Integrating Computational Cytomorphological Analysis of Bone Marrow Smears and Clinical Data for Predicting Response to Venetoclax and Hypomethylating Agents in AML

Introduction

Combination therapy by venetoclax and hypomethylating agents (Ven+HMA) is one of the most promising treatments for acute myeloid leukemia (AML) patients unfit for or relapsing from intensive chemotherapy. Novel predictive biomarkers are needed for patients receiving Ven+HMA, as only a proportion of patients have an optimal response. Despite advances in computer vision, the utility of bone marrow (BM) morphology as a resource of treatment biomarkers has not been sufficiently studied due to the lack of high-resolution imaging data and clinical grade image analysis algorithms.

Methods

We collected clinical data and cytomorphological BM slides from 100 AML patients treated with Ven+HMA, including 79 patients from the nationwide VenEx trial (NCT04267081, Kuusanmäki, Haematologica 2023) by the Finnish AML Study Group and 21 patients treated at the Helsinki University Hospital outside the trial. Out of these patients, 49 had de novo AML, and 51 had relapsed/refractory (R/R) or secondary AML (sAML). The median progression-free survival (PFS) time was 406 [range 1-2256] days and 174 [1-1006] days for de novo and R/R or sAML patients, respectively.

BM aspirate smears collected from patients before Ven+HMA treatment were digitized at a whole-slide level at 10x magnification and representative regions at 100x magnification with the Vision Hema Ultimate slide scanner (West Medica). We analyzed images with the Hematoscope application operating 15 deep learning-based image analysis algorithms performing BM sample segmentation, cell detection, classification, dysplasia detection, and cell morphometry (e.g., size, shape, color) analysis (unpublished). In median, 3351 [1124-5520] single cells were analyzed for each sample. The cytomorphological data was combined with mutational, demographic, and clinical data. To predict treatment response (PFS) and to find the most predictive features, we trained ElasticNet-penalized multivariate Cox regression models using a repeated nested 3-fold cross-validation (CV) setup. For each of the 400 repeats, a mean time-dependent area under the receiver operating characteristic curve (iAUC) over 3 evaluation points was calculated for model evaluation.

Results

The median iAUC score was 0.78 [Interquartile range 0.76-0.80] for the models fit on all data. The most predictive clinical factors included R/R or sAML disease (median HR 1.52 [1.42-1.82]) indicating poor response, as well as IDH2 mutation (0.94 [0.90-1.00]), and NPM1 mutation (0.89 [0.85-0.98]) as biomarkers of treatment sensitivity. Elevated bone marrow cellularity quantified as the whole-slide proportion of eukaryotic cells (1.10 [1.05-1.25]) was associated with a poor response. In blasts, increased variability of nuclei eccentricity (e.g., deviation of circularity) (1.07 [1.03-1.20]) and increased variability in nuclear-cytoplasmic ratios (1.14 [1.09-1.32]) were associated with a poor response while a higher median nuclear-cytoplasmic ratio (0.94 [0.91-1.00]) was associated with an improved response.

Lastly, we repeated the CV setup with unpenalized Cox regression models using the most predictive features. A model fit with disease etiology, NPM1, and IDH2 statuses achieved a median iAUC of 0.80 [0.79-0.81], while with disease etiology and cytomorphological features iAUC was 0.81 [0.80-0.82]. Disease etiology alone achieved an iAUC of 0.73 [0.72-0.73]. An iAUC of 0.84 [0.82-0.85] was achieved by combining disease etiology, NPM1, IDH2, and cytomorphological features.

Conclusions

We developed a framework for predicting Ven+HMA response in AML. Recent developments in computer vision permit reliable quantitative analysis of BM morphology and add a novel data modality to predictive modeling. Future work will focus on further exploring quantitative cytomorphology and linking it with other modalities and predictive assays.

Rimpiläinen:from AbbVie, Pfizer and Sanofi: Other: Travel costs. Siitonen:Abbvie, Amgen, GSK, Jansen-Cilag, Novartis, Novo-Nordisk, Takeda: Consultancy. Pyörälä:Abbvie, Amgen, Bristol-Myers Squibb, Pfizer, Servier: Other: Personal Fees/Travel costs. Porkka:Novartis: Research Funding; Incyte: Research Funding; Roche: Research Funding. Kontro:Faron Pharmaceuticals: Consultancy; AbbVie: Consultancy, Membership on an entity's Board of Directors or advisory committees, Research Funding; BMS: Membership on an entity's Board of Directors or advisory committees, Speakers Bureau; Novartis: Membership on an entity's Board of Directors or advisory committees; Pfizer: Membership on an entity's Board of Directors or advisory committees, Speakers Bureau; Astellas: Consultancy, Membership on an entity's Board of Directors or advisory committees; Servier: Membership on an entity's Board of Directors or advisory committees, Speakers Bureau; Immedica: Membership on an entity's Board of Directors or advisory committees. Brück:Gilead Sciences: Research Funding; Pfizer: Research Funding; Amgen: Consultancy; GSK: Consultancy; Roche: Consultancy; Sanofi: Consultancy; Novartis: Consultancy; Hematoscope: Current equity holder in private company.