Mapping the Drug Combination Landscape for AML: An Integrative Ex Vivo Functional and Computational Approach

Introduction

Despite advances in AML treatment, the persistent challenge of treatment resistance and relapse highlights the critical need for effective drug combinations to improve outcomes. Cell line-based predictions often fail clinically due to inter- and intra-tumor heterogeneity in patients. To address the need for clinically relevant combinations, we expanded our BeatAML study (Bottomly et al. 2022; Eide et al. 2023) to capture combination responses and integrated multiomics data to identify molecular determinants of combination responses.

Method

Ex vivo functional screening was conducted on 596 primary AML samples (384 from BeatAML, 212 new samples) using 170 two-drug combinations derived from 50 unique monotherapies. Drug sensitivity scores (DSS) (Yadav et al. 2014) were derived from cell viability after 3 days. Random forest analysis identified clinical and genetic features determining combination sensitivity. Unsupervised clustering was performed on 341 RNAseq AML samples. Gene signatures for cellular hierarchy and inflammation (Zeng et al. 2022; Lasry et al. 2022) were measured and further validated in published single-cell data (van Galen et al. 2019). The independent drug action (IDA) principle (IDACombo package) was applied to predict combination efficacy from monotherapy (Palmer et al. 2017; Ling et al. 2020). Predicted DSS was compared with ex vivo using Spearman correlation. Findings were validated on another cohort using 168 monotherapy data from 377 uninvolved BeatAML samples. Clinical outcomes and ex vivo Venetoclax+Azacitidine (VenAza) DSS were derived from 12 patients treated with VenAza (Chow et al. Blood 2023).

Result

We generated 170 combination response data in 596 samples, making it the largest drug combination dataset in AML primary tumors. Combination DSS exceeded monotherapy DSS (p < 2.2x10^-6), with 86 combinations showing combination index, CI<1, outperforming constituent monotherapies. Notably, 68 combinations surpassed VenAza in both efficacy and CI, including approved drug partners like Ven (n=12), Trametinib (n=11) and Idelalisib (n=11). Given the substantial molecular and clinical heterogeneity in AML, we integrated multiomics data with combination responses to uncover molecular determinants driving combination sensitivities. This revealed FAB subtype as major predictive biomarker compared to other clinical and genetic features. Transcriptomic profiling of 341 samples yielded 6 distinct clusters (C1-C6) with associations in mutations, fusions, and FAB subtypes (C1 – M3 & FLT3-ITD; C2 – Residual; C3 – M0; C4 – M5; C5 – M1 & NPM1; C6 – CBFB-MYH11). Cell maturation status emerged as the primary driver of cluster formation (C1, 3, 5 – primitive; C2, 4, 6 – mature). Pathway analysis identified enrichment of inflammatory-related pathways in mature clusters relative to primitive clusters. Single cell transcriptomics further confirmed higher inflammatory gene signatures in promonocyte-like, monocyte-like, cDC-like cells compared to LSPC-like malignant cells. These clusters showed differential response towards 77 combinations, with biases in primitive clusters for Ven + Aza/Ibrutinib and mature clusters for Ven + Palbociclib/Quizartinib. To expand our analysis beyond the 170 experimentally tested combinations, we employed the IDACombo approach, enabling predictions for a larger set of combinations (n=11,097). IDACombo predictions closely matched actual results (Spearman ρ = 0.74) and displayed robustness (n=146/170, Spearman ρ = 0.71) across independent BeatAML samples. This analysis identified Ven as a top combination partner with kinase inhibitors (PI3Ki, MEKi, mTORi) and epigenetic modifiers (BETi, HDACi). Finally, ex vivo VenAza DSS served as good predictor of clinical outcome (ROC-AUC=0.8), validating functional screens as a clinically meaningful approach.

Conclusion

By integrating ex vivo functional drug screening and computational IDA prediction model, we identified effective drug combinations for molecularly and clinically heterogeneous AML patients. By integrating multiomics data and rigorously testing over 100 different variables, our findings reinforced cell maturation states as determinants of combination drug responses. A novel finding reveals maturation status is accompanied by activation of inflammatory states in AML, indicating a role of differentiation block in shaping immune response.