1135 - Predictive modeling of treatment outcomes in chronic lymphocytic leukemia based on functional profiles at baseline

Background: Although targeted therapies have revolutionized the management of chronic lymphocytic leukemia (CLL), treatment efficacy varies from patient to patient and no treatment is given with curative intent. To avoid short treatment durations or toxicities of therapy, there is a need to identify biomarkers that can guide optimal treatment decisions for the individual patient.

Aim: To develop computational models that predict treatment outcomes in CLL patients based on functional analyses performed on primary CLL cells sampled before treatment start.

Methods: Peripheral blood mononuclear cells (PBMCs) were collected at baseline from relapsed/refractory CLL patients enrolled in three phase 2 clinical trials [NCT03226301 (ibrutinib + venetoclax cohort; n=186), NCT02742090 (umbralisib cohort; n=55), and NCT04624633 (umbralisib + acalabrutinib cohort; n=12)]. Multi-color flow cytometry with fluorescent cell barcoding was applied to the PBMCs for (phospho)protein profiling (31 proteins) of resting CLL cells.

Two computational models were developed to predict treatment outcomes on the clinical trials. Model 1 used (phospho)protein profiles from the ibrutinib + venetoclax cohort to predict MRD level in peripheral blood at cycle 15. The cohort was divided into a training set (n=139; 75%) and a test set (n=47; 25%), stratified based on MRD level, IGHV mutational status, and TP53 aberration status. After 15 cycles of ibrutinib + venetoclax treatment, 63, 44, and 7 patients in the training set had obtained an undetectable ( < 10-4, i.e. less than 1 CLL cell detected in 10000 leukocytes by flow cytometry), intermediate (10-4 to 10-2), or high MRD (>10-2) status, respectively. MRD sample or (phospho)protein profile was missing for the remaining 25 patients.

Model 2 used (phospho)protein profiles from the umbralisib cohort (training set) to predict tumor objective response rate (ORR). The umbralisib + acalabrutinib cohort was used as an independent test set. Patients who achieved partial response (PR) or complete response (CR) to umbralisib were classified as responders (n=15). Patients who achieved stable disease (SD) or progressive disease (PD) were classified as non-responders (n=40).

Statistical associations between the (phospho)protein profiles and treatment outcomes were assessed with Wilcoxon non-parametric test.

Results: Patients with high MRD after 15 cycles of ibrutinib + venetoclax therapy showed significantly higher (phospho)protein levels than patients with intermediate or undetectable MRD for 8 proteins in the training set (p < 0.05), including Bim, ERK1/2 (pT202/Y204), and STAT3 (pY705). There were no significant differences in (phospho)protein profiles between intermediate and undetectable MRD groups.

In the umbralisib cohort, the (phospho)protein levels were significantly higher in CLL cells from responders than from non-responders for 8 proteins (p < 0.05), including MEK1 (pS298), mTOR (pS2448), and p90 RSK (pS380).

The (phospho)protein profiles from the two training sets were used as input data to construct two independent LASSO logistic models with leave-one-out cross validation to predict patients with either high MRD (ibrutinib + venetoclax cohort; Model 1) or objective response (umbralisib cohort; Model 2). In the test sets, we obtained ROC-AUC (Receiver Operating Characteristic/Area Under the Curve) scores of 0.70 and 0.82 for Model 1 and Model 2, respectively. The proportion of the positive (responders) class in the umbralisib cohort (15/55; 0.27) was used as cut-off for the probability to predict a responder. The model correctly predicted all patients in the test set (umbralisib + acalabrutinib cohort) to be responders (median probability of 0.52, range=0.33-0.86), suggesting that it is feasible to predict treatment outcomes in CLL based on functional profiles at baseline.

Summary/Outlook: We present a novel approach to model treatment outcomes for CLL patients based on functional analyses and show that baseline (phospho)protein profiles of CLL cells have predictive value in independent CLL cohorts treated with targeted therapies. To further increase the robustness of the models, we will include additional types of functional data (immune phenotyping and drug sensitivity screening data), as well as clinical data. These analyses will be presented. A parallel strategy will be to expand the training set by combining input data from various patient cohorts. The goal is to develop a robust model and apply it in a biomarker-driven clinical trial to guide individual treatment selection.