Fig. 2: Benchmarking algorithms on simulated data.

a Simulation framework for scMTNI. We first simulate GRNs for cell types across a cell lineage tree. Next, we generate in silico single-cell gene expression data for each cell type using BoolODE using the simulated GRNs and add 80% zeros in the simulation data. Then, we apply five multi-task learning algorithms and three single-task learning algorithms for GRN inference to the simulated datasets and predict networks in stability selection framework. We compare the performance of these algorithms based on area under precision and recall curve (AUPR) and F-score of top edges. b AUPR comparing inferred networks to ground truth networks of simulated datasets 1, 2, 3. c F-score comparing top K edges in the inferred networks to those in the ground truth networks of simulated datasets 1, 2, 3 (cell type 1: K = 202, cell type 2: K = 217, cell type 3: K = 239). The brighter and larger the circle the better the performance of the algorithm. Source data are provided as a Source Data file.