Predicting response to naratuximab emtansine, an anti-CD37 antibody-drug conjugate (ADC), in combination with rituximab in diffuse large B-cell lymphoma (DLBCL), by analyzing the spatial arrangement of CD37 and CD20 positive cells using deep learning.

Abstract

e19565 Background: DLBCL is the most common type of Non-Hodgkin’s lymphoma, accounting for 30-40% of cases. Despite improvements in response and survival with standard of care treatment, up to 40% of patients have relapsed and/or refractory (R/R) disease. A phase 2 study (NCT02564744) evaluated the efficacy of naratuximab emtansine, an anti-CD37 ADC, in combination with rituximab, in 80 patients with R/R DLBCL. We performed an exploratory analysis of the study to find pathology-based biomarkers predictive of response. Deep learning (DL) models were used to extract spatial features from digitized slides stained with CD37 and CD20 and their predictive role was evaluated. Methods: A cohort of 47 DLBCL patients from the study were selected based on availability of CD20/CD37 IHC staining and were used for analysis. Patient characteristics of the analyzed cohort were similar to those of the full study cohort. Overall response rate (ORR) of the analyzed cohort was 44.7%, similar to the ORR of the full study cohort. For each patient, whole slide images (WSI) of two slides from a pre-treatment biopsy, one stained for CD20 and one for CD37, were analyzed. DL models were used to classify cells as positive or negative to the two markers and CD20/CD37 co-expression was analyzed. Over 140 spatial features were pre-defined based on drug mechanism of action and biological hypotheses, and were calculated for each patient using cell classifications. A repeated 5-fold cross-validation analysis was performed to identify features predictive of objective response. Results: Two spatial features related to the proximity of CD37 and CD20 positive cells, demonstrated a significant correlation with clinical outcome. Each feature identified patients in the cohort as having either a positive or a negative response. Feature performance was estimated using 2000 repetitions of 5-fold cross-validation. One feature classified on average 21% of patients as responders. Patients positive for this feature had an average ORR of 78% (95% CI 64%-82%), an increase of 34% in ORR relative to the original cohort (p < 0.05). The second feature classified on average 35% of patients as positive. Patients positive for this feature had an average ORR of 67% (95% CI 62%-71%), an increase of 23% in ORR relative to the original cohort (p < 0.05). Conclusions: DL analysis of the co-expression and spatial arrangement of CD37 and CD20 in pre-treatment biopsies of DLBCL patients could potentially be used as a predictive biomarker for a response to a combination treatment of anti-CD37 and anti-CD20 drugs in DLBCL, and may improve patient stratification for further clinical trials.