A functional prognostic model predicts progression‐free survival in patients with relapsed chronic lymphocytic leukemia treated with ibrutinib + venetoclax Authors: Johanne U Hermansen, Weikaixin Kong, Andrea M Brodersen, Yanping Yin, Aleksandra Urban, Idun D Rein, Liye He, Rudi Agius, Rebecca S Teglgaard, Juho Rousu, Christian Brieghel, Sabina Kersting, Mark‐David Levin, Hoa T T Tran, Mattias Mattsson, Juha Ranti, Gerrit‐Jan Veldhuis, Caspar da Cunha‐Bang, Rogier Mous, Julie Dubois, Arnon P Kater, Jorrit M Enserink, Carsten U Niemann, Tero Aittokallio, Sigrid S Skånland Published: 2026 ABSTRACT AbstractTargeted therapies have improved outcomes for patients with chronic lymphocytic leukemia (CLL); yet, not all achieve durable responses. Existing prognostic models, developed for chemoimmunotherapy, have limited predictive value for newer targeted combinations. Here, we used data from a randomized phase II clinical trial to identify predictors of progression‐free survival (PFS) in response to Bruton tyrosine kinase (BTK) + B‐cell lymphoma 2 (Bcl‐2) inhibitor combination therapy. We studied the functional, genetic, and clinical characteristics of patients with relapsed/refractory (R/R) CLL treated with ibrutinib + venetoclax in the VISION/HO141 clinical trial (NCT03226301). Ex vivo drug sensitivity testing, (phospho)protein profiling, and immunophenotyping were performed on baseline peripheral blood mononuclear cells from 177 patients. Ex vivo drug sensitivity correlated with clinical responses to single‐agent therapies. (Phospho)protein patterns reflected intracellular pathway activity and predicted ex vivo drug sensitivity. Using stringent feature selection and regularized ridge regression to prevent overfitting, we integrated functional, genetic, and clinical data sets and identified six features—sensitivity to VEGFR‐2‐, JAK1/2 + Bcl‐2‐, and BTK + PI3K‐inhibitors; phosphorylation ratios of Bcl‐2:BTK and p90RSK:PLCγ2; and frequency of CD8+ T cells—that together outperformed the International Prognostic Index for CLL (CLL‐IPI) and other clinical scores in predicting PFS. These findings demonstrate that a prognostic model based on the functional features of CLL cell signaling and interaction with the microenvironment has the potential to improve upon conventional prognostic scores, warranting further clinical validation also in the frontline setting. INTRODUCTION Chronic lymphocytic leukemia (CLL) is a B‐cell malignancy characterized by heterogeneity in both clinicobiological features and clinical outcomes. 1 Prognostic markers include mutational status of the immunoglobulin variable heavy chain gene (IGHV) and TP53 aberrations. These genetic markers were defined in the era of chemoimmunotherapy (CIT), with unmutated IGHV and TP53 aberrations being associated with a poor prognosis. Targeted therapies were approved for CLL a decade ago as an alternative to CIT and included inhibitors of phosphatidylinositol 3‐kinase (PI3K), Bruton tyrosine kinase (BTK), and B‐cell lymphoma‐2 (Bcl‐2). These proteins are key regulators of cell proliferation and survival, and the novel drugs were proven to be very effective. Next‐generation inhibitors have since been developed and approved, and combinations of targeted therapies are under investigation. 2 The combination of the BTK‐ and Bcl‐2‐inhibitors ibrutinib + venetoclax was approved as frontline treatment of CLL by the European Medicines Agency in 2022, as the first all‐oral, fixed‐duration regimen. The approval was based on the phase III GLOW study, 3 and the fixed‐duration cohort of the phase II CAPTIVATE study that showed deep and durable responses to ibrutinib + venetoclax, also in patients with poor prognostic markers. 4 This observation suggested that traditional prognostic markers have lost predictive power in the era of fixed‐duration combination targeted therapies. This was confirmed in the phase II VISION/HO141 trial, which evaluates the efficacy of minimal residual disease (MRD)‐guided combination treatment with ibrutinib + venetoclax in relapsed/refractory (R/R) CLL. 5 , 6 In agreement with these observations, a recent evaluation of the CLL international prognostic index (CLL‐IPI) demonstrated that its impact in predicting survival with targeted therapies is diminished compared to CIT, and that targeted therapies are effective across established risk groups. 7 However, since response to treatment remains heterogeneous also in the era of targeted therapies, novel prognostic markers are needed to optimally tailor treatments. Here, we aimed to investigate the predictive power of genetic, functional, and clinical features on treatment outcomes to fixed‐duration ibrutinib + venetoclax combination therapy in the VISION/HO141 clinical trial. We performed functional profiling by ex vivo drug sensitivity screening, immunophenotyping, and (phospho)protein profiling on CLL cells collected before the start of treatment from 177 enrolled patients with R/R CLL. We established statistical associations between functional features and traditional prognostic markers to understand the phenotypic landscape of low‐ and high‐risk disease, and then used both functional and genetic features to predict patient‐specific PFS. Our multi‐modal data‐driven machine learning model selected a combination of six functional features as most predictive of PFS, and the resulting functional prognostic model (fPM) demonstrated an improved and robust performance compared to CLL‐IPI and other clinical scores, 8 , 9 , 10 , 11 indicating its potential to guide personalized management strategies and improve conventional prognostic scores. METHODS Reagents Drug sensitivity testing was performed with a custom‐made drug library (Tables S1–S2). Drugs were either approved for CLL, under clinical investigation, reported to show ex vivo efficacy in CLL, or derivatives of these. 12 The antibodies and reagents for flow cytometry are listed in Supporting Information S4: Table S3. Patient material and ethical considerations Peripheral blood mononuclear cells (PBMCs) were collected and cryopreserved at baseline from 177 patients with R/R CLL enrolled in the VISION/HO141 clinical trial (NCT03226301, Table 1). All participants signed a written informed consent prior to sample collection. Research on patient material was carried out in agreement with the Declaration of Helsinki. The study was approved by the Regional Committee for Medical and Health Research Ethics of South‐East Norway (28507). Table 1Patient characteristics.Characteristic n = 177Median age (IQR)69 (62–72)Median progression‐free survival (95% confidence interval)2.95 (2.76–2.97)Median β2‐microglobulin (mg/L) (IQR)5 (3.8–6.8)Sex—no. (%)Male121 (68.3)Female56 (31.7)Binet classification—no. (%)Stage A26 (14.7)Stage B73 (41.2)Stage C76 (43.0)Data not available2 (1.1)Cytogenetic subgroup—no. (%)11q deletion55 (31.0)Trisomy 1220 (11.2)13q14 deletion120 (67.8)17p13 deletion30 (16.9)No abnormalities18 (10.1)IGHV mutation status—no. (%)Mutated56 (31.6)Unmutated112 (63.2)Could not be evaluated9 (5.0) TP53 mutation status—no. (%)Mutated45 (25.4)Unmutated127 (71.7)Could not be evaluated5 (2.8)Minimal residual disease (C15)—no. (%)Low (<10−4)84 (47.5)Intermediate (≥10−4 and <10−2)62 (35.0)High (≥10−2)13 (7.3)Unknown18 (10.2) Patient characteristics. CellTiter‐Glo luminescent cell viability assay Drug sensitivity experiments were performed as previously described. 13 , 14 The drug library consisted of 95 single agents and 87 combinations (Tables S1–S2), which represented multiple target classes (Supporting Information S1: Figure S1a). Briefly, each single agent or combination was printed into four 384‐well cell culture microplates using the Echo 550 liquid handler (Labcyte Inc.). Each compound was tested at five different concentrations in 10‐fold increments ranging from 1 to 10,000 nM (0.1–1000 nM for copanlisib and dasatinib; 1 µM−10 mM for sodium salicylate). Combinations were designed using the fixed molar concentration series identical to those used for single agents, that is, the “diagonal” design. 14 PBMCs from CLL patients were first co‐cultured with irradiated (50 Gy/125 Gy/125 Gy, respectively) APRIL‐, BAFF‐, and APRIL + CD40L‐expressing fibroblasts (ratio 1:1:1) for 24 h to mimic the tumor microenvironment and to prevent induction of spontaneous apoptosis, as previously described. 15 The primary CLL cells were then separated from the adherent fibroblast layer by carefully resuspending the culturing medium and transferring it to a separate tube. A single‐cell suspension (5000−10,000 cells/well) was distributed to each well of the 384‐well assay plates using the CERTUS Flex liquid dispenser (Fritz Gyger). The cells were incubated with the compounds at 37°C for 72 h. Cell viability was measured using the CellTiter‐Glo luminescent cell viability assay (Promega) following the manufacturer's instructions. Luminescence was recorded using an EnVision 2102 Multilabel Reader (PerkinElmer). The response readout was normalized to the negative (0.1% DMSO) and positive (100 μM benzethonium chloride) controls. The measured concentration–response data were processed with KNIME software (KNIME AG). The quality of the experiments was monitored by calculating the Z‐prime for each drug plate. This measure describes the separation of the positive and negative controls on each plate. Experiments with Z‐prime values > 0.5 are described as excellent, values ranging from 0 to 0.5 are acceptable, and values below 0 indicate poor quality. 16 The Z‐prime was above the threshold in the majority of the experiments (Supporting Information S1: Figure S1b). To quantify the compound responses, a modified drug sensitivity score (DSS) was calculated for each sample and compound separately. 17 , 18 An area under the concentration–response curve was constructed using an activity window from 100% to 10%, and a window from the minimum concentration tested to the concentration where the viability reached 10%. The DSS3 metric was used, without the division by the logarithm of the upper asymptote of the logistic curve. The score ranges between 0 and 100, and a higher DSS indicates higher sensitivity to the treatment. For each patient sample, the DSSs were standardized relative to the mean DSS and standard deviation in that patient sample, reported as the Z‐scored DSS (Supporting Information S1: Figure S1c). The calculated DSS values quantify the relative cell viability (Figure S1d). Investigation of ex vivo drug sensitivities Drug sensitivity data were split into a single‐agent data set and a drug combination data set. Next, the standardized drug sensitivity data were further split into groups using K‐means clustering on the Z‐scored DSS. Preliminary observations of the data alluded to three groups of relative drug sensitivities; we therefore chose K = 3 for the clustering. These groups were visualized as a heatmap using the complete hierarchical clustering method and the Euclidean distance within each K‐means cluster separately. The clusters were later labeled with genetic subtypes, and we describe the three groups as high, intermediate, and low drug sensitivity (separated by a space in the heatmap). Immunophenotyping with (phospho)protein profiling and fluorescent cell barcoding Experiments were performed as previously described in detail. 19 , 20 , 21 Briefly, PBMCs from CLL patients were thawed, washed with flow wash (phosphate‐buffered saline [PBS] supplemented with 2% fetal bovine serum [FBS]), and stained with fixable viability stain. Further, the cells were fixed and washed with PBS supplemented with 0.02% saponin before staining with a fluorescent cell barcoding (FCB) matrix. The cells were then combined in one tube and stained with the indicated surface markers for immunophenotyping before permeabilization. The cells were stored in the permeabilization buffer at −80°C until further processing. After thawing, the cells were washed with flow wash and stained with antibodies against intracellular proteins. The samples were analyzed with a BD FACSymphony A5 cytometer (BD Biosciences) equipped with 355, 405, 651, and 640 nm lasers. The data were analyzed in Cytobank (https://cellmass.cytobank.org/cytobank/) as previously described. 19 , 20 , 21 Processing of clinical features Genetic abnormalities included 11q deletion, trisomy 12, 13q14 deletion, 17p13 deletion, TP53 mutation, and IGHV mutational status. 5 Clinical features included in the analyses were Binet stage, MRD in peripheral blood at 15 cycles, and PFS. Statistical association analysis of functional, genetic, and clinical data The clinical features of Binet stage, MRD, and PFS were binarized to perform the statistical analysis (Wilcoxon Rank sum test). Binet stage was split into groups “A&B” and “C”; MRD was split into “undetectable/low” (<10−4, i.e., from less than 1 CLL cell detected in 10,000 leukocytes by flow cytometry) and “intermediate/high” (≥10−4 to >10−2); and PFS was split into “predicted high risk” and “predicted low risk” based on the fPM model described in the section “Development of the functional prognostic model (fPM)”. Prediction of ex vivo drug sensitivities using a random forest model The Z‐scored DSS values were predicted for each drug or drug combination from binarized clinical features (see section “Statistical association analysis of functional, genetic, and clinical data”, above), genetic subtypes, and the Z‐scored (phospho)protein data using random forest regression. Missing values were handled by excluding patient samples with any missing data point from the analysis, as the vast majority of missing values occurred for binary features and were thus not suitable for imputation. The exclusion resulted in 131 patient samples with complete data sets. A random forest model was fitted for each drug or drug combination using the random forest package in R with the complete data set to generate 500 decision trees randomly using 13 out of 41 features sampled in each split. To assess model performance and detect potential overfitting, the difference between the predicted and measured values was assessed using root mean squared error (RMSE). RMSE was reported for each drug or drug combination from both an out of bag (OOB) estimate using the complete data set and a 10‐fold testing strategy. 22 , 23 In the 10‐fold cross validation, samples were randomly assigned to the training (90%) and test (10%) sets. As the outcome variable (Z‐scored DSS) was continuous, cross validation folds were not stratified. Given that OOB errors were slightly higher than the corresponding 10‐fold test errors, we considered the OOB estimates to be conservative and therefore relied on the OOB strategy using the complete patient cohort for downstream analyses of feature importance (Supporting Information S1: Figure S6c,d). Feature importance analysis The percent increase in the mean squared error (MSE) when leaving a feature out of the model (%IncMSE) was calculated as a measure of importance for each feature in the random forest models functioning as an OOB sampling error estimator. These values were extracted using the importance function from the DALEX package in R. To assess the significance of feature importance, a permutation test was conducted for each feature. A null distribution of %IncMSE values was generated by randomly permuting the target feature and refitting the random forest model. The p‐value for each feature was computed as the proportion of %IncMSE values in the permutated data that was greater than or equal to the original value, indicating that the importance of the features was not random (Supporting Information S1: Figure S7). SHapley Additive exPlanations (SHAP) analysis Shapley values were calculated to understand how the values of each feature affected the directionality and importance of each feature in predicting each ex vivo DSS using the SHAP library and the RandomForestRegressor from the sicikit‐learn library in Python. For the interpretability analysis using SHAP, the random forest model was fitted by randomly splitting the data into association training and test sets using an 80:20 ratio. This common split allows for a substantial portion of the data (80%) to be used for training the model to learn complex patterns and computing the SHAP values, while reserving a sufficient portion (20%) for an unbiased evaluation of how well the trained model generalizes to unseen data. The model generated 500 decision trees and the random forest data were visualized using SHAP summary plots. Feature importance data were collected and summarized, including the mean, standard deviation, maximum importance, and the corresponding drug associated with the maximum importance for each feature. Processing of functional data Protein profiling data were normalized by subtracting the mean fluorescence intensity (MFI) of the isotype control from the other MFI measurements for each patient sample, before log10‐transformation. For association analyses, the normalized protein profiling data and drug sensitivity data were standardized the same way across proteins or drugs for each patient to compute Z‐scores: where S is either log10 (MFI) of the protein profiling data or the DSS from the drug screening data 18 ; i is either a specific protein or DSS; μ pt is the mean protein expression or DSS per patient sample; and σ pt is the standard deviation per patient sample across the proteins or drugs. Statistical association analyses Statistical associations of drug sensitivity or (phospho)protein features with clinical features and genetic subtypes were investigated using the Wilcoxon rank sum test. P < 0.05 were considered significant. False discovery rate correction was performed (Tables S4–S5). Codes: https://github.com/Andreabrodersen/CLL-association-study. Development of the fPM The patient cohort (n = 177) was divided into two batches; the first batch (n = 157) was stratified into training (n = 80) and test (n = 77) sets based on IGHV and TP53 aberration status. The second batch, for which functional data were generated later, served as a blinded validation set (n = 20). The combined test and validation sets constituted a “combined test set”; these data were not used in training of the fPM model. Please refer to Supporting Information S11: Supplementary Methods for a detailed description of model development and validation. RESULTS Genetic sub‐groups of CLL present distinct ex vivo drug sensitivity patterns To understand phenotypes of CLL and their associations with prognostic markers, we performed multivariate analyses to identify correlations between clinical, genetic, and functional measurements from 177 patients with R/R CLL, before treatment with ibrutinib + venetoclax in the VISION/HO141 clinical trial. 6 The patients represented, as expected, a high‐risk population with 17% del(17p), 25% mutated TP53, and 63% IGHV unmutated CLL (Table 1). The setup for genetic, drug sensitivity, immune phenotyping, and (phospho)protein profiling is outlined in Figure 1. Ex vivo sensitivity to single‐agent treatments (Supporting Information S2: Table S1) showed significant differences between genetic sub‐groups of CLL (Supporting Information S5: Table S4). We confirmed previous findings that TP53 aberrations (17p13 deletion or TP53 mutation) affect drug sensitivity (Figure 2A). CLL with wild‐type TP53 presented with higher sensitivity to chemotherapeutic agents (2‐chlorodeoxyadenosine, bendamustine, chlorambucil, and etoposide), the p53‐MDM2 antagonist nutlin 3a, and the nuclear export receptor inhibitor selinexor, compared to TP53 aberrant CLL (Figure 2A). 24 Interestingly, CLL cells with TP53 aberrations showed significantly increased sensitivity to targeted agents, including the BTK inhibitor zanubrutinib and the PI3K inhibitors buparlisib, compound 7n, and idelalisib (Figure 2A). The study design. Schematic outline of the study. Peripheral blood mononuclear cells (PBMCs) collected from patients (n = 177) with relapsed/refractory chronic lymphocytic leukemia (CLL) enrolled in the clinical trial VISION/HO141 were subjected to genetic and functional measurements. Clinical information about the patients was available. All measurements were subjected to data integration and biomarker discovery studies using multivariate analyses and machine learning. Associations between genetic markers and ex vivo sensitivity to single agents and drug combinations in R/R CLL. (A) A Wilcoxon rank sum test was performed to investigate the relationship between changes in the Z‐scored drug sensitivity score (DSS) and assess the statistical significance (−log10(p)) for each single drug in response to the different genetic subtypes and binarized clinical features. The results were visualized using volcano plots, where each point represents a drug, with colors indicating significance levels (black for nonsignificant and blue for significant at a threshold of P < 0.05). The horizontal line indicates the significance threshold (P = 0.05) and the vertical line indicates the zero‐change line. Drug labels are provided for the most significant associations. Data points in the upper right quadrant indicate an association between a high DSS for the given drug and the presence of the indicated phenotype, whereas data points in the upper left quadrant indicate a lower DSS for that phenotype. (B) The analysis was performed as described in (A), but for drug combinations. The ex vivo sensitivity to the treatment administered in the trial, ibrutinib + venetoclax, is highlighted in dark gray. BTK inhibitors are labeled in green; Bcl‐2 inhibitors are labeled in burgundy. We next found that IGHV‐mutated CLL showed increased sensitivity to chemotherapeutic agents (Figure 2A), which is in agreement with clinical observations. 25 IGHV‐mutated CLL further showed increased sensitivity to venetoclax (Figure 2A). To study the impact of the two prognostic markers, IGHV and TP53 mutational status, on ex vivo sensitivity, we compared the profiles of CLL cells that harbored both aberrant TP53 and unmutated IGHV (n = 31) to CLL cells with TP53 aberration and mutated IGHV (n = 48). We found that these cohorts showed a similar drug sensitivity profile (Supporting Information S1: Figure S2a), suggesting that TP53 mutational status is dominant over IGHV mutational status regarding drug sensitivity. We next studied the impact of different karyotypic aberrations on drug sensitivity profiles. CLL cells with trisomy 12 or 13q14 deletion were less sensitive to the Bcl‐2 family inhibitors navitoclax or obatoclax, respectively, compared to CLL cells without these aberrations (Figure 2A). The single agents with overall highest potency, bortezomib, romidepsin, panobinostat, and ixazomib (Supporting Information S1: Figure S1c), did not induce significantly different responses in any of the genetic sub‐groups (Figure 2A), indicating that they may be effective across genetic sub‐groups. Together, these results demonstrate that genetic sub‐groups of CLL present with distinct ex vivo drug sensitivity profiles and that ex vivo drug sensitivities reflect clinical efficacy. Presence of poor prognostic markers is associated with increased sensitivity to BTK inhibitor combinations Since monotherapy is uncommon in CLL, we expanded our study to include drug combinations (Supporting Information S3: Table S2). We found that the CLL cells were most responsive to drug combinations that included a Bcl‐2 inhibitor, including six Bcl‐2 + BTK inhibitor combinations and 15 Bcl‐2 + PI3K inhibitor combinations (Supporting Information S1: Figure S3 and Tables S4–S5). Interestingly, there was an overall separation in sensitivity to Bcl‐2 inhibitor combinations and BTK inhibitor combinations between CLL cells with a favorable prognostic profile (low risk; TP53 wild type, mutated IGHV) and CLL cells with an unfavorable prognostic profile (high risk; TP53 aberrant, unmutated IGHV; Figure 2B and Supporting Information S6: Table S5). IGHV‐mutated CLL cells showed significantly increased sensitivity to Bcl‐2 inhibitor combinations (9/12 combinations) and reduced sensitivity to BTK inhibitor combinations (13/16 combinations) compared to cells with unmutated IGHV (Figure 2B). CLL cells with TP53 aberration showed significantly increased sensitivity to BTK inhibitor combinations (6/7 combinations for TP53‐mutated CLL and 3/4 combinations for 17p13 deleted CLL) compared to cells with wild‐type TP53 (Figure 2B). In agreement with the observed responses to monotherapies, CLL cells with the combined presence of aberrant TP53 and unmutated IGHV showed a similar sensitivity profile to CLL with aberrant TP53 and mutated IGHV (Figure 2B and Supporting Information S1: Figure S2b). It is noteworthy that the ex vivo sensitivity to ibrutinib + venetoclax, administered in the trial, was found to be statistically different only for CLL cells without 11q deletion (Figure 2B, marked in gray), and for patients with a predicted long PFS (see section “Construction and validation of a multi‐fPM for PFS prediction” and Supporting Information S1: Figure S2b). The lack of association with other prognostic markers, including IGHV and TP53 mutational status, aligns with the clinical report showing that treatment outcome did not correlate with predefined prognostic markers. 5 Trisomy 12 confers intermediate prognostic risk in CLL. 26 CLL cells with trisomy 12 showed increased sensitivity to BTK inhibitor combinations and reduced sensitivity to Bcl‐2 inhibitor combinations relative to CLL cells without this aberration (Figure 2B). Deletion of 13q14 is the most common cytogenetic aberration in CLL and is associated with a good prognosis if it is the only karyotype aberration. 27 CLL cells with 13q14 deletion showed increased sensitivity to the Bcl‐2 inhibitor combinations venetoclax + pilaralisib and venetoclax + pimasertinib (see Supporting Information S6: Table S5 for P‐values), in agreement with deletion in 13q14 being associated with increased expression of Bcl‐2, the target of venetoclax, 28 , 29 and confirming that low‐risk CLL may show increased sensitivity to Bcl‐2 inhibitor combinations (Figure 2B). Taken together, our findings demonstrate that low‐risk CLL is associated with increased ex vivo sensitivity to Bcl‐2 inhibitor combinations, while intermediate‐ and high‐risk CLL is associated with increased sensitivity to BTK inhibitor combinations, in agreement with clinical guidelines. 2 Low‐ and high‐risk CLL show significantly different protein expression and phosphorylation profiles Having established that genetic sub‐groups of CLL present with distinct drug sensitivity profiles, we next aimed to investigate whether these sub‐groups presented with different immunophenotypes or protein profiles, since such features may impact drug sensitivity. 30 , 31 To this end, we performed combined immunophenotyping and (phospho)protein profiling on the patient cells (Figure 1). The gating strategy is shown in Supporting Information S1: Figure S4, and a representative t‐SNE plot is shown in Supporting Information S1: Figure S5a. Among the immune cell populations that were analyzed (CD19+ CLL cells, CD3+ T cells, CD4+ T cells, CD8+ T cells, CD56+ NK cells, and CD56+CD16+ NK cells) and stratified for genetic sub‐groups of CLL (11q deletion, trisomy 12, 13q14 deletion, 17p13 deletion, TP53, and IGHV), we found few statistically significant differences (Supporting Information S1: Figure S5b‐g), suggesting that the immunophenotype is generally preserved between genetic sub‐groups of CLL. However, CD56+ NK cells showed a significantly higher proportion in CLL with 11q deletion compared to CLL without this aberration (Supporting Information S1: Figure S5b), and CD8+ T cells showed a significantly higher proportion in CLL with unmutated IGHV than CLL with mutated IGHV (Supporting Information S1: Figure S5f). The latter is of particular interest, since we identified a high proportion of CD8+ T cells as a poor predictive marker of PFS (see section “Development and validation of a functional prognostic model for prediction of clinical response to ibrutinib + venetoclax treatment” below). We next analyzed the expression or phosphorylation level of 31 proteins related to B‐cell signaling (Supporting Information S4: Table S3). For each patient sample, the normalized log10‐transformed MFI was standardized, resulting in a Z‐score (log10 MFI) (Supporting Information S1: Figure S6a–b). These Z‐scores (log10 MFI) were clustered. The Bcl‐2 family proteins Bcl‐2, Bcl‐xL, and Bim were the top three proteins with the highest relative expression across patient samples, followed by BTK (pY223) and S6‐ribosomal protein (pS235/236) (Figure 3A). We observed that the phosphorylation level of PLCγ2 (pY759), which is downstream of BTK, was significantly higher in cells from patients with genetic aberrations associated with a favorable disease (13q14 deletion, mutated IGHV) and lower in cells from patients with genetic aberrations associated with unfavorable disease (17p13 deletion, TP53 mutation, 11q deletion, and IGHV mutated and TP53 aberrant; Figure 3B and Supporting Information S1: Figure S2c). Additionally, we observed that cells from patients with poor prognostic markers 11q deletion or unmutated IGHV showed higher expression of several of the anti‐apoptotic Bcl‐2 family proteins Bcl‐2, Bcl‐2 (pS70), Bcl‐xL, and Mcl‐1, while patient cells with trisomy 12 or 13q14 deletion showed significantly lower expression of Bcl‐2, Bcl‐2 (pS70), and/or Mcl‐1 (Figure 3B). Interestingly, CLL cells from patients who showed undetectable or low MRD in response to the trial treatment also showed significantly higher levels of Mcl‐1, underscoring the lack of association between reported sensitivity markers and response to ibrutinib + venetoclax treatment (Supporting Information S1: Figure S2c). Associations between genetic markers and (phospho)protein profiles. (A) Peripheral blood mononuclear cells (PBMCs) from 177 chronic lymphocytic leukemia (CLL) patient samples were fixed, barcoded, and stained with antibodies against surface markers and intracellular (phospho)proteins and analyzed by flow cytometry. The results are shown for CD19+ B cells. The signals were clustered using complete hierarchical clustering and Euclidean distances. Information about genetic markers (see key) was added after clustering. (B) The volcano plot illustrates the relationship between changes in (phospho)protein signals [Z‐score log10(MFI)] and statistical significance (−log10(p)) for each protein in response to different genetic markers. Each point represents a protein, with colors indicating significance levels (black for nonsignificant and blue for significant at a threshold of P < 0.05). Dashed lines indicate the significance threshold (P = 0.05) and the zero‐change line. Proteins on the left side of the vertical line are more abundant in cells without cytogenetic changes (“not present”) or with unmutated TP53 or IGHV. Proteins on the right side are more abundant in cells with cytogenetic changes (“present”) or mutated TP53 or IGHV. Protein labels are provided for the most significant associations. Taken together, genetic subgroups of CLL display significantly different protein expression and phosphorylation profiles. Proteomic features predict ex vivo drug sensitivity Having shown that genetic sub‐groups of CLL present with distinct drug sensitivity and proteomic profiles, we next investigated whether genetic or proteomic features are more predictive of ex vivo drug sensitivity. These features could serve as a proxy to drug responses, since we showed that ex vivo drug sensitivities reflect reported clinical efficacy. Here, random forest modeling was used to predict ex vivo sensitivity to single agents and drug combinations separately. We found that the top five predictors of single‐agent sensitivity were pSAPK/JNK, predicted PFS (see section “Development and validation of a functional prognostic model for prediction of clinical response to ibrutinib + venetoclax treatment” below), pMEK1, TP53 mutational status, and pp38 MAPK (Supporting Information S1: Figure S7a). The top five predictors of sensitivity to drug combinations were pMEK1, pSAPK/JNK, pSYK, pSTAT6, and pp38MAPK (Supporting Information S1: Figure S7b). To confirm that the findings were real and not random artifacts generated by the model, a null distribution was created for each feature by randomly permutating the target feature and repeating the analysis (Supporting Information S1: Figure S8). The number of times a drug showed up as more significant in the real data than in the permutation data was counted for significance cutoff of empirical P < 0.01 from the permuted data. We found that pp38 MAPK most often predicted sensitivity to single agents (Supporting Information S1: Figure S7c), while pSAPK/JNK was most predictive for drug combinations (Supporting Information S1: Figure S7d). Taken together, proteomic features were most predictive of ex vivo sensitivity to both single agents and drug combinations. Investigating the directionality of ex vivo drug sensitivity features Since we established that proteomic features are predictive of ex vivo drug responses, we next asked what is the directionality of these features? In other words, is the absolute protein level associated with high or low drug sensitivity? For validation purposes, we also included the binary genetic and clinical features in the same analysis, and we chose to focus on the sensitivity to ibrutinib + venetoclax, since these drugs were administered in the trial. We then performed a SHAP analysis on ex vivo drug sensitivities predicted by the random forest model. We found that higher sensitivity to ibrutinib was associated with reduced phosphorylation of AKT, a predicted short PFS (see section “Development and validation of a functional prognostic model for prediction of clinical response to ibrutinib + venetoclax treatment” below), and reduced expression of Bcl‐2 (see Figure 4A and Supporting Information S1: Figure S9a for the test set and the training set, respectively). Next, we found that higher sensitivity to single‐agent venetoclax was associated with low expression of Bcl‐xL, high phosphorylation of Tyrosine residues, and low phosphorylation of MEK1 (Figure 4B and Supporting Information S1: Figure S9b). Finally, higher sensitivity to ibrutinib + venetoclax was associated with low phosphorylation of p53, high phosphorylation of TBK1, and low phosphorylation of AKT (Figure 4C and Supporting Information S1: Figure S9c). However, the separation of the features was most pronounced for ibrutinib as a single agent and less pronounced for the ibrutinib + venetoclax combination (Figure 4A,C). Genetic features, including IGHV and TP53 mutational status, did not show a clear correlation with any of the drug responses (Figure 4A–C). This trend was also observed for other clinically relevant BTK‐ and Bcl‐2 inhibitor single agents and combinations (Supporting Information S1: Figure S10). Together, these findings suggest that single features may be insufficient to stratify the response to the ibrutinib + venetoclax combination. This limitation may be overcome by developing a predictive biomarker signature, which we aimed to do next for the clinical efficacy of ibrutinib + venetoclax treatment. Single proteomic features are less predictive of ex vivo sensitivity to the ibrutinib + venetoclax combination than to single agents. (A–C) SHapley Additive exPlanations (SHAP) values were calculated to assess the impact of and directionality for each protein profile, genetic subtype, and clinical feature on predicating the Z‐scored DSSs. The model was trained using random forest regression with 500 decision trees. The data set was split into association training and test sets, with 20% of the data used for testing (n = 27) and the remainder used for training (n = 104). SHAP summary plots for the test set are shown for the drugs ibrutinib (A), venetoclax (B), and the combination ibrutinib + venetoclax (C). Each feature is represented by a horizontal bar, where the placement of an observation on the bar indicates the magnitude of the feature's contribution to the model's output as denoted on the x‐axis; values on the left‐hand side are associated with predicting lower drug sensitivity scores and values on the right are associated with predicting higher drug sensitivity scores. A wider band indicates a larger variation of that feature across the data set. Features are sorted by their importance, with the most influential features appearing at the top. Each dot in the plot represents a value for a specific feature for a specific patient. The dots are colored based on the feature's value. High values and the presence of binary features are marked in red, whereas low values or the absence of binary values are blue. Intermediate continuous values for the protein profiles are shown in varying shades of purple. The directionality of the drug sensitivity is indicated in gray below the X‐axis. Development and validation of a functional prognostic model for the prediction of clinical response to ibrutinib + venetoclax treatment To systematically investigate how clinical, genetic, and functional profiles could be leveraged to predict the clinical efficacy of ibrutinib + venetoclax treatment, we developed multi‐modal feature selection methods and machine learning models (Figure 5A). The patient cohort was stratified into a training set (n = 77) and a test set (n = 80), based on IGHV and TP53 aberration status, with similar PFS distributions (Supporting Information S1: Figure S11 and Supporting Information S7: Table S6) and clinical profiles (Supporting Information S8: Table S7). We found that regularized ridge regression was best at predicting PFS, both in the training set and in the test set, compared to other machine learning methods tested (Figure 5B,C and Supporting Information S1: Figure S12). The six most robust and predictive multi‐modal features were ex vivo sensitivity to the VEGFR‐2/EGFR inhibitor vandetanib, the JAK1/2 inhibitor ruxolitinib + venetoclax, and the BTK inhibitor acalabrutinib + the PI3K inhibitor ZSTK474; relative expression or phosphorylation of the protein pairs Bcl‐2:BTK (pY223) and p90RSK (pS380):PLCγ2 (pY759); and a high proportion of CD8+ T cells (Figure 5D and Supporting Information S1: Figure S13). Notably, these are all functional features, and we therefore named the model “functional prognostic model” (fPM). The coefficients estimated in the training set indicated that patient cells with higher ex vivo sensitivity to vandetanib, ruxolitinib + venetoclax, and acalabrutinib + ZSTK474, as well as cells with relatively higher Bcl‐2 expression compared with BTK phosphorylation and higher p90RSK phosphorylation relative to PLCγ2 phosphorylation, were associated with more favorable PFS. In contrast, a higher proportion of CD8+ T cells was associated with increased risk of progression, suggesting that elevated cytotoxic T‐cell abundance at baseline may reflect an unfavorable immune state in R/R CLL (Figure 5D). Development of a functional prognostic model to predict progression‐free survival. (A) Overview of the progression‐free survival (PFS) prediction workflow. (B, C) Survival curves of the functional prognostic model (fPM)‐based low‐ and high‐risk groups in the training set (n = 80, B) and the test set (n = 77, C). (D) Coefficients and hazard ratio (HR) of the variables of the fPM score estimated in the training set. The median value of the fPM score in the training set (−1.01687) was used as the cutoff to separate the high‐risk and the low‐risk patient groups. The statistical significance of the PFS differences between the two groups was assessed with the log rank test. The coefficients estimated in the training set indicated that the immunophenotype feature “proportion of CD8+ T‐cells” was unfavorable, while the other features were favorable for PFS outcome. For instance, coefficients of the drug sensitivity features were all negative, that is, high sensitivity to the treatments led to a lower fPM score and a tendency to classify the patient into the low‐risk group. The fPM model further outperformed the other machine learning methods in the test set with a higher discriminatory accuracy (Harrell's C‐index 0.766, time‐averaged ROC‐AUC 0.812), indicating its robust and predictive performance (Supporting Information S1: Figure S14). COX, RSF, and GBM models also showed relatively good performance in the test set, whereas the nonlinear support vector machine (SVM) performed poorly (Supporting Information S1: Figure S14). As expected for the relatively small sample size, all models demonstrated a degree of overfitting, indicated by lower performance in the test set than the training set. However, the blinded validation set (n = 20) confirmed that the selected functional features, when combined with the regularized ridge regression, could predict the risk of progression with a significant discriminatory performance (Harrell's C‐index 0.815, time‐averaged ROC‐AUC 0.805; Figure S15). The distribution of the fPM score was relatively similar across the different data sets (Supporting Information S1: Figure S16). Taken together, we developed an fPM that predicts response to ibrutinib + venetoclax treatment based on a small panel of six functional features. The model is available for testing in other cohorts (https://github.com/weikaixi/CLLproject). Comparative evaluation of the fPM score against established prognostic scores and unimodal models To investigate the predictive performance of the fPM score against established prognostic models and clinical scores, we used sub‐sets of the test set (n = 49), the validation set (n = 17), and the combined test and validation set (n = 66), which included patients with all required information about the genetic and clinical parameters available for calculation of the scores (Supporting Information S1: Figure S17). The fPM score demonstrated a stable and predictive performance in each of the independent data sets (Figure 6A,B). While the Charlson comorbidity index (CCI), 32 not addressing CLL risk factors but general comorbidity, showed good performance in the test set only (Harrell's C‐index 0.74, time‐averaged ROC‐AUC 0.68, Figure 6A,B), its performance deteriorated in the validation set (Harrell's C‐index 0.27, time‐averaged ROC‐AUC 0.28, Figure 6A,B), leading to near‐random accuracy in the combined test set. This was due to differences in the relationship between PFS time, CCI, and the fPM score in the two independent sets (Supporting Information S1: Figure S18). In contrast, the fPM score provided generalizable prediction of PFS for the patients in the combined test set (Figure 6C–G). The adjusted P‐values between the groups are shown in Supporting Information S9: Table S8. Predictive performance and survival curves of the functional prognostic model score and established prognostic scores. In the test set, the validation set, and the combined test set, there were 49, 17, and 66 patients, respectively, with complete genetic data (TP53 and IGHV status) and clinical information to calculate the established prognostic scores. (A) Time‐dependent receiver operating characteristics (ROC) analysis. (B) Harrell's C‐index. (C–G) Progression‐free survival (PFS) curves in the combined test set. The statistical significance of the PFS differences between the lowest‐risk group and the highest‐risk group was assessed with the log rank test and the Benjamini–Hochberg procedure. Rank‐based Harrel's C‐index was used to compare the mixed types of models and scores that either classify patients into various prognostic groups or output continuous prognostic scores. Ahn et al, Ahn et al 4 factor score; Binet, Binet stage; CCI, Charlson comorbidity index; CLL‐IPI, chronic lymphocytic leukemia international prognostic index. Compared to ridge regression models that utilized only one type of functional data (i.e., ex vivo drug sensitivity, protein profiles, or immunophenotype), the fPM score showed the highest performance (Supporting Information S1: Figure S19). The model based only on the protein profiling data (PM) showed worse than random accuracy in the independent test and validation sets (Supporting Information S1: Figure S19). In addition, we performed ablation experiments, in which one type of input data was removed at a time. Compared with the full fPM score, the model without drug sensitivity features (NDM) performed the worst (Figure 7A–F), indicating that ex vivo drug sensitivity data contributed most to the predictive power. Importantly, the PFS differences between the high‐ and low‐risk groups became insignificant after removal of any of the three input data sets (Figure 7C–F), demonstrating that all three types of input features are required and contribute to the model robustness. Contribution of multi‐modal functional features and comparison with genetic prognostic markers. (A–F) The ablation experiment results on removing one type of input data at a time from the multi‐modal functional Prognostic Model (fPM) score. (A) Time‐dependent receiver operating characteristics (ROC) analysis on removing or not removing one type of input data at a time. (B) Harrell's C‐index analysis on removing or not removing the three types of input data at a time. NDM, no‐drug‐sensitivity‐feature model; NIMM, no‐immunophenotype‐feature model; NPM, no‐protein‐profile‐feature model. (C–F) Survival curves in the combined test set. Statistical significance of the progression‐free survival (PFS) differences was assessed with the log rank test. (G–I) Comparative association analyses of the genetic features and the fPM score. (G) Heatmap of genetic features in the full patient set (n = 177). The patients are sorted in ascending order from left to right according to their fPM score. There were a total of 16 patients with missing genetic features (white color). The chi‐square test was used to assess the enrichment of binary genetic features between the low fPM score and high fPM score groups. n.s., not significant; *P < 0.05. (H) Association of the genetic features and PFS status in the training set (n = 69). (I) Association of the genetic features and PFS status in the combined test set (n = 92) using univariate Cox regression analysis. (J) Six functional features (drug sensitivity to VEGFR‐2‐, JAK1/2+Bcl‐2‐, and BTK + PI3K‐inhibitors; phosphorylation ratio for Bcl‐2:BTK and p90RSK:PLCγ2; and CD8+ T cells) are required to run the fPM model and risk‐stratify the patient. The functional features are collected by performing three biomarker assays (white boxes) on peripheral blood mononuclear cells from the patient. The fPM model needs to be validated in prospective clinical studies to prove the feasibility of clinical implementation (gray boxes). The model may then be applied to guide treatment decisions. In summary, integration of data from multiple functional assays through regularized regression techniques offers a robust prediction of PFS in R/R CLL treated with ibrutinib + venetoclax. The fPM score improved the predictive performance for PFS both in the test set and in the validation set, indicating its potential to guide personalized management strategies and improve upon conventional prognostic scores. Ex vivo sensitivity to ibrutinib + venetoclax alone does not predict clinical outcome To evaluate whether ex vivo sensitivity to the therapeutic agents used in the VISION/HO141 clinical trial could predict patient outcome, we performed univariate Cox regression analyses in the combined test set using ex vivo sensitivity scores for venetoclax, ibrutinib, and the ibrutinib + venetoclax combination as individual predictors of PFS. None of the three treatment‐related ex vivo sensitivity measurements showed a significant association with PFS (Supporting Information S1: Figure S20), with hazard ratios close to 1 for venetoclax (HR = 0.993, P = 0.852), ibrutinib (HR = 0.830, P = 0.224), and the ibrutinib + venetoclax combination (HR = 0.968, P = 0.391). These findings suggest that ex vivo sensitivity to the treatment drugs alone is insufficient to predict clinical outcome, supporting the added value of integrating multiple functional biomarkers in the fPM. Comparative evaluation of the fPM score against genetic features To broaden the perspective, we next compared the fPM score to traditional genetic markers such as IGHV mutation and TP53 aberrations. Although these features were included in the data integration, none were selected into the final fPM model, suggesting that functional parameters held greater predictive value. However, given their historical role in CLL prognosis, 33 we conducted a direct comparison to investigate whether the fPM score offers improved prediction of PFS. When investigating cytogenetic features and IGHV mutational status in the full patient cohort (n = 177), trisomy 12 aberration was enriched in the high‐risk fPM group (Chi‐square test, P < 0.05; Figure 7G). This suggests that the patients in the high‐risk and low‐risk fPM groups may have different cytogenetic backgrounds. However, we compared the fPM score against genetic features in terms of PFS prediction and performed univariate Cox regression analyses in the training and combined test sets, which showed that the assessed genetic features (IGHV mutational status, TP53 aberration, and cytogenetics) were not individually associated with the PFS in R/R CLL (Figure 7H,I), consistent with other reports. 5 , 34 This supports the observation that none of the genetic features were selected by the feature selection methods for the fPM risk score. Prior treatment history does not improve the predictive performance of the fPM score Since previous treatment exposure may influence disease biology in R/R CLL, we investigated whether prior CIT history contributed additional prognostic information beyond the established fPM. We first assessed the association between three major prior treatment regimens (FCR, BR, and chlorambucil‐based therapy) and the six selected fPM biomarkers using the chi‐square test in the combined test set. No significant associations were observed (Supporting Information S10: Table S9), suggesting that these functional biomarkers are largely independent of treatment history. To further evaluate the potential prognostic contribution of treatment history, we incorporated these three treatment variables together with the six original fPM features and re‐trained the ridge regression model in the training set. The coefficients of this regression model are shown in Supporting Information S1: Figure S21. When evaluated in the independent combined test set, this new model showed decreased predictive performance with relatively low time‐dependent ROC values (Supporting Information S1: Figure S22a), and Kaplan–Meier analysis failed to significantly stratify patients into distinct PFS groups (Supporting Information S1: Figure S22b). These findings suggest that prior treatment history does not provide additional prognostic information beyond the functional biomarkers already captured by the fPM. External validation supports the predictive ability of the fPM score in a mechanistically distinct treatment regimen To further evaluate the generalizability of the fPM, we applied the model to an independent external cohort of 26 R/R CLL patients treated with umbralisib monotherapy (NCT02742090) and with treatment response information, and all six required fPM input markers available. 30 Patients who responded to treatment tended to show lower fPM scores compared with nonresponders (Supporting Information S1: Figure S23a), indicating that patients predicted to have lower functional risk were more likely to benefit from therapy. Although this cohort lacked PFS data, the fPM demonstrated predictive ability for treatment response, achieving an ROC AUC of 0.706 (P = 0.043, permutation test) and an overall classification accuracy of 65.4% (Supporting Information S1: Figure S23b–d). Interestingly, comparison of fPM score distributions revealed that scores in the external cohort were shifted toward lower risk values and showed a narrower distribution compared with the original VISION/HO141 cohort (Supporting Information S1: Figure S23e). As a result, when applying the original risk threshold derived from the training cohort, all umbralisib‐treated patients were classified into the low‐risk group. Despite this compressed score range, the continuous fPM score remained predictive of treatment response, while the risk threshold could not be tested. These findings suggest that the biological signals captured by the fPM remain informative even when patient score distributions are restricted, providing preliminary support for the model's predictive ability when transferring to a mechanistically distinct treatment regimen. DISCUSSION Combination therapy with ibrutinib + venetoclax is approved as first‐line treatment of CLL and has been studied in several clinical trials in treatment‐naïve and R/R CLL, with consistently high efficacy across prognostic sub‐groups. 3 , 4 , 5 , 35 Known prognostic factors in CLL do not correlate well with outcome across targeted combination treatments, in particular, not for R/R CLL. Here, we used the VISION/HO141 trial to develop the data‐driven fPM model, demonstrating that functional phenotypes of primary CLL cells and their microenvironment in terms of drug sensitivity for single‐agent and combination treatment, phosphorylation balance between intracellular signaling pathways, and T cell distribution are predictive for PFS upon ibrutinib + venetoclax treatment in R/R CLL (Figure 7J). Features most predictive of PFS included ex vivo sensitivity to vandetanib (VEGFR‐2/EGFR inhibitor), ruxolitinib (JAK1/2 inhibitor) + venetoclax (Bcl‐2 inhibitor), and acalabrutinib (BTK inhibitor) + ZSTK474 (PI3K inhibitor); relative expression/phosphorylation level of Bcl‐2:BTK (pY223) and p90RSK (pS380):PLCγ2 (pY759); and a high proportion of CD8+ T cells, as integrated in the final fPM (Figure 7J). Our findings demonstrate that patterns of ex vivo drug sensitivity profiles can predict response to a given but different treatment, as previously reported also for other hematologic malignancies. 36 The drugs that we identified as predictive share a common mechanism wherein they inhibit signals from the tumor microenvironment that the CLL cells depend on for survival. 37 For example, binding of VEGF secreted from bone marrow stromal cells to VEGFR‐2 on CLL cells provides survival support for the CLL cells. 38 The general targeting of the tumor microenvironment may therefore explain why drugs other than Bcl‐2‐ and BTK‐inhibitors were also identified as predictive. It may also explain the treatment‐agnostic predictive power of fPM, that is, the modest predictive accuracy for umbralisib therapy, although the model was trained on ibrutinib + venetoclax data. Interestingly, the relative expression of Bcl‐2 to phosphorylation of BTK (Y223) contributed to the prediction. These are targets of the therapies administered in the trial, and preliminary data modeling of MRD status indicated that these proteins also contributed to MRD prediction. 39 Furthermore, our model showed that a high proportion of CD8+ T cells is unfavorable for PFS. This aligns with previous reports. 40 , 41 , 42 The inverted CD4:CD8 T‐cell ratio observed in CLL has been reported to be associated with exhaustion of CD8+ T cells, and high amounts of dysfunctional CD8+ T cells can contribute to progressive disease. 41 , 43 Exploration of the exhaustion status of the CD8+ T cells in such patient cohorts would thus be of interest for future studies. The fPM model, although developed in a fully data‐driven fashion, ends up representing features of the microenvironment, intracellular signaling, and drug sensitivity that are all representing known mechanistical features of CLL cell survival. A limitation of the model is that it was developed based on predefined functional profiles. It could be that other, yet unknown, features could further improve the predictions. Although targeted therapies have significantly improved CLL management, the disease is still considered incurable, and there remains a clinical need to identify effective therapies for traditional high‐risk CLL. We found that CLL cells with unmutated IGHV or a TP53 aberration were more sensitive to BTK inhibitor combinations ex vivo, while low‐risk CLL cells were more sensitive to Bcl‐2 inhibitor combinations. The MURANO and CLL14 trials showed that while venetoclax regimens are beneficial in CLL, patients with TP53 aberration or unmutated IGHV have a shorter PFS than patients without TP53 aberration or patients with mutated IGHV, respectively. 44 , 45 Ex vivo sensitivity to the combination of ibrutinib + venetoclax was not associated with these genotypes, which is in agreement with the clinical studies, which, despite a trend toward shorter PFS in the CAPTIVATE trial of ibrutinib + venetoclax for patients with TP53 aberrations, demonstrated noninferiority for ibrutinib + venetoclax as compared to continuous ibrutinib monotherapy in the CLL17 trial. 3 , 4 , 5 , 46 The ex vivo sensitivity to the next‐generation BTK inhibitor and Bcl‐2 inhibitor zanubrutinib + sonrotoclax was also not significantly associated with IGHV mutational status, TP53 aberration, or the analyzed karyotypes. This demonstrates that traditional prognostic markers are not associated with sensitivity to the BTK inhibitor + Bcl‐2 inhibitor combination ex vivo in general, as was also found for clinical efficacy in the VISION/HO141 trial used as a basis for the development of the fPM model. 5 Zanubrutinib + sonrotoclax is currently being studied in a phase I/II clinical trial in CLL, with encouraging results at a median follow‐up of 8.5 months, but stratification on prognostic sub‐groups has not yet been reported. 47 Considerable efforts are being made toward understanding how different regimens with targeted therapies compare to each other. For example, fixed‐duration ibrutinib + venetoclax was compared to ibrutinib monotherapy and fixed‐duration venetoclax + obinutuzumab in previously untreated CLL patients in the phase III clinical trial CLL17 (NCT04608318) demonstrating in early data noninferiority for the time‐defined combination treatments compared to continuous ibrutinib monotherapy. 46 While it is important to gain knowledge on the comparative efficacy of the available treatment regimens, it is equally important to understand which patient sub‐groups will benefit the most from the different regimens so that the treatment can be tailored to the individual patient. It has been reported that hyperphosphorylation of Bcl‐2 is a mechanism of resistance to venetoclax in CLL. 48 Our analysis showed that a high level of phosphorylated Bcl‐2 is associated both with high and low sensitivity to venetoclax, suggesting that there is a threshold for phosphorylation of Bcl‐2 that changes the random forest models' prediction of venetoclax sensitivity. The same effect was not evident for the second‐generation inhibitor sonrotoclax, where a high level of phosphorylated Bcl‐2 was primarily associated with higher drug sensitivity. This trend was also observed when venetoclax was paired with ibrutinib, indicating potentially different mechanisms of resistance to treatment with venetoclax as a single agent or combination. The findings presented here underscore the significant potential of integrating multi‐modal data from multiple functional assays using regularized regression techniques to predict PFS. The importance of correlation with clinical outcomes rather than in vitro phenotypes is emphasized by the differences between the ex vivo drug sensitivity, assessed as a pattern here, and the clinical outcomes. 49 This approach can be adapted to treatment regimens beyond ibrutinib + venetoclax, as the biomarker signature that we identified and operationalized in the fPM model may have predictive power for related regimens, as we showed for the umbralisib regimen, although fine‐tuning of the model in the context of different therapies may be required. Our study shows that a small number of functional features reflecting signaling pathways and mechanisms implemented in CLL development can predict PFS in R/R CLL treated with ibrutinib + venetoclax. The features are detected by flow cytometry, which is already implemented as a diagnostic tool, and by a luminescent cell viability assay, which has been used in clinical trials. 50 The test itself is relatively low cost, where the salary of the operator is the largest expense. Implementation of the biomarker in routine clinical practice is therefore expected to be technically and financially feasible, although it requires standardized and validated procedures, and expertise in sample handling and data analysis. Since our study was limited to retrospective analyses of R/R CLL, the model needs to be validated in independent patient cohorts, including targeted combination regimens for frontline CLL, and in a prospective clinical trial, before it may be implemented in clinical practice (Figure 7J). AUTHOR CONTRIBUTIONS Johanne U. Hermansen: Investigation; data curation; formal analysis; visualization; writing—original draft (supporting); writing—review and editing. Weikaixin Kong: Investigation; data curation; formal analysis; visualization; writing—original draft (supporting); writing—review and editing. Andrea M. Brodersen: Investigation; data curation; formal analysis; visualization; writing—original draft (supporting); writing—review and editing. Yanping Yin: Investigation; writing—review and editing. Aleksandra Urban: Investigation; writing—review and editing. Idun D. Rein: Methodology; writing—review and editing. Liye He: Investigation; writing—review and editing. Rudi Agius: Investigation; writing—review and editing. Rebecca S. Teglgaard: Investigation; writing—review and editing. Juho Rousu: Supervision (supporting); writing—review and editing. Christian Brieghel: Resources; writing—review and editing. Sabina Kersting: Resources; writing—review and editing. Mark‐David Levin: Resources; writing—review and editing. Hoa T. T. Tran: Resources; writing—review and editing. Mattias Mattsson: Resources; writing—review and editing. Juha Ranti: Resources; writing—review and editing. Gerrit‐Jan Veldhuis: Resources; writing—review and editing. Caspar da Cunha‐Bang: Resources; writing—review and editing. Rogier Mous: Resources; writing—review and editing. Julie Dubois: Resources; writing—review and editing. Arnon P. Kater: Conceptualization; resources; writing—review and editing. Jorrit M. Enserink: Supervision; writing—review and editing. Carsten U. Niemann: Conceptualization; supervision; resources; writing—review and editing. Tero Aittokallio: Conceptualization; supervision; writing—original draft (supporting); writing—review and editing. Sigrid S. Skånland: Conceptualization; supervision; methodology; investigation; formal analysis; visualization; writing—original draft (lead); writing—review and editing. CONFLICT OF INTEREST STATEMENT R.S.T. received financial research support from AstraZeneca outside of this work. C.B. has received travel grants from Octapharma outside of this study. S.K. reports research funding from AbbVie, AstraZeneca, Janssen, and Roche/Genentech. M.‐D.L. reports advisory board compensation from Janssen, AbbVie, and Roche, and travel reimbursement from AbbVie, Janssen, and Roche. H.T.T.T. reports consultancy fees from AbbVie, AstraZeneca, Bayer, Janssen‐Cilag, and Novartis. M.M. reports lecture remuneration from Janssen and advisory board compensation from AbbVie. J.Ra. reports consultancy fees from AbbVie, AstraZeneca, and Janssen. C.d.C.‐B. has received honoraria from Octapharma. J.D. received research funding from AbbVie. A.P.K. reports research grants from AbbVie, AstraZeneca, Celgene, Janssen, and Roche/Genentech; speaker's fees from AbbVie and AstraZeneca; and participation in advisory boards of AbbVie, AstraZeneca, Janssen, and Roche. C.U.N. reports research grants from AbbVie, AstraZeneca, Genmab, Janssen, and Octapharma, and consultancy compensations from AbbVie, AstraZeneca, BeOne Medicines, CSL Behring, Galapagos, Genmab, Janssen, Lilly, MSD, Roche, Synamics, Takeda, and Octapharma. T.A. has received unrelated research funding from Mobius Biotechnology GmbH. S.S.S. has received consulting fees from BeOne Medicines and Janssen, and research support from BeOne Medicines. The other authors declare no competing financial interests. ETHICS STATEMENT PBMCs were collected and cryopreserved at baseline from 177 patients with R/R CLL enrolled in the VISION/HO141 clinical trial (NCT03226301, Table 1). All participants signed a written informed consent prior to sample collection. Research on patient material was carried out in agreement with the Declaration of Helsinki. The study was approved by the Regional Committee for Medical and Health Research Ethics of South‐East Norway (28 507). FUNDING WK was supported by the Orion Research Foundation and K. Albin Johanssons Stiftelse. LH was supported by the Orion Research Foundation, Cancer Foundation Finland, and The Osk. Huttunen Foundation. JME was supported by grants from the Norwegian Health Authority South‐East, grant numbers 2017064, 2018012, and 2019096; the Norwegian Cancer Society, grant 208012; and the Research Council of Norway through its Centers of Excellence funding scheme (262652) and through grants 294916 and 314811. CUN was supported by the Innovation Fund Denmark under the frames of ERA PerMed (CLL‐CLUE) and EP PerMed (CLL‐OUTCOME), the Alfred Benzon Foundation, the Danish Cancer Society, and Sygesikringen Denmark. TA was supported by the Norwegian Cancer Society (grants 216104 and 273810), the South‐Eastern Norway Regional Health Authority (grants 2020 026 and 2023 105), the Radium Hospital Foundation, the Finnish Cancer Foundation, and the Research Council of Finland (grants 340 141 and 345 803), under the frames of ERA PerMed (CLL‐CLUE, grant 344 698) and EP PerMed (CLL‐OUTCOME, grant 367 855). SSS was supported by the Research Council of Norway under the frames of ERA PerMed (CLL‐CLUE, grant 322 898) and EP PerMed (CLL‐OUTCOME, grant 332 640), the Norwegian Cancer Society (grant 333 784), Fondsstiftelsen, Stiftelsen Kristian Gerhard Jebsen (grant 19), and the Norwegian Centre for Clinical Research (MATRIX: Research Council of Norway and the Norwegian Cancer Society joint grant 328 827). The VISION/HO141 clinical trial was funded by AbbVie and Janssen. These companies have commented on the manuscript, but not participated in conception or writing of the study. Supporting information Supporting File 1. Supporting File 2. Supporting File 3. Supporting File 4. Supporting File 5. Supporting File 6. Supporting File 7. Supporting File 8. Supporting File 9. Supporting File 10. Supporting File 11. ACKNOWLEDGMENTS We thank all the patients who contributed to this study. We thank Mentowa Fürst Bright and Nora Rojahn Bråthen at Oslo University Hospital for technical assistance. We thank Aram N. Andersen (Oslo University Hospital) for feedback and suggestions related to random forest, SHAP analyses, and permutation tests. We thank the High Throughput Biomedicine Unit at the Institute for Molecular Medicine Finland (FIMM), University of Helsinki, and HiLIFE for assistance with drug sensitivity screens. We thank the Flow Cytometry Core Facility at Oslo University Hospital for expert assistance and access to instrumentation. BeOne Medicines (Beijing, China) provided the compounds BGB‐10188, sonrotoclax, and zanubrutinib (ISR‐NCL‐BGB3111‐NonBGNE‐0003). DATA AVAILABILITY STATEMENT The nonsensitive functional data and codes for the methods that were used for the ex vivo and PFS predictions in individual samples are available on GitHub: https://github.com/Andreabrodersen/CLL-association-study and https://github.com/weikaixi/CLLproject, respectively. The full data set with pseudonymized patient data can be accessed on a collaborative basis at a secure data platform, ensuring the required data privacy protection upon reasonable request to the corresponding author: sigrid.skanland@ous-research.no. REFERENCES Associated Data Supplementary Materials Supporting File 1. Supporting File 2. Supporting File 3. Supporting File 4. Supporting File 5. Supporting File 6. Supporting File 7. Supporting File 8. Supporting File 9. Supporting File 10. Supporting File 11. Data Availability Statement The nonsensitive functional data and codes for the methods that were used for the ex vivo and PFS predictions in individual samples are available on GitHub: https://github.com/Andreabrodersen/CLL-association-study and https://github.com/weikaixi/CLLproject, respectively. The full data set with pseudonymized patient data can be accessed on a collaborative basis at a secure data platform, ensuring the required data privacy protection upon reasonable request to the corresponding author: sigrid.skanland@ous-research.no.