A personal reference curve for brain wiring: BRIDGE cuts brain-age error by 3%–9% and ties deviation scores to ADOS in autistic boys
Synopsis
Brain-connectivity differences in autism and other neurodevelopmental disorders have mostly been read from group-average case-control contrasts, which flatten individual variation. BRIDGE combines a graph variational autoencoder with normative modeling to learn an age-varying reference trajectory of structural connectomes from 1,164 neurotypical participants, then reports a global deviation score from brain-age difference and an individualized regional deviation map from regional embeddings. In 229 autistic children, the global score correlated with ADOS (Spearman R=−0.291) and IQ (Pearson R=0.195) in males, and brain-age prediction error was 3%–9% lower than comparable deep-learning models.
Interpretation
Developmental deviation in structural brain connectivity can now be quantified per person rather than only as a difference between group averages. Earlier connectome studies mostly used a case-control paradigm, comparing group-average connectivity between a neurodivergent group and matched neurotypicals, so within-group variation was not captured. Across 968 neurotypical participants from the Philadelphia Neurodevelopmental Cohort (PNC) and 196 neurotypical plus 229 autistic participants from the Center for Autism Research, connectomes generated by BRIDGE showed the lowest Kolmogorov–Smirnov dissimilarity to actual connectomes in degree, clustering coefficient, betweenness centrality and edge length, ahead of GCN-VAE, FC-VAE and 13 classic generative models.
The model compresses each person's structural connectome into a low-dimensional latent space that is organized by age. Traditional normative models struggle with the network structure of connectomes: modeling edges directly ignores topological interdependence, and predefined graph measures are insensitive to developmental change. An encoder maps the connectome's upper triangle into a 40-dimensional latent space, and a decoder generates connectomes under two wiring constraints, cost minimization and Hebbian learning, with brain age predicted by a linear layer on the latent features; training used neurotypical participants only, with PNC pretraining and CAR fine-tuning.
Distance from the reference trajectory is split into one global score and a region-by-region map. Group-average methods poorly explain individuals who share a diagnosis but differ in presentation, and region-level deviation had not been reported per individual on connectomes. Gaussian process regression fits the neurotypical age trajectory and variance for each region's embedding, and a Z score gives regional deviation; the global score is the age-bias-corrected brain-age difference, which in autistic males correlated with ADOS (Spearman R=−0.291, 95% CI [−0.414,−0.158]) and IQ (Pearson R=0.195, 95% CI [0.049,0.333]), with the ADOS correlation remaining significant after controlling for IQ (R=−0.221).
Perspective
The demonstrated setting is autistic and neurotypical children and adolescents aged 6 to 19, with structural connectomes built from diffusion MRI probabilistic tractography, and all downstream analyses conducted within the CAR cohort acquired on a single scanner under one protocol. Correlation analysis included only autistic males because the female sample was limited.
The authors note that normative models may need to be tailored to population groups defined by sex, race/ethnicity and social determinants of health; CAR's 30 gradient directions are fewer than PNC's 64, and pretraining plus fine-tuning does not remove the limits imposed by the target data's lower angular resolution; geometric distance came from an adult template aged 22 to 37 and may not capture geometric constraints and metabolic costs that vary across developmental stages; and the cohort excludes children under 6.