Fig 1: Active learning workflow with synthetic data.(A, B) Effect of the number of conditions added per round using different acquisition functions. Results are averaged across ten models per round; error bars represent standard deviation. (A) Hit discovery per round. From the total number of hits, colors indicate how many were correct versus incorrect. (B) Model generalization, represented by the Pearson correlation coefficient R2, after the final round (round 5). Mean R2 values are given, including error bars representing standard deviation across ten models. (C–E) In-depth characterization of models trained with four conditions added per round. (C) Comparison of mean CXCL9 activity and confidence interval width across acquisition functions. Each data point represents a condition from the selection pool in the corresponding round. (D) Comparison of model simulations with true CXCL9 values from the synthetic data set. Each data point represents a condition added to the training set in that round. Dashed lines indicate hit boundaries, dividing the space into regions for true/false hits and non-hits. (E) Inhibitor combinations of added conditions during the first three rounds. (F–H) Effect of the initial training set on (F) hit discovery (mean and SD across n = 10 models), (G) correct hit discovery (mean and SD), and (H) model generalization (mean and SD, n = 10 models). Round 1 is compared with round 5 to illustrate how the starting set influences outcomes. (I) Frequency of selection for each condition characterized by one or two inhibitors across five different initial training sets (Manual, All, Random1, Random2, Random3). Source data are available online for this figure.
Fig 2: Prior knowledge network (PKN) and perturbation screenings.(A) Overview of input preparation for CXCL9 logic-ODE model. Prior knowledge was used to set up the perturbation screening and prior knowledge network, also giving an idea of what possible effects could be of inhibitors (red arrow = inhibitory, green arrow = stimulatory, question mark = unknown). (B) Measured CXCL9 levels from wet lab screenings for all screened conditions, reported as Mean Fluorescent Intensity (MFI), averaged across two independent biological replicates with technical duplicates. (C) Statistical analysis of IFNγ and TNFα synergy. A Wilcoxon rank-sum test compared measured values of the IFNγ + TNFα condition from the wet lab with additive values, calculated as the sum of CXCL9 responses from IFNγ and TNFα only conditions (result either not significant (p > 0.05) or significant (p < 0.05 and p > 0.01)). Each data point for the measured data represents a replicate (two technical replicates for both biological replicates), while the line representing the additive result is the average from summing IFNγ or TNFα only conditions. Source data are available online for this figure.
Fig 3: Integration of wet lab screenings in active learning workflow.(A) CXCL9 activities measured in AsPC1 and BxPC3 wet lab screenings for four conditions selected by different acquisition functions. Each data point represents the mean value obtained from duplicate measurements. For each set of conditions per cell line and acquisition function, box plots show the median (center line), interquartile range (IQR, box bounds; 25th–75th percentiles) and the 1.5 x IQR (whiskers). Individual data points are overlaid on each box plot, also illustrating the minima and maxima. (B) Cumulative distribution functions of CXCL9 activity for each cell line and acquisition function. An early rise indicates that selected conditions yielded higher CXCL9 activity values. Greater distances between curves reflect larger differences in measured activity. (C) Confidence interval widths from models (n = 10) trained on conditions sampled in rounds 1 and 2. Violin plots show the distribution of widths from model simulations across all conditions in the selection pool. Embedded box plots indicate the median (center line), interquartile range (box bounds; 25th–75th percentiles), and whiskers extending to 1.5 x the interquartile range (IQR). (D) Frequency of inhibitor selection, either as single agents or in combinations, across rounds and acquisition functions. A value of four indicates that the inhibitor was chosen by all acquisition strategies. Source data are available online for this figure.
Fig 4: Parameter analysis of AsPC1- and BxPC3-specific models.(A, B) Effects of in silico single-edge knockouts on CXCL9 responses under IFNγ + TNFα stimulation. Significant effects (Wilcoxon rank-sum test across ten models) are highlighted separately per cell line, indicating whether CXCL9 increased or decreased. (C) Effect of in silico single-edge knockouts on IFNγ and TNFα synergy. Following each knockout, a significance test (Wilcoxon rank-sum test across ten models) was performed to determine whether synergy was retained (T) or lost (F). In the original (non-knockout) models, synergy was present across all visualized conditions (no inhibitor, RASi, or MEKi). (D) Effect of in silico knockouts of different sets of edges (batches) on IFNγ and TNFα synergy. Synergy was assessed after the removal of each batch. (E) Parameter comparison between AsPC1 and BxPC3 models using bootstrapped parameter values. Significant differences are highlighted in the network. Source data are available online for this figure.
Supplier Page from BioLegend for LEGENDplex™ Human CXCL9 (MIG) Capture Bead B3, 13X