AI-enabled multi-omics pharmacogenomic modeling guides resistance-aware multitarget optimization of venetoclax therapy in acute myeloid leukemia

Abstract

Acute myeloid leukemia (AML) remains difficult to treat because profound molecular heterogeneity enables rapid adaptive resistance to targeted therapy. We developed an AI-driven pharmacogenomic pipeline that integrates genomic, transcriptomic, epigenetic, and clinical features to predict patient-specific response to venetoclax and translate resistance-associated signals into actionable therapeutic hypotheses. Across internal and external cohorts, the deep learning model showed strong discrimination and generalization (AUROC 0.84-0.90), produced calibrated probabilities, and stratified overall survival independent of standard prognostic factors. Explainable modeling (SHAP) identified mechanistically coherent drivers of sensitivity and resistance: high BCL2 and apoptotic priming favored response, whereas MCL1 upregulation, TP53 disruption, and RAS/MAPK activation were dominant resistance programs. Translating these findings into therapy design, network-based modeling and in silico perturbation prioritized rational combinations expected to block escape routes, including venetoclax plus MCL1 inhibition (e.g., AZD5991-class inhibitors), venetoclax plus FLT3 inhibition (e.g., gilteritinib-class agents) in signaling-driven disease, and venetoclax plus p53-axis modulation (e.g., MDM2 inhibition) in TP53-altered contexts. Structural candidate evaluation using ensemble docking provided supportive drug-target interaction evidence for prioritized dependencies. Together, these results establish a clinically interpretable, resistance-aware AI framework for precision optimization of venetoclax-based combination therapy in AML.

Introduction

Leukemia remains a major global health burden, with acute myeloid leukemia (AML) showing five-year survival rates below 30% in adults despite therapeutic advances¹. Large genomic studies have identified at least 11 recurrent molecular subtypes, with FLT3-ITD, NPM1, TP53, and RAS pathway mutations defining distinct prognostic groups and therapeutic vulnerabilities². The dramatic success of targeted therapy in chronic myeloid leukemia, where BCR-ABL1 inhibition yields near-normal survival, illustrates the promise of precision oncology for hematologic malignancies³.

Artificial intelligence (AI) and machine learning (ML) are increasingly being used to compress high-dimensional biomedical data into actionable drug-design decisions, including target prioritization, resistance forecasting, and candidate evaluation^4–7. In precision oncology, this is particularly important because tumor heterogeneity and pathway redundancy frequently lead to rapid therapeutic failure when treatment strategies rely on single mechanisms. Recent advances in AML therapy, including FLT3 inhibitors such as midostaurin⁸, IDH1/2 inhibitors^9,10, and the BCL-2 inhibitor venetoclax^11–13, have significantly improved response rates. However, resistance and relapse remain common, driven by TP53 loss, MCL1/BCL-XL upregulation, and clonal evolution^14–19. Functional genomic studies and large-scale initiatives such as Beat AML further highlight the need for predictive models that integrate multi-omics data with drug sensitivity profiles^20,21.

However, few frameworks systematically integrate multi-omics data, model interpretability, and resistance-aware combination prioritization within a single clinically oriented pipeline for venetoclax-treated AML.

In this study, we developed an AI-enabled pharmacogenomic model that integrates genomics, transcriptomics, and clinical data to predict venetoclax response in AML, uncover molecular drivers of drug sensitivity and resistance, and rationally prioritize combination therapies such as dual BCL-2/MCL-1 or FLT3/BCL-2 inhibition. Our approach exemplifies how AI and systems biology can transform multi-omic data into clinically actionable insights.

Here, we present an AI-enabled multi-omics pharmacogenomic framework to (i) predict venetoclax response and resistance, (ii) identify resistance-associated pathways and actionable target sets, and (iii) prioritize resistance-aware multitarget therapeutic hypotheses for downstream candidate evaluation. The workflow integrates model interpretability to convert predictions into drug-design decisions and couples these decisions with in silico evaluation (e.g., structural/interaction evidence) while treating toxicity and safety signals as constraints during prioritization.

Results

Cohort characteristics and multi-omics integration

primary endpoint = venetoclax response in AML, other subtypes = contextual/robustness. We assembled a multi-center cohort (N = 580) with a primary focus on AML (58%), which constituted the main population for venetoclax response modeling, while ALL, CML, and CLL were included for contextual comparison (Fig. 1A). Within the AML subset, a total of n = 336 patients were included, comprising 168 responders and 168 non-responders based on integrated response definitions derived from ex vivo IC₅₀ thresholds and clinical remission status. Patients contributed rich molecular data (Fig. 1B) with >80% coverage for mutation and transcriptome profiles; ~50% had all major modalities like genomics, transcriptomics, and clinical data, enabling integrative modeling (Fig. 1C, D). Common alterations included FLT3-ITD, NPM1, and DNMT3A in AML, BCR-ABL1 in CML, and both B- and T-ALL subtypes.

Fig. 1. Comprehensive cohort characterization, multi-omics data integration, and therapy response landscape for AI-enabled pharmacogenomic modeling in leukemia.

A Cohort composition showing patient distribution across acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), chronic myeloid leukemia (CML), and chronic lymphocytic leukemia (CLL), reflecting real-world prevalence and clinical representation. B Patient-level data availability heatmap displaying coverage of six major modalities-somatic variants, RNA-seq, copy number variation (CNV), methylation, clinical annotations, and drug response data-highlighting heterogeneity typical of multi-omics studies. C Modality coverage summary indicating the proportion of patients contributing each data type, with >80% coverage for genomic and clinical features, ensuring robust downstream modeling. D Data overlap analysis demonstrating combinatorial distribution of modalities across patients, emphasizing integrative potential for multi-omics AI modeling. E Therapeutic class distribution across the cohort, dominated by tyrosine kinase inhibitors (TKIs), with additional representation of BCL-2 inhibitors, hypomethylating agents, and thiopurines. F Drug-wise patient counts showing exposure frequencies for key agents including imatinib, dasatinib, venetoclax, and azacitidine. G Therapy outcomes summarizing responder vs. non-responder proportions per drug, providing a basis for supervised drug-response modelling. H UMAP visualization of transcriptome profiles before (left, colored by data source) and after batch-effect correction (right, colored by leukemia subtype), showing improved biological clustering and subtype separation critical for downstream predictive modeling.

Therapeutic exposure profiling (Fig. 1E–G) showed TKIs as the dominant class (32% of treatments), followed by venetoclax (15%), hypomethylating agents (14%), and IDH1/2 inhibitors. Key agents included imatinib (n = 45), venetoclax (n = 80), and gilteritinib (n = 20). Responses were defined per drug (e.g., CR/clearance for venetoclax in AML), with an overall response rate of ~48%. In AML patients, venetoclax was primarily administered in combination with hypomethylating agents (e.g., azacitidine) or low-dose cytarabine, consistent with standard-of-care regimens, providing clinically relevant therapeutic context for response modeling.

To harmonize data, we applied batch correction; UMAP plots (Fig. 1H) shifted from study-driven to biologically driven clustering, cleanly separating AML, ALL, CML, and CLL, with further resolution of AML subtypes and CML phases. This integration produced a unified dataset suitable for downstream predictive modeling.

Mutation landscape and pharmacogenomic associations

We examined recurrent somatic mutations by response status (Fig. 2A). BCR-ABL1 enrichment is from the broader therapy-response landscape (TKI strata), and is not interpreted as venetoclax resistance biology. (70% vs. 0%, p < 0.001), reflecting TKI-treated patients labeled as venetoclax non-responders. FLT3-ITD (27% vs. 12%, p = 0.002), TP53 (19% vs. 5%, p = 0.004), and RUNX1 (15% vs. 5%, p = 0.009) were also overrepresented in non-responders, consistent with adverse-risk biology, whereas NPM1 mutations were twice as frequent in responders (30% vs. 15%, p = 0.012), supporting their favorable prognostic value. The patterns were as expected based on mutation-drug association heatmaps (Fig. 2B): BCR-ABL1 was expected to predict a high response to TKI and poor venetoclax sensitivity, an association was expected between FLT3-ITD and FLT3 inhibitors, and TP53 was expected to be negatively associated with a range of drugs. Responders had minor enrichment of IDH1/2 mutation, which is consistent with reported venetoclax sensitivity.

Fig. 2. Integrated pharmacogenomic landscape linking driver mutations, drug associations, and metabolizer-dependent toxicity outcomes in leukemia patients.

A Mutation prevalence by therapy response status showing significant enrichment of BCR-ABL1 (p < 0.001), FLT3-ITD (p = 0.002), TP53 (p = 0.004), and RUNX1 (p = 0.009) in non-responders, whereas NPM1 shows higher prevalence in responders (p = 0.012), suggesting its potential as a favorable prognostic biomarker. B Mutation-drug association heatmap displaying standardized association Z-scores (red = positive, blue = negative) between recurrent driver mutations and major therapeutic classes. Strong positive associations are observed for BCR-ABL1 with TKIs (Imatinib, Dasatinib) and FLT3-ITD with Venetoclax, whereas TP53 mutation exhibits negative associations across multiple agents, implicating broad resistance. C Hepatotoxicity analysis (ALT levels) across TPMT/NUDT15 metabolizer phenotypes using Kruskal-Wallis testing (p < 0.0001), demonstrating a stepwise increase from NM → IM → PM consistent with impaired thiopurine clearance. D Relative dose intensity (% of planned dose) showing significant dose reduction in IM and PM groups (one-way ANOVA, p < 0.0001), highlighting clinically relevant underdosing risk in genetically predisposed patients. E Distribution of CTCAE v5.0 toxicity grades across phenotypes, with poor metabolizers displaying the highest frequency of grade ≥ 2 events (Chi-square test, p < 0.001), reinforcing the clinical necessity of genotype-guided dosing. F Co-mutation network of resistant patients visualized using Cytoscape, with node size proportional to mutation frequency and edge thickness reflecting co-occurrence strength. Hub nodes such as TP53, RUNX1, FLT3, and KRAS form densely connected sub-networks, representing potential multitarget therapeutic opportunities for overcoming acquired resistance.

Pharmacogenetic determinants of thiopurine toxicity were evaluated as a proof-of-principle to confirm that the integrated dataset recapitulates established genotype–phenotype relationships. TPMT/NUDT15 poor metabolizers showed significantly increased toxicity and reduced dose tolerance (p < 0.001), consistent with known pharmacogenomic evidence. This analysis supports the biological validity of the dataset used for downstream AI modeling.

AI model predicts venetoclax response and identifies key resistance mechanisms

Using the integrated multi-omics feature set, we trained a deep neural network classifier to predict venetoclax sensitivity. Differential expression analysis (Fig. 3A) revealed upregulation of MCL1, PIK3R1, and AKT3 in non-responders, whereas pro-apoptotic genes BCL2L11 (BIM) and PMAIP1 (NOXA) were enriched in responders, suggesting an apoptosis-primed state. Gene set enrichment analysis (Fig. 3B) confirmed significant enrichment of p53 signaling and mitochondrial apoptosis in responders, while anti-apoptotic MCL1 programs dominated resistant cases (NES = −2.18, q < 0.001). UMAP visualization (Fig. 3C) demonstrated clear separation of high- vs. low-sensitivity patients, with responders clustering by NPM1 mutation and high BCL2 expression, and non-responders enriched for TP53, RAS, and FLT3-ITD mutations. SHAP analysis (Fig. 3D) ranked BCL2 expression and BIM as top positive predictors, while MCL1 overexpression and TP53 mutation had the strongest negative contributions; FLT3-ITD, DNMT3A, and IDH1/2 mutations contributed positively in favorable contexts, whereas NRAS/KRAS activation reduced sensitivity. Model performance (Fig. 3E) showed an AUROC 0.84 ± 0.03, outperforming logistic regression and tree ensembles, with moderate calibration (Brier = 0.20), supporting interpretable probability estimates while indicating scope for further calibration refinement. Together, these results demonstrate that the model captures a biologically meaningful venetoclax response signature, recapitulates known resistance mechanisms, and nominates actionable strategies such as dual BCL2 + MCL1 inhibition.

Fig. 3. Multi-layer integrative modeling identifies transcriptomic signatures, pathway dependencies, and predictive features underlying Venetoclax response in leukemia.

A Volcano plot showing differentially expressed genes between responders and non-responders (FDR < 0.05, |log₂FC|> 1); resistance genes (MCL1, BCL2L11, PIK3R1, SAMD9L) are upregulated in non-responders, highlighting anti-apoptotic and PI3K/AKT signaling. B GSEA reveals negative enrichment of MCL1 escape programs in responders (NES = −2.18, FDR < 0.001) and upregulation of apoptotic/p53 pathways, supporting model validity. C UMAP shows clear separation of predicted high- vs. low-sensitivity patients by AI-derived drug scores. D SHAP analysis ranks top 15 predictors, with apoptotic genes (BCL2L11, MCL1), TP53, and epigenetic regulators (DNMT3A, IDH1/2) driving Venetoclax response prediction. E ROC curves compare algorithms; the deep neural network (AUC = 0.84 ± 0.03) outperforms logistic regression (AUC = 0.71) and tree-based models (AUC = 0.78–0.80), confirming deep learning superiority.

Clinical validation and survival stratification

We next assessed the model’s clinical relevance by testing survival stratification (Fig. 4A–C). In the discovery cohort, AI-predicted venetoclax responders (score > 0.5) had markedly longer overall survival (median OS 18.7 vs. 9.3 months; HR = 0.46, 95% CI 0.34-0.61, p = 4.2 × 10⁻⁹). This survival advantage persisted across four external cohorts (TCGA-AML, GSE12417, GSE6891, BeatAML), with consistent hazard ratios (~0.4–0.6) and significant log-rank p < 0.01 (Fig. 4B). These results indicate that the model generalizes across platforms and captures a biologically favorable AML subset characterized by NPM1 mutation, high BCL2, and low MCL1; however, given the lack of uniform venetoclax treatment across cohorts, the observed survival differences may primarily reflect underlying prognostic biology rather than direct predictive validation of treatment response. Importantly, multivariable adjustment confirmed that the model’s prediction is an independent prognostic factor, suggesting it reflects mitochondrial priming rather than simply recapitulating standard risk stratifiers. Accordingly, survival stratification should be interpreted as supportive evidence of biological relevance rather than a direct measure of treatment-specific predictive performance. UMAP projections (Fig. 4C) further support biological plausibility, showing that predicted responders form a distinct, conserved transcriptomic cluster across cohorts.

Fig. 4. AI-predicted Venetoclax response stratifies survival outcomes and generalizes across independent AML cohorts.

A Kaplan-Meier survival analysis comparing AI-predicted responders (blue) and non-responders (orange) in the discovery cohort. Responders demonstrate significantly prolonged overall survival (median OS: 18.7 vs. 9.3 months, hazard ratio [HR] = 0.46, 95% CI: 0.34-0.61; log-rank p = 4.2 × 10⁻⁹), with shaded bands indicating 95% confidence intervals. B External cohort validation in TCGA-AML, GSE12417, GSE6891, and BeatAML datasets, showing consistent survival stratification with similar effect sizes (all log-rank p < 0.01). This confirms the robustness and transferability of the model across different patient populations and sequencing platforms. C 3D UMAP projections of transcriptomic data from each cohort, colored by AI-predicted sensitivity scores (0-1). Predicted responders cluster together in a high-sensitivity region, supporting the biological plausibility of the prediction and indicating that the model captures a conserved transcriptomic response phenotype.

Comparative modeling and modality contribution

To assess the value of multi-omics integration, we compared models trained on single modalities with the integrated model (Fig. 5A). The multi-omics model achieved the highest cross-validated performance (AUROC ≈ 0.90, PR-AUC ≈ 0.72), significantly outperforming genomics-only (0.76), transcriptomics-only (0.82), and clinical-only models (0.65), while performance on the independent test set remained robust (AUROC 0.84 ± 0.03). This highlights the complementary value of combining genomic alterations with expression signatures such as BCL2, BIM, and MCL1 levels.

Fig. 5. Comparative performance, calibration, and clinical utility of the AI-enabled multi-omics pharmacogenomic model.

A Performance by data modality expressed as AUROC and PR-AUC, demonstrating superior discriminative ability of the integrated multi-omics model (AUROC = 0.90, PR-AUC = 0.72) compared to single-modality models (genomics, transcriptomics, proteomics, or clinical data alone). Error bars represent 95% confidence intervals across 10-fold cross-validation. B Calibration curve of the top-performing model using 20-bin reliability analysis, showing close alignment between predicted probabilities and observed response frequencies at higher probability bins. The Brier score (0.200) and expected calibration error (ECE = 0.322) indicate acceptable calibration and probabilistic interpretability, crucial for clinical decision-making. C Decision curve analysis (DCA) illustrating net clinical benefit across a range of treatment probability thresholds. The multi-omics model (blue) provides the highest net benefit over both “treat all” and “treat none” strategies, with maximum benefit observed near a threshold probability of 0.05 (NB = 0.323), supporting its potential to inform risk-adapted therapy selection in real-world settings.

Calibration analysis (Fig. 5B) showed reasonable agreement between predicted probabilities and observed responses, particularly at higher probability bins (Brier = 0.20; ECE = 0.322), indicating moderate calibration performance and supporting probabilistic interpretability while highlighting the need for threshold-aware clinical application.

Decision curve analysis (Fig. 5C) confirmed net clinical benefit over “treat-all” or “treat-none” strategies across a wide probability range, with maximal benefit near a 5% treatment threshold (NB ≈ 0.32). The 5% threshold was selected to reflect a clinically conservative decision boundary aimed at maximizing sensitivity for identifying potential responders while minimizing the risk of excluding patients who may benefit from combination therapy. These findings demonstrate that integrating all available data yields more accurate, calibrated, and clinically actionable predictions, enabling risk-adapted therapy selection.

External validation and subtype-specific performance

Across independent cohorts, the model reproduced key molecular and clinical patterns. Immune profiling (Fig. 6A) showed that responders exhibited lower expression of inhibitory checkpoint markers (p < 0.01), suggesting a less immunosuppressive state consistent with apoptosis-sensitive phenotypes. External validation (Fig. 6B) in BeatAML confirmed higher MCL1 in resistant cases and higher BCL2 in sensitive ones (p < 0.001), with FLT3 mutations associated with greater predicted sensitivity. Pediatric TARGET-AML validation (Fig. 6C) showed similar expression trends (MCL1/MYC ↑ , BCL2 ↓ , FDR < 0.05) and AUROC ≈ 0.79, demonstrating cross-age generalizability. Subtype analysis (Fig. 6D) yielded AUROCs of 0.83 (CLL), 0.82 (AML), 0.81 (ALL), and 0.73 (CML), reflecting biological dependence on BCL-2: CLL scored uniformly high while CML scored low, aligning with clinical venetoclax sensitivity patterns^3,22. Collectively, these results confirm that the model generalizes across cohorts and ages, captures immune and apoptotic biology, and remains context-aware across leukemia subtypes.

Fig. 6. Immunogenomic correlates, cross-cohort validation, and subtype-specific model performance.

A Violin plots of immune checkpoint gene expression (CD274/PD-L1, CTLA4, HAVCR2/TIM-3, TIGIT, LAG3, PDCD1/PD-1, PDCD1LG2/PD-L2, SIGLEC15) stratified by responder status. Responders show significantly lower expression of multiple inhibitory checkpoints, suggesting a less immunosuppressive tumor microenvironment conducive to Venetoclax sensitivity. Statistical significance assessed by Wilcoxon rank-sum test (***p < 0.001, ****p < 0.0001). B External validation in the BeatAML cohort showing differential expression of Venetoclax resistance-associated genes (MCL1, BCL2, FLT3) with concordant directionality as in the discovery cohort, reinforcing biological reproducibility. C TARGET-AML volcano plot confirming upregulation of key pro-survival and signaling genes in non-responders (FDR < 0.05, |log₂FC|> 1), extending findings to pediatric/young adult AML populations. D Subtype-specific model performance quantified by AUROC with 95% confidence intervals. Model shows robust and consistent discrimination across CLL (AUROC = 0.83), AML (AUROC = 0.82), and ALL (AUROC = 0.81), with slightly lower performance in CML (AUROC = 0.73). Pooled performance across all subtypes remains high (AUROC = 0.80), demonstrating broad applicability of the model across major leukemia subtypes.

Global feature contributions and patient-specific explanations

SHAP analysis provided both global and individual interpretability. Global feature ranking (Fig. 7A) identified BCL2 and BCL2L11 (BIM) as top positive predictors, while MCL1 overexpression and TP53 mutation were dominant negative contributors. Dependence plots (Fig. 7B) confirmed monotonic, biologically consistent effects: high BCL2 and FLT3-ITD shifted predictions toward response (ρ = 0.71 and 0.29, p < 0.001), whereas elevated MCL1 or TP53 loss drove resistance (ρ = −0.38 and −0.51, p < 0.001). Patient-level explanation (Fig. 7C) illustrated a mixed-risk AML case (predicted probability ≈ 0.55) where strong positive drivers (high BCL2, FLT3-ITD) were partly offset by negative contributions (TP53 mutation, high MCL1). The resulting intermediate score transparently reflects competing biologic forces and suggests potential combination strategies, e.g., adding MCL1 or p53-targeted therapy. Collectively, these insights confirm that the model learns clinically meaningful relationships and enables case-specific decision support. Our finding that TP53 loss drives venetoclax resistance is consistent with studies showing synthetic lethality from concurrent BCL-2 inhibition and p53 activation^14,23.

Fig. 7. Global and patient-level model interpretability using SHAP analysis.

A Global SHAP summary plot ranking the top predictive features by mean absolute SHAP value. BCL2 expression, BCL2L11 (BIM), PMAlP1 (NOXA), FLT3 mutations, and TP53 mutations emerged as the most influential predictors of Venetoclax response, with directionality indicating that higher BCL2 expression favors sensitivity, whereas TP53 mutations drive resistance. B SHAP dependence plots for four key predictors (BCL2, MCL1, FLT3-ITD, TP53), showing smooth, monotonic relationships between feature values and model output. Higher BCL2 expression and FLT3-ITD positivity increase predicted probability of response (ρ = 0.71, ρ = 0.29, both p < 0.001), whereas MCL1 and TP53 dysregulation negatively impact response (ρ = −0.38 and −0.51, p < 0.001). C Patient-level SHAP waterfall plot demonstrating explainability for an individual case (predicted response probability = 0.55). Negative contributions from TP53 and MCL1 partially offset positive drivers such as BCL2 expression and FLT3-ITD, highlighting the utility of this approach for case-specific decision support. The accompanying table quantifies feature values and SHAP contributions for the top 10 features, allowing transparent interpretation of model output.

Mechanistic network modeling and combination therapy insights

We built an integrative network linking drugs, key genes, pathways, and outcomes to translate model findings into therapy (Fig. 8A). Venetoclax connected strongly to the BCL2-BIM-NOXA apoptotic module (green edges, response-associated), while TP53-MCL1 formed a resistance hub (red edges), highlighting that p53 loss or MCL1 overexpression blunt venetoclax efficacy. FLT3 inhibitors linked to the FLT3-RAS-MAPK module, which also stabilizes MCL1, rationalizing synergy between FLT3 and BCL-2 inhibition. IDH1/2 modules tied into BCL2, consistent with IDH-mutant AML’s BCL-2 dependence²⁴. Expanded gene-gene networks (Fig. 8B) revealed coherent clusters: apoptotic priming (BCL2, BIM, NOXA), anti-apoptotic resistance (MCL1, TP53), myeloid signaling (FLT3, KRAS, NRAS), epigenetic modulators (DNMT3A, IDH1/2), and immune checkpoints (PD-L1, LAG3). Notably, BCL2 and MCL1 showed an inverse expression pattern, suggesting targetable apoptotic dependency, and TP53 mutation correlated with immune checkpoint upregulation, indicating an “immune-cold,” resistant state. In silico perturbations (Fig. 8C) suggested that dual MCL1 inhibition and BIM upregulation may shift network topology toward a response-like state (+36% predicted sensitivity), and co-targeting BCL-2 + FLT3 or BCL-2 + IDH1 produced additive effects, providing hypothesis-generating evidence consistent with known synergistic mechanisms. More over like approach builds on recent machine-learning frameworks predicting AML drug response from multi-omic data, which have demonstrated strong predictive power and clinical relevance⁴. Together, this systems-level approach nominates rational combination strategies particularly venetoclax plus MCL1 or FLT3 inhibitors as promising routes to overcome resistance.

Fig. 8. Integrated systems-biology framework linking mechanistic insights, network interactions, and dynamic perturbation analysis.

A Mechanistic network: Hierarchical, multi-layer network connecting therapeutic drugs (left), key genes (BCL2, MCL1, TP53, FLT3, DNMT3A, IDH1/2, KRAS, NRAS, BCL2L11, PMAIP1), pathway modules (middle; apoptosis, anti-apoptotic escape, DNA damage/p53, myeloid signaling, epigenetic regulation), and clinical outcomes (Predicted Response ↑, Resistance ↑). Edge thickness reflects SHAP importance, and edge color denotes directionality (green = response-associated, red = resistance-associated). B Expanded interaction network: Force-directed co-expression and co-mutation network of the top 16 predictive genes. Node size represents SHAP importance, node color indicates association with response (blue), resistance (orange), or immune regulation (purple), and edge width/color encodes pairwise correlation strength (|r|≥ 0.5). Distinct functional modules emerge, including apoptotic priming (BCL2, BCL2L11, CASP3, BAX), anti-apoptotic resistance (MCL1, TP53), myeloid signaling (FLT3, KRAS, NRAS, NPM1), epigenetic modulators (DNMT3A, IDH1, IDH2), and immune checkpoints (CD274/PD-L1, TIGIT, LAG3).

Resistance mapping and multi-target combination design

A resistance map (Fig. 9A) revealed that venetoclax non-responders were enriched for MCL1 overexpression (21%), NRAS/KRAS mutations (~25%), and BCL2 alterations (27%), suggesting escape via anti-apoptotic reprogramming or target downregulation. FLT3-TKD mutations (~30%) and RAS pathway activation were common in FLT3 inhibitor resistance, while IDH2 mutations appeared in 29% of venetoclax-resistant cases, hinting at clonal selection. Using these data, we constructed a multitarget escape network (Fig. 9B) prioritizing rational combinations: BCL-2 + MCL-1 inhibition had the highest synergy score (1.59), followed by IDH1 + BCL-2 (1.55), both supported by preclinical and early clinical results. Other promising strategies include combining venetoclax with MEK inhibition to suppress MCL-1 stabilization in RAS-mutant AML, or adding p53 reactivators/MDM2 inhibitors to resensitize TP53-mutant disease. Structural mapping (Fig. 9C) localized recurrent FLT3 resistance mutations (ITD, TKD D835) and IDH1 hotspot mutations, informing future drug design and resistance monitoring. Together, these findings nominate multitarget strategies, particularly dual apoptotic blockade and pathway co-inhibition as high-priority approaches to overcome therapeutic escape.

Fig. 9. Resistance mapping, multitarget network analysis, and structural localization of mutational hotspots.

A Resistance map: Clustered heatmap of resistance-associated mutations across major targeted therapy classes (TKIs, IDH inhibitors, BCL2i, FLT3i). Prevalence (%) is color-coded, with asterisks marking statistically significant associations (q < 0.05). Notable enrichment was observed for BCL2 (27%), MCL1 (21%), FLT3 (27%), and IDH2 (29%) in resistant patients. B Multitarget escape network: Mechanistic map illustrating escape routes through parallel signaling nodes (PI3K-AKT-mTOR, MAPK, JAK-STAT) and cross-talk between BCR signaling and apoptotic priming. Suggested combination therapies are annotated, with BCL2 + MCL1 inhibition and IDH1 + BCL2 co-targeting receiving the highest combination synergy scores (1.59 and 1.55, respectively). The right panel shows ranked combination scores to prioritize rational drug development. C Domain-level mutation map: Lollipop plot depicting amino acid positions of mutational hotspots in FLT3 (ITD, TKD D835, AB domain) and IDH1 (Rossmann-like domain, catalytic pocket). Mutations are annotated with prevalence (%), highlighting actionable regions that may guide drug design and resistance monitoring.

Model optimization and component contribution

Transfer learning markedly enhanced model generalization across BeatAML, TARGET-AML, and internal cohorts, improving AUROC by +0.07-0.10 and PR-AUC by +0.05-0.09 (Fig. 10A). Quantitative accuracy was confirmed by a strong correlation between predicted and observed IC₅₀ values for multiple agents (R² = 0.91, RMSE = 0.26 log₁₀ μM, Fig. 10B). Ablation analysis (Fig. 10C) revealed that multi-omics integration (ΔAUROC−0.09) and transfer learning (ΔAUROC−0.06) were the most critical contributors, highlighting that both comprehensive feature inclusion and cross-cohort adaptation are essential for maximal predictive performance.

Fig. 10. Model performance improvement via transfer learning, predictive calibration, and ablation study.

A Transfer learning evaluation: Bar-and-marker plot showing AUROC (bars) and PR-AUC (markers) across BeatAML, TARGET-AML, and internal validation cohorts. Transfer learning consistently improves discrimination performance compared to baseline models (ΔAUROC + 0.07 to +0.10, ΔPR-AUC + 0.05 to +0.09), demonstrating superior generalization across cohorts. B Predicted vs observed IC₅₀: Log-scale scatter plot comparing model-predicted drug sensitivity (IC₅₀) versus observed ex vivo IC₅₀ values for Venetoclax, Enasidenib, Midostaurin, and Ibrutinib. High correlation (R² = 0.91, RMSE = 0.26 log₁₀ μM) confirms accurate quantitative prediction of patient-specific drug response. C Ablation study: Horizontal tornado plot showing performance deltas (ΔAUROC, ΔPR-AUC) after systematically removing each model component. Multi-omics integration (ΔAUROC−0.09) and transfer learning (ΔAUROC−0.06) contributed most to performance, underscoring their critical importance in the final model.

Multi-cohort validation and resistance landscape

Across four independent cohortsTCGA-AML, BeatAML, GSE12417, and an internal validation set predicted venetoclax sensitivity scores were significantly higher in responders than non-responders (p < 0.001; Fig. 11A–D), confirming the model’s robustness across diverse populations. Resistance profiling (Fig. 11E) revealed that KRAS, DNMT3A, and TP53 mutations, as well as MCL1 overexpression, were 2-3× more frequent in non-responders (p < 0.001), emphasizing their role in therapeutic escape and supporting rational combination strategies aimed at co-targeting these resistance drivers.

Fig. 11. Multi-cohort validation, resistance landscape, and survival impact of AI-predicted Venetoclax sensitivity.

Boxplots showing predicted Venetoclax sensitivity (0-1 scale) in Responders vs. Non-responders across four independent cohorts: A TCGA-AML, B BeatAML, C GSE12417, and D an internal validation cohort. Across all datasets, predicted sensitivity was significantly higher in Responders (p < 0.001, Mann-Whitney U test), demonstrating the robustness of the model across diverse patient populations. E Stylized lollipop chart depicting the prevalence of resistance-associated alterations (KRAS-mut, DNMT3A-mut, MCL1-high, TP53-mut) in Responders (teal) vs. non-responders (red). Non-responders consistently showed 2-3× higher prevalence of these alterations (p < 0.001), highlighting their role in therapeutic escape mechanisms and supporting the rationale for combination strategies targeting MCL1 and TP53 pathways.

Drug-target network and docking validation

An AI-predicted drug-target network revealed high-confidence interactions between key targets (BCL2, MCL1, FLT3, IDH1, MEK, BTK) and approved or investigational agents, highlighting opportunities for rational multi-target therapy design (top panel). Ensemble docking supported the predicted interaction between venetoclax and BCL2 (Fig. 12A), with consistent poses (A1-A5) occupying the canonical anti-apoptotic pocket. Key hydrogen bonds (ASP178, ASP267) and π interactions (Fig. 12B) validated strong and specific engagement, providing structural support for the model’s predictions and reinforcing BCL2 as a primary therapeutic dependency.

Fig. 12. Predicted drug-target network and structural docking validation.

(Top panel) In silico drug-target association network showing predicted interactions between key molecular targets (MEK, BTK, IDH1, FLT3, MCL1, BCL2) and FDA-approved or investigational drugs. Solid lines represent high-confidence associations (score ≥ 0.90) and dashed lines represent moderate-confidence associations, highlighting potential multi-target therapeutic opportunities. A Molecular docking pose of Venetoclax within the BCL2 binding pocket, depicted in cartoon representation (teal) with the ligand shown as sticks (CPK coloring). A1-A4 illustrate alternate binding conformations from ensemble docking. B Key residue interactions at the BCL2 binding site, with hydrogen bonds (yellow dashed lines) and π interactions (gray dashed lines) annotated. Distances between key residues (e.g., ASP178, ASP267) and ligand atoms are shown, confirming high-affinity engagement of Venetoclax with the anti-apoptotic pocket.

In summary, through our AI-driven integrative analysis, we have charted a map of leukemic vulnerabilities and escape routes. This allowed us to nominate combination therapies that concurrently hit multiple nodes of the network, thereby blocking the leukemia’s “escape hatches.” The highest-priority combinations emerging from this study - BCL-2 + MCL-1 dual inhibition for venetoclax-resistant AML, BCL-2 inhibition + FLT3/AXL inhibition for high-risk AML, and BCL-2 + IDH1 inhibition for IDH-mutant AML - are all either supported by preclinical evidence or in early clinical testing, lending credence to our model’s predictive power. Our analysis also underscores the importance of patient-specific factors: for instance, a TP53-mutant patient might benefit more from a venetoclax + MDM2 inhibitor combo, whereas a RAS-activated patient might do better with venetoclax + a MEK inhibitor. Such precision combination strategies are the next frontier in AML therapy, and AI models like ours will be instrumental in guiding their development and application.

Discussion

This study establishes an interpretable AI-enabled pharmacogenomic framework for resistance-aware optimization of venetoclax-based therapy in AML, extending beyond conventional response prediction toward actionable drug-design guidance. By integrating multi-omics and clinical covariates, the deep learning framework achieved robust discrimination and clinically meaningful calibration for venetoclax sensitivity across independent cohorts (AUROC ~ 0.84-0.90), indicating that the learned signature generalizes across platforms and populations. Importantly, the model’s dominant determinants align with established apoptotic biology: high BCL2 and BIM favor response, whereas TP53 disruption and MCL1 upregulation drive resistance^14–19,25. This is consistent with mechanistic evidence and functional perturbation studies showing that loss of p53 control and compensatory pro-survival buffering can blunt BCL-2 dependency^14,23. The model also recovered clinically observed contexts such as enhanced venetoclax sensitivity in IDH1/2-mutant AML, supporting the view that the framework is extracting actionable dependency patterns rather than fitting cohort-specific noise²⁴.

A central methodological contribution of this study is the translation of model outputs into drug-design decisions. Explainability (SHAP and patient-level decompositions) converts probabilistic predictions into ranked resistance drivers and pathway modules that can be acted upon, enabling target prioritization and resistance-aware optimization. The resistance programs identifiedTP53-linked apoptotic failure, MCL1-mediated escape, and RAS/MAPK-associated survival rewiring point to rational multitarget strategies that directly address pathway redundancy and adaptive escape. In this setting, the model functions as an in-silico screening layer that narrows the therapeutic search space from broad multi-omic signals to a concise set of druggable nodes and combination hypotheses, improving the feasibility of downstream lead evaluation and experimental validation.

The immune-associated signal adds an additional, clinically relevant dimension to therapy design. Venetoclax-sensitive cases displayed lower expression of inhibitory checkpoints, suggesting that immune suppression may co-segregate with resistance programs and potentially reduce the effectiveness of apoptosis-targeted strategies. While this association does not establish causality, it motivates clinically testable hypotheses: resistance-associated immune states may justify combination approaches that incorporate immune modulation alongside apoptosis-directed treatment, particularly in patients with high checkpoint expression. Within the SI framing, this is not a digression into immunology; rather, it provides a biological context that can inform patient-tailored combination design and stratified trial logic for venetoclax-based regimens.

The multitarget prioritization layer provides a concrete bridge from resistance mapping to candidate selection. Network modeling and perturbation simulations converged on combinations that block escape routes through complementary nodes, with venetoclax + MCL1 inhibition emerging as a top-ranked strategy and venetoclax + FLT3 inhibition or venetoclax + p53-axis modulation prioritized in genetically defined subgroups. The relevance of MCL1 targeting is supported by strong preclinical evidence, including the activity of the MCL1 inhibitor AZD5991 in hematologic malignancies²⁶, and several resistance-aware combinations consistent with our ranking are being investigated clinically²⁷. These findings highlight how AI can be used to propose multitarget solutions that remain effective as tumors evolve, matching the Collection’s emphasis on resistance prediction and multitarget drug design.

Several limitations frame the next steps for translation. First, although retrospective multi-cohort validation supports generalizability, prospective testing is needed to confirm clinical utility under real-world decision constraints. Second, multi-omics coverage remains incomplete across datasets, and future work should integrate proteomic and metabolomic layers and incorporate longitudinal measurements (e.g., MRD dynamics) to capture temporal adaptation. Third, while structural docking provides supportive candidate-evaluation evidence, more rigorous lead-optimization steps such as molecular dynamics, binding free-energy refinement, and PK/PD-ADMET-constrained scoring will strengthen deployability and patient safety. In addition, the proposed synergy predictions are based on in silico perturbation and network modeling and have not yet been experimentally validated in ex vivo or in vivo systems, and should therefore be considered hypothesis-generating. Overall, the present work demonstrates an end-to-end blueprint in which AI predicts venetoclax response, exposes resistance mechanisms, and prioritizes rational, resistance-aware multitarget strategies, thereby advancing the drug-design pipeline for precision oncology and supporting integration of AI-guided prioritization into prospective clinical workflows. In a clinical implementation setting, this framework could be deployed as a decision-support tool that integrates patient-specific multi-omics data to guide personalized combination therapy selection and stratified clinical trial enrollment.

This study demonstrates that AI-enabled pharmacogenomic modeling can predict venetoclax response in leukemia while uncovering biologically grounded determinants of sensitivity and resistance. By integrating multi-omic and clinical features, the deep learning framework achieved strong and transferable performance across independent cohorts and produced interpretable outputs that reflect established apoptotic control. High BCL2/BIM signaling and preserved p53-associated programs were linked to response, whereas MCL1 upregulation and TP53 alterations emerged as dominant resistance drivers, supporting that the model captures meaningful mechanisms rather than cohort-specific correlations and can function as a decision-support layer for treatment selection.

Beyond prediction, the framework advances an AI-guided drug-design logic by converting resistance signals into actionable multitarget strategies. Network-based prioritization and in silico perturbation analyses nominated rational combinations aimed at blocking escape routes, including dual BCL-2/MCL-1 inhibition, co-targeting BCL-2 with FLT3 in signaling-driven disease, and p53-axis modulation in TP53-disrupted contexts. Collectively, these results support the use of AI-driven multi-omics pipelines to forecast response, anticipate resistance, and prioritize resistance-aware combination hypotheses, providing a practical route toward more effective and personalized venetoclax-based therapy development in precision oncology.

Methods

Study objective and drug-design outputs

This study was designed as an AI-guided pharmacogenomic framework to predict venetoclax response and enable resistance-aware multitarget therapy optimization in acute myeloid leukemia. The primary outputs were: (i) patient-level probability of venetoclax response/non-response (operational resistance), (ii) ranked molecular determinants of resistance mapped to actionable pathways, (iii) prioritized multitarget/combination hypotheses to overcome resistance (e.g., BCL2 co-dependencies), and (iv) in silico candidate evaluation using docking as a structural support layer. Safety signals were treated as constraints during prioritization when available.

Patient cohorts and data sources

This study was conducted in accordance with the Declaration of Helsinki. De-identified clinical and molecular data were compiled from The Cancer Genome Atlas (TCGA-AML), the Beat AML Master Trial, Gene Expression Omnibus (GEO) datasets GSE12417 and GSE6891, and the TARGET pediatric AML program^{20,21,28–32}. Ethics approval and informed consent for the original studies were obtained by the original investigators under their respective institutional review boards or ethics committees. Because the present study was a secondary analysis of publicly available de-identified data and no identifiable human participant information was accessed, additional informed consent was not required. The requirement for additional ethics approval was waived by the Ethics Committee of Fuzhou University Affiliated Provincial Hospital, waiver/reference number K2026-03-025.

The primary analysis focused on AML patients with venetoclax sensitivity/response labels, while other leukemia subtypes were used for contextual comparison and robustness assessment of BCL-2-dependent programs. These additional subtypes were not used to drive the primary venetoclax response modeling but were included to evaluate the generalizability of learned biological dependencies and to benchmark BCL-2-related signaling contexts across leukemia types. Cohort splitting and evaluation were performed at the patient level to prevent information leakage across training and test partitions.

Multi-omics feature compilation

For each patient, we integrated six data modalities:

1. Genomic mutations: Somatic single-nucleotide variants and indels were obtained from whole-exome or targeted sequencing calls, with a focus on recurrent leukemia drivers (FLT3, NPM1, TP53, DNMT3A, IDH1/2, KRAS/NRAS, RUNX1, BCR-ABL1).

2. Transcriptomics: RNA-seq or microarray data were normalized (log-transformed, TPM or RMA) and batch-corrected across cohorts using ComBat³³. Dimensionality reduction (UMAP)³⁴ was applied for exploratory visualization.

3. Copy number variation: Arm-level and focal CNVs were inferred from SNP arrays or sequencing and binarized for recurrent events (TP53 deletion, 5q-).

4. DNA methylation: Beta values from TCGA Illumina HumanMethylation450 data were averaged for tumor suppressor gene promoters.

5. Clinical data: Age, sex, cytogenetic risk, white blood cell count, treatment regimens, and outcomes (response, survival) were harmonized from source publications.

6. Drug sensitivity: Ex vivo cytotoxicity assays from BeatAML and other studies were used to classify venetoclax sensitivity (IC₅₀ below vs. above a predefined threshold) or, where available, clinical response (complete remission vs. refractory). For AML-specific analyses, response labels were derived from BeatAML ex vivo IC₅₀ data and harmonized with available clinical remission annotations to ensure consistency across datasets.

7. Response and resistance definition: Venetoclax non-response was treated as operational drug resistance and defined by either (i) IC₅₀ above a cohort-normalized threshold (upper quartile of ex vivo sensitivity distributions) or (ii) refractory/failed remission where clinical response data were available. To harmonize ex vivo and clinical labels, IC₅₀-based classifications were aligned with response categories by mapping low IC₅₀ values to responder status and high IC₅₀ values to non-response, ensuring consistency across datasets. Responder status was defined conversely by IC₅₀ below the threshold or achievement of complete remission.

Data were stored in a relational schema indexed by patient ID. Features with >20% missingness were excluded; continuous variables were z-scored. The final dataset included >80% coverage for genomic/transcriptomic data and ~60% for multi-modality profiles.

Model architecture and training

A deep feed-forward neural network was implemented to predict patient-specific drug response probabilities. Inputs comprised concatenated feature vectors of 512 features per patient after preprocessing and feature selection, including mutations (one-hot encoded), expression of differentially expressed and apoptosis-related genes, CNVs, and key clinical covariates. The network contained fully connected layers with ReLU activation, dropout regularization, and a sigmoid output for probability estimation. Model training used the Adam optimizer (learning rate = 1 × 10⁻³) with binary cross-entropy loss. Hyperparameters (number of layers, neurons per layer, dropout rate) were optimized via 10-fold cross-validation to maximize AUROC. Training was performed using a batch size of 32 with L2 regularization applied to network weights to reduce overfitting. Early stopping based on validation loss was implemented with a patience of 10 epochs. The dataset was split at the patient level into training (70%), validation (15%), and test (15%) sets. A fixed random seed (seed = 42) was used to ensure reproducibility of model training and evaluation. To address class imbalance, class weights were applied, and Gaussian noise augmentation was used for expression data.

To support clinical decision use, predicted probabilities were calibrated using post-hoc calibration on the validation folds (Platt scaling), and calibration performance was evaluated using Brier score and expected calibration error (ECE), and assessed with calibration curves and Brier score. Generalization was further checked by evaluating performance across independent cohorts and by reporting confidence intervals via bootstrapping.

Validation and performance metrics

Model performance was evaluated on a 30% held-out internal test set and on independent cohorts (TCGA-AML, Beat AML validation set, TARGET-AML). Primary metrics included AUROC, area under the precision-recall curve (PR-AUC), accuracy, and calibration (Brier score, reliability plots). For samples with quantitative drug sensitivity data, Spearman correlation between predicted probabilities and observed IC₅₀ values was computed. Survival analyses were performed by stratifying patients into predicted responder vs. non-responder groups and comparing overall survival using Kaplan-Meier and log-rank tests; multivariable Cox models assessed independence from known prognostic covariates.

Drug response and multi-drug design

Drug response data (IC₅₀, AUC) for venetoclax, FLT3 inhibitors, IDH inhibitors, and CDK9/MCL1 inhibitors were harmonized across cohorts using log-transformation and quantile normalization. Venetoclax sensitivity was binarized by IC₅₀ threshold or clinical remission status. For multi-drug design, we built a drug-target interaction matrix from DrugBank³⁵, ChEMBL³⁶, and literature, mapping each drug to its key targets and pathways. A network-based synergy scoring approach was used to simulate dual inhibition of resistance nodes (BCL-2 + MCL-1, BCL-2 + FLT3) in silico. Predicted probabilities under single vs. dual inhibition were compared to generate patient-specific synergy scores. Dual inhibition was simulated in silico by jointly perturbing feature representations corresponding to the targeted pathways (e.g., downregulation of MCL1 expression and/or FLT3 signaling activity) within the input feature space, followed by re-evaluation of model-predicted response probabilities. The resulting change in predicted response relative to single-target perturbation was used as a proxy for combination benefit, formalized as a synergy score defined by the increase in predicted response probability under dual-target perturbation compared with the best-performing single-target condition. To account for model uncertainty, confidence intervals for synergy scores were estimated via bootstrap resampling of input features and model predictions. This framework enabled rational prioritization of multitarget therapeutic strategies. Toxicity constraints were incorporated by penalizing combinations involving targets associated with known hematologic or systemic toxicity, based on curated drug safety annotations, ensuring prioritization of clinically feasible combinations.

Model interpretation

To elucidate key drivers of prediction, SHapley Additive exPlanations (SHAP)³⁷ were computed per patient and per feature. Mean absolute SHAP values ranked the global importance of predictors, and dependency plots were generated for highly influential variables (BCL2, MCL1, TP53 mutation). SHAP waterfall plots illustrated patient-specific feature contributions, enabling transparent interpretation of model-derived therapeutic recommendations.

Network and combination therapy analysis

The situational predictive features were projected to known protein-protein and signaling networks over STRING³⁸ and curated literature. Community detection using the Louvain modularity optimization algorithm³⁹ identified functional modules such as apoptosis regulation, RAS/MAPK signaling, epigenetic control, and immune checkpoint regulation. The degree of pathway perturbation was simulated in silico by performing therapeutic interventions (MCL-1 inhibition, FLT3 blockade) by modifying corresponding feature values and re-evaluating response probabilities, enabling rational prioritization of combination treatment strategies.

Molecular docking

Molecular docking was performed using AutoDock Vina⁴⁰ to evaluate interactions between venetoclax and model-prioritized targets (e.g., BCL-2, MCL-1, IDH1). Protein structures were obtained from the Protein Data Bank⁴¹ and prepared by protonation, removal of non-essential heteroatoms, and energy minimization. Docking grids were centered on known binding pockets, with an exhaustiveness parameter of 8 and up to 10 binding poses generated per ligand. Docked conformations were ranked by binding affinity and consistency of key interactions. Validation was performed by re-docking co-crystallized ligands into their respective binding sites to confirm pose accuracy. Interaction profiles, including hydrogen bonds and hydrophobic contacts, were analyzed using PyMOL to support structural interpretation of predicted drug-target interactions.

Statistical analysis

All statistical tests were two-sided unless otherwise specified. Student’s t-test or Wilcoxon rank-sum test were used to compare continuous variables, while the chi-square test or Fisher’s exact test were used for categorical variables. Kaplan–Meier analysis and Cox proportional hazards regression were performed for survival analyses, with hazard ratios reported along with 95% confidence intervals. Confidence intervals for model performance metrics were estimated using bootstrapping (1000 iterations). Statistical significance was defined as p < 0.05. Predictive modeling was reported in accordance with TRIPOD recommendations⁴², and prognostic analyses were described following REMARK guidelines⁴³.

Author contributions

C.W. and Z.Y. conceived and designed the study and contributed equally as co-first authors. C.W., Z.Y. and S.C. developed the study methodology and analytical framework. S.C., Y.H., Xin Chen (Author 5), and S.Q. curated the data and performed bioinformatic and statistical analyses. S.J. contributed to network modeling, structural docking analysis, and interpretation of drug-target interactions. C.W. and Z.Y. drafted the original manuscript. S.C., Y.H., X.C. (Author 7), S.Q. and S.J. critically reviewed and edited the manuscript. C.W. supervised the project and acquired funding. All authors read and approved the final manuscript.

Data availability

All datasets used in this study are publicly available from TCGA, BeatAML, GEO (GSE12417, GSE6891), and TARGET AML repositories. Processed data matrices and model-derived outputs supporting the findings of this study are available from the corresponding author upon reasonable request.

Code availability

The code used for model development and analysis is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.