High PD-L1 expression (≥50%) is a routine biomarker but is incompletely predictive, with response rates to PD-1 monotherapy only 35-45% in patients with lung cancer. Beyond PD-L1, additional individual pre-treatment variables, including clinical (smoking history, BMI), genomic (TMB, STK11, EGFR), and laboratory features (baseline dNLR), individually associate with response but have not been comprehensively examined in combination. We hypothesized that a multifactorial model incorporating routinely available clinical, pathologic, and genomic variables could improve prediction of response in high PD-L1 patients receiving first line anti-PD-(L)1 monotherapy.
190 patients from MSKCC with advanced, PD-L1 high NSCLC (PD-L1 ≥50%) treated with PD-1 or PD-L1 inhibitor were identified and separated into training (n=134, 70%) and validation cohorts (n=56, 30%). In addition to PD-L1 expression, 39 variables were collected, including histology, clinical (age, gender, performance status, smoking, clinical trial vs standard of care treatment), molecular (TMB, EGFR, KRAS, STK11, KEAP1, TP53, ALK, ROS1, BRAF), and baseline CBC (including dNLR). Radiologic response assessments were performed according to RECIST 1.1. To distinguish responders vs. non-responders, a logistic regression classifier with an elastic net penalty was used to restrict the number of variables considered and to optimize generalizability to independent cohorts. The parameters of the model were optimized using only the training cohort and its performance was measured on the validation cohort.
In PD-L1 high NSCLC patients treated with PD-(L)1 blockade, the ORR was 43%. In the training cohort, 5 features (PD-L1 expression, current smoking status, lymphocyte count, platelets, total WBC) associated with response. Three features (EGFR mutation, STK11 mutation, standard of care treatment) associated with lack of response. TMB was not predictive within this selected PD-L1 high cohort. In the training cohort, the eight identified features were used to develop a multifactorial model which improved BOR prediction (AUC 0.83) compared to PD-L1 alone (AUC 0.65), p=0.02. Improved performance of the model was confirmed in the validation cohort (AUC 0.66 for multifactorial model vs. AUC 0.52 for PD-L1 alone).
Among patients with high PD-L1 expression, multiple clinical, molecular, and baseline laboratory features impact response to PD-(L)1 monotherapy. The addition of these routinely available variables to PD-L1 in a multifactorial model improves prediction of response to PD-(L)1 blockade in patients with high PD-L1. This approach may help further stratify patients within the PD-L1 high population and identify which patients are likely to benefit from PD-(L)1 monotherapy vs those who should consider chemotherapy + immunotherapy.
Non-Small Cell Lung Cancer, PD-1 inhibitor, High PD-L1 expression
© 2019 Published by Elsevier Inc.