Fig. 5

Panel (A) shows the ROC curves for predicting pCR status using the V3 total tumor cellularity (TTC). The red curve is from the TTC measured from the V3 data. The blue, green, and teal curves are from the V3 TTC predicted from calibrations using V1 and V2 data with a global, ADC + MSI habitat-informed, or local proliferation rate, respectively. We obtained AUC (95% CI) values of of 0.77 (0.67–0.86), 0.72 (0.61–0.82), 0.79 (0.68–0.87), and 0.77 (0.65–0.85), respectively. Using DeLong’s test, we do not find any significant difference between these AUC values. Panel (B) shows the ROC curves for predicting pCR status using the V3 total tumor volume (TTV). The red curve is from the TTV measured from the V3 data. The blue, green, and teal curves are from the V3 TTV predicted from calibrations using V1 and V2 data with a global, ADC + MSI habitat-informed, or local proliferation rate, respectively. We obtained AUC values of 0.77 (0.67–0.85), 0.72 (0.62–0.82), 0.79 (0.67–0.87), and 0.76 (0.66–0.84), respectively. Using DeLong’s test, we do not find any significant difference between these AUC values. Overall, we do not find that the AUC values from the predictions using any of the model calibration options (i.e., global, habitat-informed, or local proliferation rates) are significantly different, and we do not find that these AUC values are significantly lower than the AUC values from the measured data. Thus, we can predict pCR from V1-V2 calibrations without significantly reducing AUC from pCR predictions using measured V3 data. In particular, this means calibrating a biology-based model to a limited number of habitats allows for accurate and interpretable predictions of patient response to NAT.