Hierarchical and non-hierarchical network flows generate complementary representational dynamics in human visual cortex

Name: Verified Rigor Record — Hierarchical and non-hierarchical network flows generate complementary representational dynamics in human visual cortex
Item: Hierarchical and non-hierarchical network flows generate complementary representational dynamics in human visual cortex
Rating: 79
Author: Αlpha¹

Tzalavras, A.; Osher, D. E.; Cocuzza, C. V.; Nallan Chakravarthula, L.; Mill, R. D.; Peterson, K. L.; Cole, M. W.

Paper source

Read PDF

Αlpha¹ Rigor Review

Rigor Review Verified

Critical

High-priority

Action items to resolve · Critical / High

Confidence: moderateEngine: 7.6.0Source: full text

Not verified in this report

No reported statistical tests were found to recompute.

This Alpha1 Rigor Review uses AI Rigor Reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

The paper presents a computational analysis of fMRI connectivity using empirical neural network models; strengths include a clear premise, detailed methods, and ethical compliance. Weaknesses include missing power analysis, vague inclusion/exclusion criteria, no outlier handling, incomplete data/code sharing, and statistical reporting gaps (assumptions, exact p-values, effect sizes). The paper is generally robust but has several reporting gaps that readers should weigh.

Three independent reviewers evaluated the published paper; their assessments were synthesized. The statistics verification component did not recompute any tests due to insufficient detail (coverage note). The citation check found no retracted or missing references. The integrity check flagged a minor concern about overly precise percentages of significant vertices.

12 major claims checked against the paper's own evidence: all adequately supported.

partialResults, p.14-15Reviewers 1, 3
Network flow distance and representational distance converge on the classic V1↔V2↔V3↔V4 hierarchy.
Both measures show the ordering V2 > V3 > V4 distance from V1, consistent with hierarchy. However, the mediation analysis is indirect; the convergence is qualitative (order) rather than a direct formal test.
Evidence: Figures 2 and 3 show communicability and representational distance ordered V2 > V3 > V4. Paired t-tests show significant differences between regions.
“Both approaches yielded the same result, with region V2 demonstrating the highest communicability with V1, followed by V3 and finally V4”
partialResults, p.10Reviewer 1
The GC-PCR method yields confound-controlled connectivity estimates that are more direct than standard Pearson correlation.
The paper validates GC-PCR connections using DWI tractography, showing increased directness. However, the main ENN and lesion results do not directly compare GC-PCR to a standard FC method, so the superiority is asserted rather than demonstrated empirically in this paper.
Evidence: Methods and Discussion: 'we also validated our connections using DWI tractography, increasing confidence that our direct pathway estimates were not false positives'.
“we also validated our connections using DWI tractography, increasing confidence that our direct pathway estimates were not false positives due to network confounding.”
partialMethods, FC estimation; Results, Connectivity-based and representation-based testsReviewer 2
The GC-PCR connectivity method yields confound-controlled and tractography-validated estimates.
The paper describes the method and states validation using DWI tractography, but no quantitative validation results are presented in the main text (e.g., correlation values between GC-PCR and tractography).
Evidence: The text states 'we also validated our connections using DWI tractography' but no numerical results are given.
“As before, we also validated our connections using DWI tractography, increasing confidence that our direct pathway estimates were not false positives due to network confounding.”
supportedResults, p.16-17Reviewer 1
Hierarchical and direct pathways make distinct contributions to visual functionality: hierarchical pathways reduce dimensionality, direct pathways preserve high-dimensional information.
The ENN lesion experiments show dimensionality decreases significantly across hierarchical steps (e.g., V1 > V4, p=0.0013) and direct pathways retain higher dimensionality across steps. The claim is supported by the presented data.
Evidence: Figure 5 and accompanying statistics: 'dimensionality decreased significantly from V1 to V4 (V1 > V4, p = 0.0013)' and 'the direct pathways retained higher dimensionality across steps, whereas the hierarchical pathways demonstrated decreases'.
“Paired Wilcoxon tests confirmed significant model differences across all steps following the initial step”
supportedResults, p.17-18Reviewer 1
Hierarchical pathways contribute more to large receptive field sizes and peripheral representations.
Correlations between hierarchy index and eccentricity/receptive field size were significant (e.g., right V3 r=0.368 for eccentricity). The complementary in silico experiment (difference-of-deviations) also showed positive correlations. The evidence supports the claim.
Evidence: Figure 6 and text: 'for regions V3 and V4, all correlations between the hierarchy index, eccentricity and receptive field size were significant'.
“Consistent with our hypothesis, for regions V3 and V4, all correlations between the hierarchy index, eccentricity and receptive field size were significant.”
supportedResults, p.16Reviewer 1
Both hierarchical and direct pathways contribute to visual function, with direct pathways being faster (early-step propagation).
The ENN lesion model shows that at early steps (1-2), the direct-pathway-only model has significantly higher R² than the hierarchical-only model, supporting faster processing. At later steps, both converge. This is directly supported.
Evidence: Figure 4 and statistics: 'demonstrating significant differences for the first 2 steps'.
“demonstrating significant differences for the first 2 steps”
supportedResults, p.15-16Reviewer 1
V1-initiated ENN accurately generates retinotopic time series and pRF parameters.
The model shows significant correlations with actual data at both parcel and vertex levels (all p<0.0001 for parcels; vertex-level significant via MaxT). pRF parameters (eccentricity, angle) are also generated accurately (except right V4 angle). The evidence supports the claim with appropriate caveats.
Evidence: Figures 3-4 and text: 'Generated-to-actual similarities were significantly above chance for all six regions (all p<0.0001)' and parcel-level correlations for pRF parameters.
“All region-level correlations were statistically significant (p < 0.0001), except for the angle comparison for right V4”
supportedResults, In silico ENN lesion experiments sectionsReviewer 2
Hierarchical and direct pathways make distinct and complementary contributions to visual functionality.
The paper provides evidence from in silico ENN lesion experiments showing distinct effects on dimensionality and receptive field size, supporting the claim.
Evidence: In silico lesion experiments show hierarchical pathways reduce dimensionality more than direct pathways, and hierarchical pathways contribute more to large receptive fields/peripheral representations.
“We found that both hierarchical and direct pathways contribute to visual function, supporting a view of early visual cortex in which hierarchical processing coexists with parallel routes of information propagation.”
supportedResults, In silico ENN lesion experiment: Hierarchical pathways compress visual representationsReviewer 2
Hierarchical pathways compress visual representations (reduce dimensionality).
The paper demonstrates significantly lower participation ratio in hierarchical vs. direct lesion models across steps, consistent with the claim.
Evidence: Figure 4 shows participation ratio differences; text states paired Wilcoxon tests confirmed significant model differences across all steps following the initial step.
“the direct pathways retained higher dimensionality across steps, whereas the hierarchical pathways demonstrated decreases as the system reached its steady state”
supportedResults, Network flow distance and Representational transformation distanceReviewer 2
The classic V1-V2-V3-V4 hierarchy is recovered by both network flow distance and representational distance.
Both communicability and representational similarity analyses show V2 closest to V1, then V3, then V4, consistent with the classic hierarchy.
Evidence: Figures 1 and 2 show the statistical comparisons; text states V2 > V3 > V4 in communicability and representational distance.
“Both approaches yielded the same result, with region V2 demonstrating the highest communicability with V1, followed by V3 and finally V4”
supportedResults, In silico ENN lesion experiments: Hierarchical and direct pathways generate visual functionality collaboratively, with more rapid processing via direct pathwaysReviewer 2
Direct pathways are faster than hierarchical pathways in generating visual representations.
The lesion experiments show that at early steps (1-2), the direct-lesion model has higher R^2, indicating faster propagation.
Evidence: Paired Wilcoxon tests show significant differences for the first 2 steps, with direct model having higher R^2.
“direct connections drive early-step propagation, whereas hierarchical connections become increasingly important at later steps”
supportedResults, In silico ENN lesion experiment: Large receptive fields and peripheral representations rely more on the hierarchical pathwaysReviewers 2, 3
Large receptive fields and peripheral representations rely more on hierarchical pathways.
The hierarchy index positively correlates with eccentricity and receptive field size, and the supplementary analysis shows hierarchical model produces more accurate large/pRF values.
Evidence: Correlations between HI and eccentricity/receptive field size are significant for V3 and V4; the model comparison shows positive correlations with actual pRF values.
“consistent with our hypothesis, for regions V3 and V4, all correlations between the hierarchy index, eccentricity and receptive field size were significant.”

1 integrity concern flagged (0 high).

Rigor

Eight Dimensions3 pass · 5 warn

12 major claims3 partial9 supported

10 copyedit issues flagged: mostly clarity, consistency, typo.

Checked 56 references: 52 verified — 4 not checked.

Reproducibility

1 data/code link checked; 1 live.

datahttps://colelab.github.io/ActflowToolbox/LIVEHTTP 200
Resolves, but the content could not be matched to the paper.

Alpha¹ Rigor Action ItemsFinal pass before journal submission

In post-publication audit, the paper is robust in its core premise and methods but has several reporting gaps that warrant a correction or clarification, particularly regarding data/code availability and statistical reporting. An informed reader should weigh the missing power analysis, lack of outlier handling, and incomplete code sharing when evaluating the reproducibility of the findings.

1.
HIGHcopyedit
Correct the swapped definitions of direct and hierarchical pathways in Methods (p.9): currently the text states 'For the direct pathways, we lesioned V2-V3 and V3-V4 connections. For the hierarchical pathways, we lesioned V1-V3 and V1-V4 connections.' but later text defines direct pathways as V1-V3 and V1-V4. Reverse the definitions to match the consistent usage.
This is a factual error that could fundamentally mislead readers interpreting the lesion experiments.
2.
HIGHdata code
Deposit custom analysis code (GC-PCR implementation, ENN lesion scripts, dimensionality analyses) in a public repository (e.g., GitHub, Zenodo) with a permanent DOI and cite it in the paper.
All three reviewers flagged missing code sharing as a major reproducibility gap.
3.
HIGHdata code
Add a dedicated data availability statement specifying the exact HCP dataset release (e.g., HCP S1200 Release) and accession number/DOI, and state that derived FC matrices and ENN outputs are available upon request or in a repository.
No data availability statement is present; this is a critical reporting requirement for a data-driven study.
4.
HIGHreporting
Report an a priori power analysis or justify the sample size of N=40 based on prior effect sizes in activity flow modeling studies.
Missing power analysis was flagged by two reviewers; sample size justification is expected for observational studies.
5.
HIGHreporting
Specify explicit inclusion/exclusion criteria for subjects (e.g., age range, head motion thresholds, data quality metrics) and for vertices/parcels (e.g., minimum signal-to-noise).
Current criteria are limited to 'selected at random'; transparency about subject selection is essential for reproducibility.
6.
HIGHreporting
Describe how outliers were handled in FC estimation and pRF modeling, or state that no outliers were excluded.
Outlier handling is not mentioned; this is a potential source of bias.
7.
HIGHstatistics
Report effect sizes (Cohen's d or equivalent) and 95% confidence intervals for all t-tests, not just for Wilcoxon tests.
Effect sizes are only reported for Wilcoxon tests; consistent reporting of effect sizes is needed for meta-analysis and interpretation.
8.
HIGHstatistics
Verify and report test assumptions (normality, equal variance) for parametric tests, or explicitly state that non-parametric alternatives were used throughout.
Assumptions are not verified; this undermines confidence in parametric test results.
9.
MEDIUMreporting
In the Vertex level model validation results, clarify the overlapping percentages of significant vertices by stating the number of vertices instead of (or in addition to) the percentages, and explain that the percentages sum to >100% because they overlap.
The percentages are unusually precise and the overlap is confusing; the integrity check flagged this as a minor concern.
10.
MEDIUMreporting
Provide version numbers for all software used (scipy, sklearn, astropy, Connectome Workbench, etc.).
Software version numbers are required for reproducibility; currently missing.
11.
MEDIUMreporting
Reference an appropriate reporting guideline (e.g., STROBE for observational neuroimaging studies) in the methods.
Reviewer 2 noted the absence of a reporting guideline; including one improves transparency.
12.
MEDIUMreporting
State whether the 'replication cohort' results mentioned in figure captions come from an independent sample or a subset of the same dataset, and describe the cohort in the main text.
The replication cohort is not described; this is needed to assess the independent replication claim.
13.
LOWcopyedit
Remove the duplicate URL in Methods (p.7): 'https://colelab.github.io/ActflowToolbox/ https://colelab.github.io/ActflowToolbox/' should appear once.
Duplicate URL is a minor copyedit issue.
14.
LOWcopyedit
Fix missing space around '<' in p-value (Results, p.15): change 'p<0.0001' to 'p < 0.0001'.
Typographical error; consistent formatting improves readability.
15.
LOWcopyedit
Fix subject-verb agreement in Methods, p.10: change 'the directness ... were validated' to 'the directness ... was validated'.
Grammatical error.
16.
LOWcopyedit
Consider aligning terminology: the title uses 'non-hierarchical' while the text consistently uses 'direct'; use one term throughout for consistency.
Terminological inconsistency may confuse readers.

Full text (66,850 chars)Analyzed: Jul 7, 2026

Αlpha¹ assesses methodological rigor, not the importance of the findings. See how we calibrate →