Fig. 2

SupirFactor benchmark and hyperparameter evaluation. Performance is evaluated on gold-standard networks consisting only of edges held out of the prior network \(\varvec{P}\), measuring recovery using AUPR. \(R^{2}\) is computed on a validation set of \(50\%\) of the data samples held out of the training data. Data sets are labeled for species, with Bacillus subtilis (B1 and B2), S. cerevisiae bulk RNA expression (S1 and S2), and S. cerevisiae single-cell RNA expression (scY). A Comparing network interpretation using model weights (\(\Theta\)) to interpretation using explained relative variance (ERV), measured using AUPR against edges held out of the prior network \(\varvec{P}\). \(R^{2}\) is calculated from the full network model. Model hyperparameters are set based on results in Additional file 1: Figs. S2-S9 as detailed in the “Model regularization” section. B Comparing normalization and activation functions for single-cell RNA expression data, as in (A). SS- is normalizing to mean of zero and unit variance, RM- is normalizing to maintain a minimum value of 0 (retaining sparsity), -L is linear activation, and -R is ReLU activation. C Benchmarking SupirFactor with optimal parameters selected from (B) against a comparable GRN inference method, the Inferelator, and two methods not using prior evidence; GRNBoost2 and GENIE3. D Comparing multi-context network performance between shallow SupirFactor, Hierarchical SupirFactor, and the multi-task Inferelator. GRNs are learned from single-cell (scY) data, with context/task groupings determined by growth condition. Global GRNs are learned from the data without separate groupings (using StARS-LASSO for the Inferelator [41]). Context networks are computed post-training in SupirFactor and split here on growth condition. E Evaluating model prediction \(R^2\) on four novel test data sets, using a GRN trained by Hierarchical SupirFactor. F Recovery of independently collected regulatory evidence not in the prior. Comparing the full Hierarchical SF with the Inferelator, GRNBoost2, and GENIE3 on using the trained single cell Yeast (scY) models. G Comparing contextual network for GRNs defining cell cycle M-phase and S-phase (Table 1). Each point is an interaction from the two contextual networks, colored by the target gene functional annotation. X and Y axis are \(\xi ^{2}\) of the S-phase and M-phase networks. GRN interactions targeting S-phase genes (purple) have higher ERV in the S-phase contextual network, and interactions targeting M-phase genes (green) have higher ERV in the M-phase contextual network