Unveiling universal aspects of the cellular anatomy of the brain

Mapping out the anatomic and connectomic structures of the brain is a major effort, with complete nanoscale reconstructions available for several small nervous systems and partial volumetric reconstructions for larger brains (human, mouse, fly, zebrafish). The largest reconstructions (~1 mm^3) already require petabyte-scale storage, raising challenges in big data analysis and, crucially, in deciding which structural aspects to focus on for meaningful cross-species comparison and for benchmarking computational models. At the cellular level, complexity has often been quantified using fractal dimensions of neurons (typical df ~1.1–1.9), reflecting trade-offs between connectivity and wiring cost. Self-similarity and scale invariance have been observed at multiple scales in brain structure and function (e.g., cortical gyrification, connectome scaling, dendritic branching). However, the relationship between structural and functional scale invariance remains incompletely understood. The authors propose that statistical physics offers a guiding framework: by treating cell fragments within sampled volumes as clusters, they test whether cellular brain anatomy exhibits signatures of collective phenomena near criticality and quantify universal scaling properties to enable principled structural comparisons across organisms.

The paper surveys work on whole-brain and large-volume nanoscale reconstructions across species (C. elegans, Ciona, Platynereis, Drosophila larva; partial volumes in human, mouse, adult fly, zebrafish; with full adult fly and more organisms forthcoming). Prior quantification of neuronal structural complexity has used fractal dimensions, with variability across organism, neuron type, and developmental stage, and methodological considerations. Scale invariance has been reported across macroscopic (cortical folding), mesoscopic (connectome self-similarity), and microscopic levels (dendritic branching; box-counting and correlation-based fractality). The authors also reference statistical physics models of clustered systems (percolation, Potts), critical phenomena, and scaling relations, as well as network-organizational principles relevant to brain structure (e.g., Rentian scaling, wiring cost optimization).

• Datasets: Public volumetric EM segmentations comprising ~1 mm^3 of human temporal cortex (4×4×33 nm native, sampled at mip 4: 128×128×132 nm), mouse visual cortex (4×4×40 nm, mip 4: 128×128×160 nm), and half of adult Drosophila central brain (8×8×8 nm, mip 4: 128×128×128 nm). Sampling at mip 4 balances data retention with computational feasibility. Sampling and analysis required >4 CPU-years.
• Sampling strategy: From segmentation (cell ID per voxel), generate random sub-samples: 3D cubes and 2D xy-slices. 3D: 1000 samples per L, with L=2^n voxels, n=4…9; largest physical sizes ~65.5×65.5×67.6 µm (human), ~65.5×65.5×81.9 µm (mouse), ~65.5³ µm (fly). 2D: 2000 samples per L; human/mouse up to L=4096 voxels (~524 µm), fly up to L=1024 voxels (~131 µm). Segments are partial cell fragments present within each sample.
• Benchmarks: Generate 3D critical site (p=0.311608) and bond percolation (p=0.24881) samples with free boundaries, matching sample counts and L ranges; analyze 2D slices for 2D measures.
• Exponent estimation and finite-size scaling (FSS): Treat segments as clusters. Measure:
  1. Largest-segment mean size µ(L) in 3D; expect µ~L^d at criticality. Use two-point slopes between successive L to obtain size-dependent d estimates; extrapolate d^f via linear fits versus 1/L using largest sizes.
  2. Box-counting fractal dimension d_box^(f): Overlay 3D grids of box size L_b; count boxes N_b(L_b) intersecting neuron reconstructions; slope of log N_b vs log(1/L_b) in linear regime gives d_box^(f). Applied to selected proofread neurons (fly: mip4; mouse/human: mip5) and largest percolation clusters.
  3. Mean segment size per voxel S(L)=V/Σ_i m_i (3D); at criticality S~L^{γ/ν}. Use FSS as for d.
  4. Segment size distribution P(m) in largest 3D samples; infer Fisher exponent τ via scaling relation τ=3−(γ/ν)/d^f.
  5. Pair correlation function C(r)=⟨S(x)S(x′)⟩−⟨S(x)⟩⟨S(x′)⟩ on 2D slices (x,y directions), averaged assuming approximate isotropy; at criticality C(r)~r^{−(d−2+η)}. FSS uses two distances r=L/8 and r=L/16 per L to estimate η(L); extrapolate via linear fits in 1/L.
  6. Higher-order correlations via gap-size statistics n(s), counting only site pairs at separation s within a segment with no intervening segment sites along the row; expect n(s)~s^{−ζ}. FSS uses s=L/2 and L/4 (brains) or L/8 and L/16 (percolation).
• Criticality proximity tests: Compute spanning fraction f_span(L): fraction of 3D samples with a single segment connected across all faces. Compare to percolation near criticality. Compute dimensionless cumulants of order parameter m (largest cluster size): U2=(⟨m^2⟩)/(⟨m⟩^2), Binder U4=1−⟨m^4⟩/(3⟨m^2⟩^2), and connected U_conn=1−⟨(m−⟨m⟩)^4⟩/(3⟨(m−⟨m⟩)^2⟩^2); examine size-independence/intersections across L.
• Neuron-only analyses: Where available, filter to neuron segments and repeat calculations. Fly: 21,739 traced, uncropped neurons. Mouse: 72,789 neurons with soma in volume; 78 extensively proofread neurons used for 3D reconstructions. Human: 15,827 neurons by soma table; 104 fully proofread neurons for 3D reconstructions.
• Scaling relations and hyperscaling tests: Verify Fisher’s identity (γ/ν=2−η) consistency by comparing η from C(r) with η derived from γ/ν. Test hyperscaling via η=2+d−2dν; define hyperscaling violation exponent δ=2+d−2dν−η (equivalently δ=2dν−γ/ν using Fisher’s identity).

• Fractal structure of largest segments: µ(L) scales ~L^d in 3D samples, consistent with criticality. Extrapolated d^f values (Table 1): Fly 1.61(5), Mouse 1.69(4), Human 1.50(5). Percolation benchmark much larger (≈2.52), highlighting distinct universality from percolation.
• Box-counting fractal dimensions of neurons (d_box^(f)) are smaller than d^f: Fly 1.42(1), Mouse 1.61(1), Human 1.39(1); percolation d_box≈2.52. Box-counting is less reliable than FSS at available sizes.
• Within fly neuron types (30 most frequent), d_box^(f) varies across types but centers near the dataset mean; fractal dimension positively correlates with neuron volume V (r=0.65).
• Collective size measure S(L)=V/Σ m_i scales as L^{γ/ν} with γ/ν consistent across organisms: Fly 1.3(1), Mouse 1.3(1), Human 1.2(2) (Table 1). Segment-size distributions show slow, heavy-tailed decay over orders of magnitude; Fisher relation yields τ≈2.2 for all three organisms, consistent with data.
• Long-range pair correlations: C(r) decays as a power law over >2 decades in 2D slices. Extrapolated η (Table 1): Fly 0.8(2), Mouse 0.6(1), Human 1.6(6) (human larger but with big uncertainty). Exponents broadly compatible between fly and mouse.
• Higher-order correlations via gap-size statistics n(s): ζ extrapolates near 2 (Fly 1.79(5), Mouse 1.9(1), Human 2.1(5)), consistent with other critical systems where ζ≈2.
• Scaling relations hold: η from C(r) is consistent with η computed via Fisher’s identity using γ/ν. This mutual consistency across independently measured exponents supports structural criticality.
• Hyperscaling violation: Estimated δ≈1 in all three organisms, indicating violation of standard hyperscaling in the cellular brain structure.
• Neuron-only analyses qualitatively agree with all-segment analyses at largest scales; small discrepancies (e.g., ζ in human at smaller L; mouse d^f differences) attributed to limited neuron identification/proofreading.
• Proximity to criticality: Spanning fraction f_span(L) increases with L and resembles percolation slightly in the ordered phase; cumulants U2 and Binder-like quantities are approximately L-independent in mouse and show finite-size trends consistent with near-criticality in other datasets.
• Cross-species compatibility: Across human, mouse, and fly, critical exponents are compatible within uncertainties, suggesting a shared structural universality class for cellular brain anatomy.

The findings indicate that cellular-level brain anatomy exhibits hallmarks of a clustered system near a critical point: fractal cluster growth of the largest segment, heavy-tailed segment-size distributions, and long-range pairwise and higher-order correlations. The agreement of independently measured exponents with standard scaling relations (e.g., Fisher’s identity) and a consistent set of exponents across species supports the hypothesis of structural criticality and universality in brain cellular structure. This addresses the challenge of selecting informative structural properties by identifying universal, scale-invariant metrics largely insensitive to microscopic details, enabling principled comparisons across organisms and with models. The observed hyperscaling violation (δ≈1) further constrains the universality class and suggests analogies to certain disordered critical systems. The authors discuss how these universal structural properties could underpin efficient yet non-minimal wiring, potentially balancing long-range connectivity demands with geometric cost, and they outline the potential to develop generative models that reproduce the measured universality to explore structure–function relationships and to benchmark physical network models against true brain anatomy.

The study introduces a statistical-physics framework showing that cellular brain anatomy in human, mouse, and fly exhibits structural criticality characterized by a consistent set of critical exponents obeying scaling relations, suggesting a shared structural universality class. These universal quantities provide robust, informative targets for cross-species comparisons and for validating or designing generative models of brain structure. Future work should: (i) analyze larger volumes and additional brain regions and organisms to assess regional and inter-individual variability; (ii) incorporate higher-quality, fully proofread datasets to refine exponent estimates; (iii) extend measurements to further exponents and universal amplitude ratios; (iv) connect structural universality with functional criticality (neuronal avalanches) and network-level properties (e.g., Rentian scaling, wiring economy); and (v) study developmental, pruning, and learning dynamics to understand how structural criticality emerges and is maintained.

• Data scope: Only a single partial cortical region was analyzed for each of three organisms; largest accessible length scales may just begin to probe long-range behavior in mammalian samples. Statistical methods could not be applied to smaller nervous systems.
• Segmentation/proofreading: Human and mouse datasets have incomplete proofreading and cell-type identification; split/merge errors and truncation at volume boundaries may impact largest-L results. Analyses suggest robustness to relabeling segments by within-sample contact, but residual biases may remain.
• Anisotropy and sampling geometry: Different voxel sizes in z (human/mouse) and smaller z extents limit 3D isotropic analyses and preclude computing C(r) along z; 2D slices used for correlation measures.
• Finite-size and resolution constraints: Computational limits motivated mip-4 sampling; lower mip (higher resolution) would be more precise but currently infeasible at largest L.
• Criticality control parameter: No tunable control parameter exists for the biological system; criticality is inferred via finite-size scaling, scaling relations, spanning fraction, and cumulant behavior rather than by tuning across a transition.
• Uncertainty estimates: Reported errors are statistical and do not include biological variability or all data-quality uncertainties; true uncertainties may be larger. Certain exponents (e.g., human η) have large error bars.