Resting-state functional magnetic resonance imaging enables the exploration of the functional brain organization and its representation via large-scale networks. Summary measures, such as functional connectivity profiles or networks' spatial topography, present inter-individual differences with applications in detecting cognitive disorders or predicting behavioral traits. However, the accuracy of these measures can be significantly disrupted due to the limited number of subjects in clinical data. Transfer learning coupled with Bayesian modeling can be leveraged to overcome this issue. We simultaneously estimate the posteriors for the individual cortical topography and functional connectivity profiles by applying a variational inference technique over probabilistic graphical models. After extracting these features, inference performance is evaluated through the regression of behavioral scores. The Bayesian formalism allows information transfer by sharing parameters pre-trained on a large dataset, especially the posteriors inferred at a population level. This way, on the HCP dataset, we show that the knowledge captured on a large sub-population ($\sim$750 subjects and 4 scans per subject) helps improve the model trained on a much smaller sample (50 to 250 subjects with 1 scan per subject), even if this sample stems from another dataset, here CamCAN. By doing it with 50 subjects, we achieve comparable performance in behavioral prediction than a three times larger dataset.