Shared Living Environment and Monozygotic Twins Connectomes: Classifying Cohabitation Status from Resting-State fMRI

Environmental neuroscience examines how the factors of the settings we live in are reflected in brain structure and function. One of the major challenges in this field is separating the effects of current lived environment from genetics and upbringing. The monozygotic twin design controls both of these sources of variance by construction. The twin pair cohabitation status at the time of brain scan acquisition is used in this work as a label that represents the environment signal potentially stored in connectome. The primary research question is whether cohabitation status can be recovered from the resting-state functional connectivity of a monozygotic twin pair, which, to our knowledge, has not been explored earlier. The second question is methodological: whether Siamese graph neural networks, which preserve brain topology, identify this signal better than linear classifiers. This analysis uses data from 127 monozygotic twin pairs that passed quality control, 35 cohabiting and 92 separated, collected within the TwinLife Environment panel. Machine learning and graph deep learning stages were evaluated on the same pairs, the same frozen cross- validation folds and the same metrics. The machine learning stage evaluated the performance of four linear classifiers (regularized logistic regression, support vector machine, ridge classifier, correlation kernel ridge regression) on 21 feature representations variants (including similarity statistics, edge- and node-level representations, and confound benchmarks). The graph-based learning stage evaluated twelve Siamese architectures that alternate in model depth, aggregation, update, or readout functions. Additionally, four loss functions, each treating class imbalance differently, were utilized for fitting the depth ladder models. The study cannot reliably distinguish no presence of cohabitation effect from a small signal, so the near-null result should not be interpreted as a complete rejection of the hypothesis. In this cohort, the cohabitation status of twin pairs is structured by their age more strongly rather than their functional connectivity. Graph deep learning models did not provide competitive advantage over regularized linear classifiers tested. Improving the study design is therefore likely to be more valuable than increasing model complexity.