-density ratios of connected areas.Source: Figure 1 · p. 31, Figures · p. 36
  • Figure 2 distributes absent and present projections across neuron-density ratio and Euclidean distance, with bars for counts and diamonds for relative frequency.Present projections become relatively less frequent as distance or structural dissimilarity increases.Source: Figure 2 · p. 32, Figures · p. 32
  • Figure 4 relates the supragranular-origin fraction to density and thickness ratios, while showing no association with Euclidean distance.The density association is strong, the thickness association moderate, and neither NSG% nor |NSG%| correlates with distance.Source: Figure 4 · p. 34, Figures · p. 34
  • Areas identified as belonging to a structural core network by Ercsey-Ravasz and colleagues (2013) had significantly lower neuron density than non-core areas; area degree decreased as neuron density increased.Source: Figures · p. 35
  • Figure 6 arranges cortical areas by neuron density, colors projections by absolute density ratio, sizes nodes by degree, and marks structural-core areas in red.Grey areas indicate unavailable density data.Source: Figure 6 · p. 36, Figures · p. 36
  • Across 100 cross-validation rounds, density ratio plus Euclidean distance performed best for projection-existence classification; adding thickness similarity did not improve performance.Figure 3 shows posterior probability and cross-validated classification; the supplementary figures compare performance across parameter combinations.Source: Figure 3 · p. 33, Figures · p. 33, Figure S1 · p. 37
Loading 1511.07222v2…