AI-derived prediction of response and relapse to venetoclax plus hypomethylating agent based therapy in Acute Myeloid Leukemia Free

Introduction

Venetoclax combined with hypomethylating agents (Ven-HMA) has become a standard of care for older or unfit patients with acute myeloid leukemia (AML). Despite high initial response rates, a significant proportion of patients experience primary refractory disease or subsequent relapse, which remain major clinical challenges. While current risk stratification systems, such as ELN guidelines, are useful for prognosis with intensive chemotherapy, their ability to predict outcomes with Ven-HMA is limited. Robust, validated biomarkers to guide treatment selection and anticipate failure are urgently needed. We hypothesized that artificial intelligence (AI) models, capable of integrating complex, multi-modal data, could accurately predict individual patient outcomes. We therefore developed and validated AI models using comprehensive clinical, genomic, and transcriptomic data from a cohort of 128 patients with AML or high-risk myelodysplastic syndrome (MDS) to predict response and relapse following Ven-HMA therapy.

Methods We retrospectively analyzed a single-institution cohort of 128 patients treated with Ven-HMA. Baseline bone marrow samples underwent targeted next-generation sequencing (NGS). Genomic data from a 302-gene DNA panel were available for all 128 patients, while transcriptomic data from a 1600-gene RNA panel were available for a subset of 37 patients, with expression quantified as transcripts per million (TPM). Response was defined as complete remission (CR) or CR with incomplete hematologic recovery (CRi) per ELN 2022 criteria. For the primary model, clinical data (age, gender, de novo vs. secondary AML), laboratory parameters (circulating blast percentage, cytogenetics), and genomic data were used. The cohort was split into training (60%) and testing (40%) sets. A Bayesian statistical approach was used for feature selection to identify the most informative variables, which were then used to construct a random forest classifier. Due to the smaller sample size of the transcriptomic cohort (n=37), models incorporating RNA expression data were developed and validated using leave-one-out cross-validation (LOOCV) to maximize data utilization.

Results

The cohort (n=128) had a median age of 73 years (range 15-89), and 52% were male. The overall response rate (CR/CRi) was 75%. Among the 96 responders, 44 (46%) subsequently relapsed. Using a machine learning approach on the testing set (n=51), an AI model integrating 10 variables successfully predicted treatment response with a sensitivity of 92.3%, a specificity of 75.7%, and an Area Under the Curve (AUC) of 0.876 (95% CI: 0.748-1.000). The top predictive features included key mutations (TP53, RUNX1, ASXL1), cytogenetic risk, disease origin (de novo vs. secondary), gender, baseline circulating blast percentage, and treatment-related factors such as the number of cycles administered. In the 37-patient subset with RNA data, a model using the expression of just 20 genes demonstrated even higher accuracy for response prediction (AUC: 0.910; sensitivity 89.5%, specificity 88.2%). Most notably, while models using clinical and genomic data alone were insufficient for building a robust relapse prediction model (AUC < 0.700), the transcriptomic-based model was exceptionally powerful. It predicted future relapse among responders with a sensitivity of 94.7%, a specificity of 94.1%, and a near-perfect AUC of 0.981 (95% CI: 0.981-1.000).

Conclusions Our study demonstrates that AI-driven models can effectively integrate routine clinical and genomic data to reliably predict initial response to Ven-HMA therapy in patients with AML and high-risk MDS. Crucially, we show that baseline bone marrow transcriptomic profiling provides superior predictive power, not only for response but also for forecasting relapse—a critical and currently unmet clinical need. The ability to identify patients at high risk of relapse before treatment initiation could enable novel clinical trial designs and personalized post-remission strategies, such as early allogeneic transplant or investigational maintenance therapies. These highly promising models warrant validation in larger, multi-institutional cohorts to establish their clinical utility in personalizing treatment for patients with AML.