(M-076) Methodology of the Exposure-Response (E-R) Analysis of Linvoseltamab in Patients with Relapsed/Refractory (RR) Multiple Myeloma (MM)

Jason Chittenden, PhD – Sr Dir Group Quantitative Pharmacology, Research & Development Sciences, Regeneron Pharmaceuticals, Inc.; Lutz Harnisch, MD, PhD – Exec Dir Quantitative Pharmacology, Research & Development Sciences, Regeneron Pharmaceuticals, Inc.; Anasuya Hazra, PhD – Sr Dir Group Clinical Pharmacology, Research & Development Sciences, Regeneron Pharmaceuticals, Inc.; Joannellyn Chiu, PhD – Assoc Dir Quantitative Pharmacologist, Research & Development Sciences, Regeneron Pharmaceuticals, Inc.

Linvoseltamab, a BCMA×CD3 bispecific antibody, induced deep and durable responses with generally manageable safety in patients (pts) with heavily pretreated RRMM in the phase 1/2 LINKER-MM1 study (NCT03761108) [1]. Treatment success may be heterogeneously distributed in a study based on inherent individual differences within the pt population and drug exposure, which are manifested as covariates in a clinical data set.

Aim: develop methods to streamline E-R analyses for time-to-event (TTE) and categorical analyses through an automated screening process to determine clinically relevant covariates and exposure metrics that would have the most significant impact on efficacy/safety endpoints.

Implemented in the R language, a series of conditional loops were assembled to screen clinically relevant covariates (including linvoseltamab exposure metrics derived from a population pharmacokinetic model, markers of disease burden, prior therapy, pt demographics, etc. from LINKER-MM1) to discern those with the most significant impact. For categorical analyses, binary endpoints were evaluated against each exposure metric by univariate logistic regression. The best exposure metric (by Akaike Information Criterion) for each endpoint was expanded by automated stepwise covariate analysis with pruning iterations to identify a multivariate logistic model. For TTE analyses, initial univariate analyses, using statistical and non-statistical tests, were followed by multivariate filtering to refine the covariate pool. Further assessments of multicollinearity and correlations, alongside rigorous model validation using several fit diagnostics, led to derivation of interquartile hazard ratios of the most significant covariates.

The developed methods efficiently filtered through a dataset from 282 pts with ≤110 clinically relevant covariates per pt from LINKER-MM1 and distilled them into ≤7 statistically significant covariates per endpoint. The total runtime per endpoint for categorical and TTE analyses was under several minutes without parallel processing. TTE and categorical analyses showed strong overlap in significant covariates, which aside from drug exposure were supported by scientific literature; pts with lower baseline tumor/disease burden (plasmacytomas, sBCMA, serum M protein, and lactate dehydrogenase) and higher erythrocyte concentrations were more likely to respond to therapy and had longer progression-free survival and overall survival. Safety analyses (categorical and TTE) for first infection and grade ≥3 neutropenia showed a lack of positive correlation between linvoseltamab exposure and hazard for these safety endpoints.

The analysis concluded with a set of covariates that confirmed dose-related improvements in efficacy and safety, and illuminated the influence of disease burden markers on outcomes. This approach offers a data-driven, mechanistic insight into RRMM, paving the way for a mechanism-based model centered around key covariates.

Citations: [1] Jagannath S, et al. AACR 2024. Presentation CT001