Abstract
Patients with metastatic triple-negative breast cancer (TNBC) show variable responses to PD-1 inhibition. Efficient patient selection by predictive biomarkers would be desirable but is hindered by the limited performance of existing biomarkers. Here, we leveraged in silico patient cohorts generated using a quantitative systems pharmacology model of metastatic TNBC, informed by transcriptomic and clinical data, to explore potential ways to improve patient selection. We evaluated and quantified the performance of 90 biomarker candidates, including various cellular and molecular species, at different cutoffs by a cutoff-based biomarker testing algorithm combined with machine learning-based feature selection. Combinations of pretreatment biomarkers improved the specificity compared to single biomarkers at the cost of reduced sensitivity. On the other hand, early on-treatment biomarkers, such as the relative change in tumor diameter from baseline measured at two weeks after treatment initiation, achieved remarkably higher sensitivity and specificity. Further, blood-based biomarkers had a comparable ability to tumor- or lymph node-based biomarkers in identifying a subset of responders, potentially suggesting a less invasive way for patient selection.