Trait-conditioned Deep Generation of Connectomes
Poster No:
1170
Submission Type:
Abstract Submission
Authors:
Yuanzhe Liu1, Caio Seguin2, Maria Di Biase3, Andrew Zalesky4
Institutions:
1The University of Melbourne, Parkville, VIC, 2University of Melbourne, Melbourne, Victoria, 3The University of Melbourne, Melbourne, VIC, 4The University of Melbourne and Melbourne Health, Melbourne, VIC
First Author:
Co-Author(s):
Introduction:
An individual's connectome is unique, and variation in connectomes associates with various individual traits (Griffa, 2013; Puxeddu, 2020). While models to predict individual traits from connectomes are abundant (Kawahara, 2017; Kopetzky, 2024), the inverse task - inferring connectome from individual traits - has not been widely studied. In this work, we aim to i) build a model that generates connectomes based exclusively on individual traits, ii) evaluate whether generated connectomes can capture empirically observed interindividual variation, iii) compare different trait categories' contribution to capturing variation, iv) explore the model utility in a data augmentation context.
Methods:
We used data of 8,086 healthy participants from the UKBiobank (Sudlow, 2015). We considered 194 individual traits, grouped into 4 categories: age and sex, lifestyle factors, body phenotypes, and cognitive measures. Previously mapped probabilistic structural connectomes in the 200-node Schaefer parcellation were used (Mansour, 2023). Participants were randomly assigned into training and test sets (80/20 split). A deep generative model (Fig. 1a) was trained to generate connectomes for test set individuals from their traits.
To determine the interindividual variation explained by the model, generated connectomes were compared to their empirical counterparts (Fig. 1b, 1c), in terms of identifiability (Amico, 2018) and network measures (Rubinov, 2010) (Fig. 1d). To assess significance, we benchmarked the results against a null where all personal information were suppressed (i.e., replaced by the training set group mean).
We next evaluated the contribution of each trait category to connectome generation, by iteratively suppressing traits by category from the model inputs. The contribution of a trait category was quantified by the drop in captured variation when it was suppressed, relative to the input-complete model.
Finally, we investigated the use of generated connectomes as augmented data to improve the training of machine learning models. We trained L1-regularized linear/logistic regression models for age and sex prediction from connectivity, with the training set comprising either i) empirical: all empirical data, ii) generated: all generated data, and iii) hybrid: equal-sized empirical and generated data, for a range of sample sizes. We evaluate whether including generated data can improve model training.
To determine the interindividual variation explained by the model, generated connectomes were compared to their empirical counterparts (Fig. 1b, 1c), in terms of identifiability (Amico, 2018) and network measures (Rubinov, 2010) (Fig. 1d). To assess significance, we benchmarked the results against a null where all personal information were suppressed (i.e., replaced by the training set group mean).
We next evaluated the contribution of each trait category to connectome generation, by iteratively suppressing traits by category from the model inputs. The contribution of a trait category was quantified by the drop in captured variation when it was suppressed, relative to the input-complete model.
Finally, we investigated the use of generated connectomes as augmented data to improve the training of machine learning models. We trained L1-regularized linear/logistic regression models for age and sex prediction from connectivity, with the training set comprising either i) empirical: all empirical data, ii) generated: all generated data, and iii) hybrid: equal-sized empirical and generated data, for a range of sample sizes. We evaluate whether including generated data can improve model training.
Results:
Evaluated with various identifiability and network measures, model-generated connectomes consistently explained a significantly greater interindividual variation in the empirical data than the null (Fig. 2a-2b; permutation test, p<0.001).
The contribution to captured variation significantly differs between the trait categories evaluated, with body phenotypes and age and sex contributing more than lifestyle factors and cognitive measures when averaged across all identifiability and network measures (Fig. 2c). Notably, while the trait contribution composition in many measures aligns with the average pattern, they can distinctively deviate from the average in others (Fig. 2d-e).
We found that models trained on hybrid data performed the best when sample sizes are small (Fig. 2f). However, performance was eventually surpassed by models trained on empirical data with large sample sizes. Comparing hybrid against empirical with the same number of empirical data, we found that including generated data can improve model performance (Fig. 2g).
The contribution to captured variation significantly differs between the trait categories evaluated, with body phenotypes and age and sex contributing more than lifestyle factors and cognitive measures when averaged across all identifiability and network measures (Fig. 2c). Notably, while the trait contribution composition in many measures aligns with the average pattern, they can distinctively deviate from the average in others (Fig. 2d-e).
We found that models trained on hybrid data performed the best when sample sizes are small (Fig. 2f). However, performance was eventually surpassed by models trained on empirical data with large sample sizes. Comparing hybrid against empirical with the same number of empirical data, we found that including generated data can improve model performance (Fig. 2g).
Conclusions:
This study introduced a deep model that successfully generates individual connectomes with empirical interindividual variation, based on solely personal profiles. While biological traits are more influential predictors, behavior/cognition also play an indispensable role. We also show that connectome generation can be a valuable data augmentation approach to small cohort studies. Our results are significant because it infers individual connectomes without brain/neuroimaging data.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
Methods Development 2
Keywords:
Computational Neuroscience
Machine Learning
Other - generative model; connectome; network neuroscience; identifiability
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Griffa, A. (2013). Structural connectomics in brain diseases. Neuroimage, 80, 515-526.
Kawahara, J. (2017). BrainNetCNN: Convolutional neural networks for brain networks; towards predicting neurodevelopment. Neuroimage, 146, 1038-1049.
Kopetzky, S. J. (2024). Predictability of intelligence and age from structural connectomes. PloS one, 19(4), e0301599.
Mansour L, S. (2023). Connectomes for 40,000 UK Biobank participants: a multi-modal, multi-scale brain network resource. Neuroimage, 283, 120407.
Puxeddu, M. G. (2020). The modular organization of brain cortical connectivity across the human lifespan. Neuroimage, 218, 116974.
Rubinov, M. (2010). Complex network measures of brain connectivity: uses and interpretations. Neuroimage, 52(3), 1059-1069.
Sudlow, C. (2015). UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS medicine, 12(3), e1001779.