Fig. 6

Inference of GRNs with CaiNet using a recurrent network training approach. a Demonstration of the inference procedure using a GRN comprising an input element and 2 genes. Left panel: sketch of the ground truth network. The equilibrium constants of transcription factor binding and gene product levels resulting from two different input levels (I and II) are depicted. Middle panel: sketch of the first training cycle. Based on the difference between a guessed network and the ground truth network and the gradient of the gene response functions, CaiNet iteratively adapts the parameters of the network until the expression levels of the guessed network match the ground truth values. Right panel: sketch of the trained network. Trajectories of the gene product levels of gene 1 (blue) and gene 2 (red) for input level I (b) and input level II (c). d Trajectories of equilibrium constants of transcription factor binding. e Sketch of the procedure to evaluate the inference approach. A ground truth network is generated by randomly omitting gene connections of a fully connected network and assigning equilibrium constants of transcription factors. Gene product levels simulated for the ground truth network are used to train a fully connected network. f Percentage of true positive gene connections (agreement between trained and ground truth network) versus percentage of false positive gene connections (exist in trained but not in ground truth network) for GRNs of 5 to 10 genes. Error bars denote s.d. of the training results of 12 different randomly generated networks. Inset: Histogram of relative errors (normalized difference in equilibrium constants between trained and ground truth network)