Fig. 3: Deep learning model with tissue-specific eRNA and canonical gene GReX as features predicts three-dimensional contact frequency.
a Grid search across 13 hyperparameters was used to find the optimal model architecture (see Methods). b The neural network was trained in cerebellum GReX and contact frequency data for 50 epochs, achieving a mean prediction R2 of 0.37 across the validation folds and 0.38 in the independent test set. Within a second tissue type not previously seen by the model, whole blood, we observed a prediction R2 of 0.18, which shows some cross-tissue portability. c Contact frequency prediction in the cerebellum test set (denoted by color) as a function of the GReX of the upstream and downstream transcript levels (x and y axes, respectively). The two-dimensional GReX space for contact pairs is constrained to lie in the colored region. d SHAP values representing the relative mean contributions of the upstream and downstream transcripts to contact frequency predictions, which we found to be 46.94% and 53.06%, respectively. Source data are provided as a Source Data file.
