Skip to main content

Table 1 Performance of ARTIVA on simulated data

From: Statistical inference of the time-varying structure of gene-regulation networks

   WT simulations KO Simulations
Parameter Changepoint sensitivity Changepoint PPV Edges sensitivity Edges PPV Changepoint sensitivity Changepoint PPV Edges sensitivity Edges PPV
  0.2 94.2 95.1 73.9 99.2 100 100 100 98.5
  0.4 90.8 94.1 79.3 97.9 100 99 100 98.2
  0.6 87.8 92.5 73.9 96.9 96 97.1 99.2 95.4
  0.8 75.4 96.3 78.8 96.4 81.4 97.8 97.6 91.4
Noise 1 80.5 96.6 74.7 97.5 69.1 95 95.1 88.8
  1.2 71.7 96.2 78.4 97.6 28.1 82.4 97.2 87.6
  1.4 58.7 94.6 79 95.9 23.6 89.7 92.7 87.5
  1.6 52.9 91.8 74.4 97.2 10.8 75.9 79.1 86.8
  1.8 60.8 94.5 76 95.6 4.8 81.8 76.8 85.6
  1 78.5 97.5 1.3 100 79 98.8 18.1 82.4
  2 92 92 24.7 98.5 97 96.5 98.7 92.5
Phase size 3 90 94.2 50.8 98.6 99.5 99 100 93.6
  4 94.5 94 74.5 98.6 99.5 99.5 100 96.4
  5 96 99 76.9 96.8 99.5 97.5 100 94.7
  12 100 99 92.6 98 99 97.1 100 97.8
  5 93.7 95.5 82.4 99.1 _ _ _ _
# of parent 10 81.1 88.3 69.7 96.4 _ _ _ _
genes 20 62.4 83.4 51.2 97.3 _ _ _ _
  40 54 77.2 33.1 96.4 _ _ _ _
  1. To evaluate ARTIVA performances, two types of data were used: WT simulations corresponding to time-series expression data with no knowledge of potential transcription factors and KO simulations corresponding to several time-series expression data of a simulated wild-type strain and knock-out strains for each known transcription factor (see Methods and Additional file 2 for more details). The default values of the parameters used for the simulation study are: # of timepoints n = 12; # of changepoints k ~ U ( { 1 , .. , n / 4 } ) ; maximal # of edges = 5; parent to target coefficient ~ U ( [ 2 , 0.1 ] [ 0.1 , 2 ] ) ; phase sizes ~ U ( { 3 , .. , n } ) ; # # of replicates r = 8 for WT simulations, r = 4 for KO simulations; noise standard deviation ~ N ( 0 , 0.5 ) ; total # of simulations = 200 for each condition. In this table, the value of noise intensity, phase size and number of parent genes change according to the parameter under study (all other parameters were set to default). In each condition, the ability of ARTIVA to detect all true phase changepoints and model edges (Sensitivity) and to detect only true positives (Positive Predictive Value, PPV) was calculated. Overall, this simulation study allows us to gain confidence in ARTIVA results ( 80% of PPV and sensitivity) for a given set of parameters (noise standard error is on the order of the mean value of the regression coefficients, number of measurements in a regulatory phase > 8 and less than 20 parent genes).