Functional connectivity recovers canine–human network homology along the evolutionary cortical hierarchy

Abstract

Mapping functional brain organization between species is a prerequisite for translating connectivity biomarkers from animal models to patients, but connectivity-based correspondence has been demonstrated almost exclusively in primates. We tested whether canine–human network homology can be recovered from functional connectivity alone, using only five network-level anatomical priors, and benchmarked a spectral manifold alignment framework (Laplacian-eigenmap embedding with Procrustes alignment on network centroids) against a simple connectivity-signature matching baseline. Data comprised resting-state fMRI from 15 canines with endovascular permanent middle cerebral artery occlusion (pre-occlusion baseline and hyperacute post-occlusion sessions) and two public OpenNeuro datasets: 29 chronic-phase post-stroke aphasia patients scanned before and after rehabilitation (ds003999) and 10 healthy controls (ds000224). Held-out canine–human network homologies were recovered well above chance (top-5 network correspondence up to 54.5% vs. permutation null $\approx$18%, $p=0.002$), and correspondence followed the evolutionary cortical hierarchy: visual and somatomotor homologies were recovered several-fold above chance, default-mode only weakly—extending a gradient previously described in primates to a carnivore. Notably, simple signature matching (54.5%) outperformed manifold alignment (35.5%), indicating that the homology signal is strong enough to be captured without embedding machinery at translational cohort sizes. Within species, canine network metrics showed no treatment-arm differences (all $p>0.3$), and chronic-phase patients did not differ from controls in modularity ($p=0.93$); a sensitivity analysis bounds these nulls. All results are fully reproducible from shared code and public data.

Highlights

• •Connectivity alone recovers canine–human network homology, 3-fold above chance.
• •Correspondence follows the cortical hierarchy: strong sensory, weak default-mode.
• •First quantitative connectivity-based homology mapping beyond primates.
• •Simple connectivity signatures capture the signal best (55% vs 36% for manifold).
• •All results, including nulls, reproduce from shared code and public data.

1. Introduction

The failure to translate promising neuroimaging biomarkers from animal models to human clinical applications represents one of the most significant obstacles in neuroscience research. This translational gap is particularly pronounced in stroke research, where the vast majority of neuroprotective therapies have failed in human trials despite successful preclinical studies. A fundamental contributor to this failure is the lack of quantitative frameworks for mapping brain organization across species.

Traditional approaches to cross-species comparison rely on anatomical correspondence, assuming homologous structures can be identified based on cytoarchitecture or evolutionary relationships. However, this approach faces significant limitations: (1) substantial differences in brain size and gyrification patterns make direct alignment problematic, (2) functional homology may not correspond to anatomical homology, and (3) pathological processes like stroke create lesions that further complicate registration.

Recent advances in manifold learning offer an alternative: mapping functional organization based on intrinsic connectivity patterns rather than dense anatomical registration. Brain networks can be represented as graphs embedded in high-dimensional spaces whose geometry reflects functional relationships; by learning mappings between these functional spaces, conserved organizational principles can be identified using only sparse, network-level anatomical priors (a handful of homologous canine–human networks) rather than voxelwise anatomical correspondence. The method does not eliminate anatomical knowledge; it substitutes minimal, well-established network homologies for the dense spatial correspondence that cross-species registration normally demands.

In parallel, a growing literature on data-driven functional connectivity modeling has shown that learned representations of connectivity carry clinically meaningful information. Deep learning approaches that fuse spatial and temporal features of adaptive dynamic functional connectivity (Li et al., 2020), aggregate features derived from dynamic effective connectivity using virtual adversarial training (Li et al., 2022), or integrate multiple parcellation templates through spatio-temporal weighted multi-hypergraph convolutional networks (Liu et al., 2024) have achieved strong within-species performance for brain disease identification. These methods presuppose a common reference frame across subjects, which does not exist across species. They also raise a question that any cross-species mapping method must answer before its machinery is credited: how much of the achievable performance is already available to simple alternatives, particularly on the modest cohorts typical of translational work?

Here we ask two questions on real data at translational-cohort scale. First, can canine–human network homology be recovered from functional connectivity alone? Connectivity-based cross-species correspondence has been demonstrated almost exclusively within primates (Mars et al., 2018, Xu et al., 2020); the dog – phylogenetically distant, gyrencephalic, and an established stroke model – is a substantially harder and more informative test of whether functional organization alone carries homology information. Second, how much methodological machinery does that recovery require? We evaluate a manifold alignment framework – graph construction, spectral embedding via Laplacian eigenmaps, and Procrustes alignment on sparse network anchors – against a permutation null and against a deliberately simple connectivity-signature baseline, using resting-state fMRI from a canine experimental-stroke model and public human stroke and control datasets, together with within-species treatment-response and group analyses. All results, including null findings, are reported in full, and every quantitative claim is reproducible from the shared code and quality-control outputs.

2. Methods

2.1. Participants

Canine cohort. Fifteen adult purpose-bred mongrel canines (25–30 kg) underwent experimental ischemic stroke via endovascular permanent middle cerebral artery occlusion (MCAO), as described in our source experimental studies (Warioba et al., 2024a, Warioba et al., 2024b): occlusion was produced under fluoroscopic guidance, anesthesia was maintained with isoflurane ($\sim$1%) with physiological monitoring (arterial pressure, EtCO₂, blood gases, temperature), and all procedures followed the approved institutional IACUC protocol #72498. Animals belonged to three arms: control/natural history (n=6), NEH flow augmentation (intra-arterial norepinephrine 0.1–$1.52\mathrm{\mu }\mathrm{g}$/kg/min titrated to 25–45 mm Hg above baseline mean arterial pressure, plus hydralazine 20 mg; n=5), and Sanguinate (pegylated bovine carboxyhemoglobin, 8 mL/kg IV over 10 min beginning 30 min post-occlusion; n=4). These 15 animals are the subset of the parent cohorts (n=40 across the source studies; 27F/13M, age 4.2 ± 3.3 years) for which complete preprocessed resting-state fMRI was retrievable for both imaging sessions; the parent studies report the full cohorts.

Canine imaging timepoints. Imaging followed a two-session design: a Day-1 pre-occlusion baseline scan and a Day-2 post-occlusion scan acquired in the hyperacute window (approximately 2 h after MCAO). All canine longitudinal contrasts in this study are therefore two-timepoint (baseline vs. hyperacute) contrasts; no additional serial timepoints were acquired.

Human cohort. Human data were drawn from two public OpenNeuro datasets (Table 1). Stroke patients came from ds003999 (Daminov et al., 2022): 29 chronic-phase post-stroke patients with aphasia (11F/18M) scanned before and after an inpatient rehabilitation course at the Pirogov National Medical and Surgical Center, at a median of 167 days post-stroke (IQR 103–254; range 31–345). All patients had aphasia (15 motor, 5 acoustic-mnestic, 9 mixed/other types); lesions were left-hemispheric in all cases with recorded laterality (12 L, 1 bilateral, 16 unrecorded); aphasia severity was assessed with the Vasserman scale before (28.2 ± 12.9) and after (22.0 ± 12.0) rehabilitation (lower is better). Healthy controls came from ds000224, the Midnight Scan Club (Gordon et al., 2017): 10 adults (5F/5M, age 29.1 ± 3.3 years), of whom we used one 30-minute resting-state session each (ses-func01). We note explicitly that the human cohort is chronic-phase and was not acquired to match the canine hyperacute paradigm; the implications of this mismatch are addressed in the Limitations.

Table 1.

Human cohort characteristics (values computed from the datasets’ participant records). Vasserman scores: lower is better. QC exclusions are per run (session), not per subject.

	ds003999 (stroke)	ds000224 (controls, MSC)
Subjects available/analyzed	29/29	10/10
Sex (F/M)	11/18	5/5
Age, years	not released	29.1 ± 3.3
Time post-stroke, days, median (IQR)	167 (103–254)	–
Aphasia type (motor/ac.-mnestic/mixed)	15/5/9	–
Lesion laterality (L/bilateral/n.r.)	12/1/16	–
Vasserman pre/post rehabilitation	28.2 ± 12.9/22.0 ± 12.0	–
Sessions per subject	2 (pre-, post-rehab)	1 (rest, 30 min)
rs-fMRI	TR 3.0 s, 2.0 × 2.0 × 3.9 mm, 163 vols	TR 2.2 s, 4 mm iso, 818 vols
Runs processed/passing QC	58/49	10/10
QC exclusions (motion/coverage)	4/5	0/0

2.2. Image acquisition and preprocessing

Canine MRI. 3T Philips system; resting-state BOLD-EPI with TR=1400 ms, TE=20 ms, 2.5 mm isotropic voxels, 300 volumes per session, under the anesthesia protocol above.

Canine preprocessing. Canine data were processed with an FSL-based, species-specific pipeline: rigid-body motion correction (MCFLIRT), non-brain removal, spatial smoothing, grand-mean intensity normalization, high-pass temporal filtering, and single-session MELODIC ICA denoising. Region-level connectivity was then derived through an audited registration chain designed for robustness on canine EPI: voxel geometry was set to the acquired 2.5 mm isotropic resolution; each session’s EPI target was brain-extracted by adaptive connected-component masking (45–110 cm³ volume prior), preventing extracerebral tissue from dominating the cross-modal cost function; and gross slice-axis orientation was determined automatically per run (Dice overlap of the EPI brain mask with the template brain mask over all axis rotations and reflections) before fine registration. Each run was registered to the Johnson et al. canine population template (FLIRT, 12 dof, normalized mutual information; followed by FNIRT), the 262-region Johnson cortical+subcortical atlas was inverse-warped to native space (nearest neighbor), and region-mean time series were extracted from the native preprocessed data. Registration quality was verified per run with automatically generated overlay images and a hemispheric-balance index; two hyperacute runs were registered through a within-subject route (Day-2 to Day-1 rigid registration composed with the Day-1 template warp), and one run (one NEH animal’s Day-2 session) failed all registration strategies and was excluded, leaving 29/30 runs and 14 animals with complete two-session data (control n=6, NEH n=4, Sanguinate n=4). All warps, orientation decisions, and QC images are provided with the code.

Human MRI and preprocessing. Acquisition parameters are given in Table 1. Both human datasets were processed with a single FSL-based pipeline mirroring the canine chain: MCFLIRT motion correction (with framewise-displacement computation), brain extraction (BET on the mean EPI; T1 neck-cropped with robustfov then BET), 6 mm FWHM spatial smoothing, grand-mean normalization, 100 s high-pass filtering, registration of the mean EPI to the subject’s T1 (FLIRT, 6 dof) and of the T1 to MNI152 (FLIRT 12 dof + FNIRT), inverse warping of the Schaefer 400-region parcellation to native space, region-mean time-series extraction, and nuisance regression of the six motion parameters and a linear trend from the region series. Functional connectivity was computed as Pearson correlation between region time series and Fisher $z$-transformed.

Quality control. Human runs were excluded for maximum framewise displacement $>5$ mm (4 runs) or parcellation coverage below 320/400 regions (5 runs, reflecting large lesions or restricted field of view), leaving 59 runs (10 control, 25 stroke pre-rehab, 24 stroke post-rehab; 23 patients with both sessions). Canine exclusions are described above. Every exclusion is listed, with its reason, in the shared results files.

2.3. Manifold alignment framework

The framework compares functional connectivity between species with different brain anatomies by embedding each species’ region-level connectome in a low-dimensional space and aligning the two spaces using sparse anatomical priors. It proceeds in three stages: graph construction from connectivity data, spectral embedding, and Procrustes alignment on network-level anchors.

Unit of alignment. The manifold is constructed over brain regions, not subjects. For each species, graph nodes are the parcellated regions (the 262-region Johnson canine atlas; the Schaefer 400-region human parcellation), and each region is described by its group-averaged connectivity profile (its row of the region-by-region Fisher-$z$ matrix). Group reference matrices were computed from the canine Day-1 (pre-occlusion) sessions (n=15) and the MSC control sessions (n=10), retaining regions present in at least 80% of contributing runs (222 canine regions; 400 human regions) and averaging available entries.

2.3.1. Step 1: Graph construction

Let ${\mathbf{C}}^{\left(s\right)}\in {\mathbb{R}}^{n_s\times n_s}$ denote the group-averaged connectivity matrix for species $s$ ($n_x=222$ canine, $n_y=400$ human regions), with row ${\mathbf{c}}_i$ the connectivity profile of region $i$. We construct an affinity matrix over regions:

$$w_{ij}=exp\left(-\frac{{\Vert {\mathbf{c}}_i-{\mathbf{c}}_j\Vert }^2}{2{\sigma }_i{\sigma }_j}\right)$$ (1)

with locally scaled bandwidths ${\sigma }_i=d({\mathbf{c}}_i,{\mathbf{c}}_i^{\left(k\right)})$, the distance to the $k$th nearest neighboring region. The affinity matrix is symmetrized, and the normalized graph Laplacian is formed:

$${\mathbf{L}}_{norm}=\mathbf{I}-{\mathbf{D}}^{-1/2}{\mathbf{W}}_{sym}{\mathbf{D}}^{-1/2},d_{ii}=\sum _j{w}_{ij}.$$ (2)

2.3.2. Step 2: Spectral embedding

Eigendecomposition of ${\mathbf{L}}_{norm}$ yields eigenvalues $0={\lambda }_1\le {\lambda }_2\le \dots$ and eigenvectors ${\mathbf{v}}_1,{\mathbf{v}}_2,\dots$; the first (constant) eigenvector is discarded and each region is embedded by the next $k$ eigenvectors:

$${\mathbf{Z}}_s=[{\mathbf{v}}_2^s,\dots ,{\mathbf{v}}_{k+1}^s]\in {\mathbb{R}}^{n_s\times k}.$$ (3)

The embedding dimension $k$ and neighborhood parameter were selected by an explicit sweep over $k\in \left\{\mathrm{2,3,4}\right\}$ and $k_{nn}\in \left\{\mathrm{5,7,10}\right\}$, choosing the configuration maximizing the correspondence score defined below ($k=4$, $k_{nn}=10$); the full sweep, including configurations that performed worse, is reported with the results.

2.3.3. Step 3: Procrustes alignment on sparse network anchors

Anchors are network centroids, not individual regions. Five a priori canine–human network homologies were used: occipital$\leftrightarrow$visual, sensorimotor$\leftrightarrow$somatomotor, cingulate$\leftrightarrow$default-mode, parietal$\leftrightarrow$ dorsal-attention, and frontal$\leftrightarrow$frontoparietal-control. Canine regions were assigned to lobar networks by spatial overlap with the Johnson atlas lobe masks; human regions carry Schaefer 7-network labels. For each mapped pair, the centroid of the member regions’ embedding coordinates was computed in each species, giving matched anchor sets $\mathbf{A},\mathbf{B}\in {\mathbb{R}}^{5\times k}$. The similarity transform (rotation $\mathbf{R}$, scale $s$, translation $\mathbf{t}$) minimizing ${\Vert \mathbf{B}-s\mathbf{R}\mathbf{A}-\mathbf{t}{\mathbf{1}}^T\Vert }_F^2$ was obtained in closed form via singular value decomposition of the anchor cross-covariance (with determinant correction to exclude reflections) and applied to all canine regions:

$${\mathbf{Z}}_x^{aligned}=s^{\ast }{\mathbf{R}}^{\ast }{\mathbf{Z}}_x+{\mathbf{t}}^{\ast }{\mathbf{1}}^T.$$ (4)

Because only network centroids inform the transform, individual regions are held out of the fit and can be used for evaluation.

2.4. Evaluation metrics, baseline, and statistical inference

Correspondence score. For every canine region belonging to the three canonical homology networks (visual, somatomotor, default-mode; $n=121$ regions), we found its 5 nearest human regions in the aligned space and scored a success when their majority 7-network label matched the homologous human network. The correspondence score CS is the fraction of successes. Significance was assessed by permutation: the canine network assignment was shuffled 500 times and CS recomputed, giving an empirical null (mean $\approx$18%) and $p$-value.

Baseline. To test whether the spectral machinery adds value, we implemented a deliberately simple alternative using no embedding: each region was described by its 5-dimensional network signature (mean Fisher-$z$ connectivity to each of the five mapped networks within its own species), and each canine region was matched to the human regions with the most correlated signatures; scoring was identical to CS, including the permutation null.

Neighborhood preservation. To quantify how much local structure the embedding retains, we computed the mean overlap between each canine region’s 10 nearest neighbors in the ambient connectivity-profile space and in the final aligned embedding.

Within-species analyses. From each run’s Fisher-$z$ matrix (restricted to the labels common to the relevant comparison set), weighted graphs were formed at 15% edge density (positive weights). We computed Newman modularity $Q$ (greedy community detection), mean connectivity strength, and mean betweenness of top-decile hubs. Canine Day-1 vs. Day-2 contrasts used exact Wilcoxon signed-rank tests within arm and Mann–Whitney $U$ tests for control-vs-treated change scores; human group contrasts used Mann–Whitney $U$ (control vs. stroke) and Wilcoxon (pre- vs. post-rehabilitation), with Spearman correlation for the association between modularity change and Vasserman improvement. Following our prespecified definition, a treated animal was classified as an imaging-defined responder if its post-occlusion $Q$ remained within one standard deviation of the control post-occlusion mean. As an exploratory analysis, each animal’s Day-1 and Day-2 connectivity profiles were embedded (subject-level Laplacian eigenmap with Nyström projection of Day-2 profiles) and the displacement direction relative to the control centroid was summarized by an angle; given the two-timepoint design this is a single displacement vector per animal, not a multi-point trajectory.

Sensitivity analysis. Because several group contrasts are reported as null, we computed the minimum effect sizes detectable at $\alpha =0.05$ (two-sided) and 80% power, incorporating the asymptotic relative efficiency of the rank-based tests: Cohen’s $d\ge 0.77$ for the canine paired Day-1/Day-2 contrast ($n=14$ pairs), $d\ge 1.55$ for the canine arm contrast (6 vs. 8 change scores), $d\ge 1.07$ for the human control-vs-stroke contrast (10 vs. 25), $d\ge 0.60$ for the human pre-/post-rehabilitation contrast ($n=23$ pairs), and $\left|\rho \right|\ge 0.56$ for the outcome correlation ($n=23$). Null results should accordingly be read as excluding large effects only. Data-quality controls supporting these analyses (motion censoring thresholds, parcellation-coverage criteria, and per-run registration verification) are described above.

2.5. Reproducibility

All analyses are deterministic (seeded permutations); the complete pipeline – download scripts for the public data, registration with persisted warps and QC images, connectivity extraction, and every statistical analysis – is available in the project repository and reproduces every number in this paper from the source data.

3. Results

3.1. Connectivity recovers cross-species network homology, most effectively via simple signatures

Canine–human network homology was recoverable from connectivity alone (Fig. 1). Simple signature matching recovered held-out homologies at 54.5% top-5 correspondence over the 121 canonical-network canine regions, three-fold above the permutation null of $\approx$18% ($p=0.002$). Manifold alignment also exceeded chance – 35.5% with the sweep-selected configuration ($k=4$, $k_{nn}=10$; most configurations 26%–36%, $p\le 0.006$; full sweep in the repository) – with neighborhood preservation of 66% at $k=10$, indicating that the aligned embedding retains most local connectivity structure.

Fig. 1.

Cross-species correspondence. Left: top-5 network correspondence for the permutation null (mean), the sweep-selected manifold alignment, and the network-signature correlation baseline; both methods exceed chance ($p=0.002$, 500 permutations), and the simple baseline outperforms the manifold method. Right: per-network correspondence of the manifold alignment across the three canonical homologies, showing strong sensory but weak default-mode correspondence.

That the simpler method scored higher is itself informative: the homology signal in region-level connectivity is strong enough to be captured by direct signature matching, and at this cohort scale the additional spectral machinery conferred no advantage. We return to the practical implications in the Discussion.

Correspondence was strongly network-dependent (Fig. 1, right): visual (42.6%) and somatomotor (43.5%) homologies were recovered several-fold above their null rates, whereas default-mode correspondence was weak (10.7%). This sensory-to-associative gradient is consistent with reports that cross-species functional correspondence degrades along the evolutionary cortical hierarchy (Xu et al., 2020) and with the contested nature of the canine default-mode network (see Fig. 2).

Fig. 2.

Aligned shared embedding. First two dimensions of the aligned space (the coordinates derived from the first two non-trivial Laplacian eigenvectors ${\mathbf{v}}_2,{\mathbf{v}}_3$; unitless). Circles: human Schaefer regions (colored for the three canonical networks, gray otherwise). Triangles: canine regions after Procrustes alignment, colored by their homologous network. Sensory networks cluster across species; default-mode regions intermix with other associative cortex.

3.2. Within-network connectivity changes differ by phase

Within-network connectivity changes moved in opposite directions in the two species’ (deliberately different) disease phases (Fig. 3): canine within-network strength increased from baseline to the hyperacute post-occlusion scan (e.g., occipital ＋0.17 Fisher-$z$), whereas chronic-phase human patients showed lower within-network strength than controls in every mapped network (occipital −0.21; somatomotor −0.20). Given the small number of networks and the phase confound, we present this as a descriptive contrast between hyperacute and chronic connectivity states rather than a test of conservation.

Fig. 3.

Within-network mean connectivity change by species and phase. Canine: hyperacute post-MCAO minus pre-occlusion baseline (n=14 paired animals). Human: chronic-phase stroke (pre-rehabilitation) minus healthy controls. The canine parietal network had insufficient coverage in the post-occlusion common label set and is omitted.

3.3. Canine treatment response: No detectable group effects

Modularity was stable from baseline to the hyperacute scan in all arms (control $0.44\pm 0.06\to 0.43\pm 0.12$, $p=1.0$; treated $0.47\pm 0.06\to 0.45\pm 0.05$, $p=0.55$), with no difference in change scores between control and treated animals ($p=0.95$; Fig. 4(a)). Mean strength rose numerically in both arms (control $2.85\to 3.82$, $p=0.16$; treated $3.95\to 4.77$, $p=0.11$; arm difference $p=0.75$), consistent with the hyperacute hyperconnectivity pattern in Fig. 3 but not statistically significant at these sample sizes. Under the pre-stated imaging-derived criterion ($Q$ within 1 SD of the control post-occlusion mean), all eight treated animals with complete data classified as responders; the criterion therefore produced no responder/non-responder stratification in the real two-timepoint data, and no classification analysis was possible. The exploratory displacement analysis was likewise null: the angle between each animal’s baseline-to-hyperacute displacement and the direction of the control centroid did not differ between control (median 94°) and treated (median 100°) animals ($p=0.95$), nor did displacement magnitude ($p=0.49$).

Fig. 4.

Modularity by group. (a) Canine per-animal modularity at baseline (Day 1) and in the hyperacute window (Day 2, $\sim$2 h post-MCAO); lines connect the same animal (control n=6, NEH n=4, Sanguinate n=4 with complete pairs). (b) Human group means ($\pm$SD): MSC controls (n=10), stroke pre-rehabilitation (n=25), stroke post-rehabilitation (n=24). No group contrast is significant.

3.4. Human chronic-phase analyses: Null group effects

Chronic-phase patients did not differ from controls in modularity (0.33 ± 0.09 vs. 0.33 ± 0.05, $p=0.93$) and showed a numerically but not significantly lower mean strength (29.6 ± 7.3 vs. 32.4 ± 6.1, $p=0.32$; Fig. 4(b)). Modularity did not change measurably across rehabilitation (pre 0.33 ± 0.09, post 0.35 ± 0.09, $p=0.46$, n=23 paired), and the change in modularity was unrelated to improvement on the Vasserman aphasia scale (Spearman $\rho =-0.04$, $p=0.87$).

4. Discussion

This study asked whether canine–human functional network homology can be recovered from connectivity alone, and how much methodological machinery that recovery requires. The first answer is clearly affirmative, and is to our knowledge the first quantitative, region-level connectivity-based demonstration of functional correspondence between a carnivore and humans: using only five network-level anatomical priors, held-out homologies were recovered up to three-fold above chance, and the correspondence hierarchy – strong for visual and somatomotor networks, weak for default-mode – extends to the dog the gradient previously described within primates, whereby functional correspondence attenuates along the evolutionary cortical hierarchy (Margulies et al., 2016, Xu et al., 2020). Functional organization alone thus carries substantial homology information even across a primate–carnivore divide, and sparse network-level priors are a viable substitute for dense anatomical registration when the goal is coarse functional correspondence.

The second answer carries the practical guidance. The homology signal proved strong enough that the simplest method captured it best: matching regions directly by the correlation of their network-level connectivity signatures – a baseline with no eigendecomposition, no alignment, and no tunable dimensionality – outperformed the manifold framework (54.5% vs. 35.5%). At the cohort sizes realistic in cross-species work (here, group connectomes built from 15 and 10 subjects), the variance cost of estimating a data-driven manifold evidently outweighs its representational benefits, and simple baselines should be reported before more complex machinery is credited. This mirrors the broader lesson from within-species connectivity modeling, where increasingly sophisticated architectures (e.g., Li et al., 2020, Li et al., 2022, Liu et al., 2024) are typically validated on cohorts an order of magnitude larger than those available translationally.

The within-species analyses were null. In the canine model, connectivity strength rose numerically in the first hours after permanent MCAO – compatible with reports of hyperacute hyperconnectivity – but neither this rise, nor modularity, differed between treated and untreated animals, and our prespecified imaging-derived responder criterion did not stratify treated animals at all (every treated animal classified as a responder). In the chronic-phase human cohort, patients were indistinguishable from controls on modularity, and connectivity change did not track rehabilitation outcome. We report these results without embellishment: with 14 paired animals and 23 paired patients, only large effects would have been detectable, and none were present.

These treatment-arm nulls warrant explicit reconciliation with the source experimental studies. In the larger parent cohorts, voxelwise ICA-based analyses reported NEH-associated preservation of connectivity within specific higher-order visual and parietal resting-state networks (Warioba et al., 2024b) and treatment-associated preservation of network integrity more broadly (Warioba et al., 2024a). The present analyses operate at a different granularity – region-averaged, whole-network graph summaries in the two-session subset (n=14; arm-contrast sensitivity $d\ge 1.55$) – at which spatially focal, component-level effects are expected to dilute. Consistent with this, within-network mean connectivity changes did not differ by arm here (all $p\ge 0.6$), while never favoring controls over treated animals. The present null results therefore neither reproduce nor contradict the published voxelwise treatment effects; they indicate that those effects, if present in this subset, are not large at the network-summary level.

Limitations. First and most fundamentally, the canine (hyperacute, experimentally controlled MCAO) and human (chronic aphasia rehabilitation cohort) datasets probe different disease phases; the cross-species comparison rests on functional-organizational homology, not on matched pathology, and the divergent direction of network changes (Fig. 3) illustrates exactly this. Public human acute-phase resting-state stroke data suitable for this design do not currently exist on OpenNeuro; a prospectively matched cohort is the necessary next step. Second, sample sizes are small throughout: per the sensitivity analysis, only large effects ($d\gtrsim 0.8$–1.6 depending on contrast) were detectable, so the null results exclude large effects only. Third, aphasia severity (Vasserman) was the only clinical outcome available for the human cohort. Fourth, canine EPI-to-template registration is intrinsically challenging (small brains, anisotropic coverage, extracerebral tissue); although every run was individually quality-controlled and one unresolvable run was excluded, residual registration error may attenuate canine connectivity estimates. Fifth, the responder criterion was imaging-derived; the absence of stratification may reflect the criterion, the two-timepoint design, or a true absence of early network-level treatment effects.

5. Conclusions

Canine–human network homology is recoverable from functional connectivity alone, using only sparse network-level anatomical priors, and the recovered correspondence follows the evolutionary cortical hierarchy previously described within primates. The homology signal is robust enough that simple connectivity signatures capture it best: at realistic translational cohort sizes, spectral manifold alignment offered no advantage over direct signature matching. Neither hyperacute canine treatment effects nor chronic human group effects were detectable at the network level within the sensitivity bounds reported. Together, these fully reproducible results establish connectivity-based cross-species mapping beyond primates and provide a benchmark – methodological and statistical – for its translational use.

CRediT authorship contribution statement

Chisondi S. Warioba: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Timothy J. Carroll: Resources, Project administration, Investigation, Funding acquisition. Gregory A. Christoforidis: Supervision, Project administration.

Ethical statement

All animal procedures were approved by the institutional Animal Care and Use Committee (IACUC protocol #72498) and conducted in accordance with institutional and national guidelines for the care and use of laboratory animals. Human data were obtained exclusively from publicly available, de-identified OpenNeuro datasets (ds003999, ds000224) collected under their originating institutions’ ethical approvals; no new human data were acquired for this study.

Funding

This work was supported by the National Institutes of Health (R01NS093908) and the National Science Foundation Graduate Research Fellowship Program (2140001).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Human data are openly available from OpenNeuro: ds003999 (doi:10.18112/openneuro.ds003999.v1.0.2) and ds000224 (doi:10.18112/openneuro.ds000224.v1.0.4). Canine raw data are available from the corresponding author on reasonable request (cwarioba@stanford.edu). All analysis code — including data download, the registration pipeline with persisted warps and per-run QC images, connectivity extraction, and every statistical analysis and figure — is available at https://github.com/csw-research/cross-species-manifold-alignment (MIT license) and reproduces every number reported here.