Background: Immunotherapy (IO) is expected to become a part of standard neoadjuvant therapy for early-stage, triple-negative breast cancer. However, only a minority of patients benefit from the addition of IO to standard neoadjuvant chemotherapy (NACT). Biomarkers of IO response in the metastatic setting, such as PD-L1 expression, fail to predict benefit from IO in the early-stage setting. Given the financial cost of IO and the potential for irreversible, immune-related toxicities, predictive biomarkers are desperately needed. SimBioSys TumorScope (TS), a commercially available biophysical model, has been previously validated as a highly accurate predictor of response to standard NAT regimens. TS has the potential to identify patients treated with IO who would have responded well to standard NACT, and those who truly benefited from IO, enabling biomarker discovery.
Methods: We identified 17 pts who received pembrolizumab in combination with standard NACT. From these cases, dynamic contrast-enhanced MRIs were assessed from multiple timepoints throughout treatment. Using a proprietary convolutional neural net, the tumor region from each MRI was segmented to assess volume over time. We compared actual tumor volumes to predicted response to standard NACT (without pembrolizumab) generated with the TS platform. By contrasting the tumor volume across the simulation and the longitudinal MRI timepoints, a metric of response directly attributable to pembrolizumab (“IO benefit”) was generated. Clinical, radiological, metabolic, tumor morphology, and microvasculature features were extracted from TS simulations, and after quality control and removal of redundant metrics, used in linear models
as predictors of IO benefit.
Results: Nine clinical, seven metabolic, 16 microvasculature, ten tumor morphology, and 77 radiological features were used as individual predictors of IO benefit in linear models. Eight out of 17 cases had positive IO benefit values. In all, 12 out of the 119 metrics of interest nominally significant predictors of IO benefit (p<0.05). Top relationships of interest included: tumor convex hull surface area to volume ratio (b=-51.4, p=0.010), maximum 2D diameter column (a metric of longest tumor dimension) (b=1.8, p=0.011), and percentage of tumor cells with low oxygen concentration at simulation outset (hypoxic fraction) (b=1.0, p=0.034). In further analyses, we observed nominally significant relationships between IO benefit and hypoxic fraction measures taken across simulation timepoints. Significant relationships were salient in hypoxic fraction measures taken at early simulation timepoints, including weeks three (p=0.017), two (p=0.024), and five (p=0.038). Importantly, between IO benefit/hypoxic fraction relationships, we observed concordant positive directions of effect (i.e., higher hypoxic cancer fraction, higher IO benefit).
Conclusion: IO is likely to soon gain regulatory approval for early-stage TNBC, but biomarkers to predict benefit are lacking. PD-L1 expression has previously been correlated with hypoxic tumor environments and high density of tumor infiltrating lymphocytes. Even within a limited dataset, TS provided a biologically explainable marker to identify patients who may benefit from IO. Work is underway to further validate these findings in a larger dataset, and to validate hypoxic tumor fraction as a predictor of IO response.
