Do independently fitted cohorts recover similar networks?
Each learned network describes a pattern of relationships among the same 100 cortical ROIs. Correlating corresponding matrix entries asks whether a control network and an ASD network place positive and negative weights on similar ROI pairs, even if their network numbers differ.
Control–ASD findings
These findings describe the default control–ASD comparison. They do not test a diagnostic-group effect or demonstrate independent recovery in every participant.
Every control network against every ASD network
Inspect a network pair
Each point is the weight of the same ROI pair in the two cohorts. A tight upward trend indicates similar spatial organization; a downward trend indicates reversed orientation. Diagonal entries are excluded from this scatter unless enabled above.
Best match for each row network
| Control | ASD | r | |r| |
|---|
What is being compared?
The matrices are equal-participant means of the subject-calibrated observation coefficient planes within each cohort’s own frozen latent basis. They are the same spatial quantities used for functional-system matching, before separating their positive and negative poles. No expression-amplitude weighting, pruning, regional summing or reference-system labels enter this calculation.
For each pair, Pearson correlation subtracts each matrix’s average entry and divides by its spread. It measures resemblance of the spatial weight pattern, not its absolute strength: a weak and a strong copy have the same correlation. Off-diagonal entries are the default; the optional diagonal includes each ROI’s self-term. The matrices are symmetric, so retaining both ordered off-diagonal entries repeats each undirected pair without changing r.
A negative r can indicate a similar pattern with opposite orientation; it does not by itself establish equivalence of the models’ dynamics. Independently fitted latent spaces can also rotate or mix components, so weak individual matches do not establish absence of a shared pattern. Transition operators and expression traces would need consistent transformation before relabeling or sign-flipping a model.
These cohorts differ in membership and are not matched for age, sex or motion. Both use the same Schaefer100 atlas, min–max normalization and rank-12 recipe. Matrices contain fitted second-moment coefficients, not ordinary covariance or direct physiological connections. Matrix entries are dependent: no edge-wise p-values or claims of diagnostic significance are attached.