Figure 2

Bacillus subtilis. (A) Comparison of predicted gene expression using OutPredict (grey dots) versus actual expression (red line) at the left-out time point. Genes are ordered by increasing actual mean expression value (red line). OutPredict predicts gene expression well at all expression levels. The accuracy of forecasting is measured by calculating the Mean Squared Error (MSE). (B) The vertical axis indicates MSE, where lower bars indicate more accurate predictions. The descriptions of the different models of the x axis can be found in Table 2. OutPredict (OP-Priors) performs significantly better (P < 0.05, based on a non-parametric paired test) than Penultimate Value (with a 30% relative improvement), DynGenie3 (with a 50% relative improvement) and Neural Network(NN). The MSE for Neural Nets is 3.75 (with standard deviation ≈0.3), which is considerably higher than for other methods (Supplementary Table S3); it is not shown here because the MSE is out of scale. Moreover, when priors from both Integrated steady-state data and prior gold standard data, are used with the OutPredict algorithm, there is a significant (P < 0.05, non-parametric paired test) improvement in predictions relative to OutPredict using only time series data. Specifically, prior gold standard data is significantly helpful, showing a 11% relative improvement (Supplementary Fig. S4). Finally, out-of-bag analysis concludes that the Time-step differencing model is better than the ODE-log.