PREDICT-AML: personalized response via drug interaction with cellular traits for acute myeloid leukemia

Acute myeloid leukemia (AML) is a heterogeneous hematologic malignancy with limited therapeutic options. The BeatAML cohort profiled specimens from \(805\) patients, integrating ex vivo drug sensitivity measurements across \(165\) drugs with clinical annotations, DNA and RNA sequencing. We present PREDICT-AML, a multi-omics driven machine learning framework that leverages this resource to generate generalizable, personalized drug response predictions from patients’ clinical, genomic, and transcriptomic profiles. PREDICT-AML employs four complementary drug representations: (a) physicochemical descriptors, (b) knowledge-distilled embeddings (KD-Embed), and chemical language model embeddings from (c) ChemBERTa and (d) MolFormer. These are combined with novel patient-specific oncogenic pathway proximity scores, cell-state enrichments, and genomic profiles. The tabular foundation model, TabPFN, was identified as the optimal architecture following systematic hyperparameter optimization via Optuna. TabPFN with knowledge-distilled embeddings (KD-Embed) achieves a Pearson correlation ( \(r_{pc}\) ) of \(0.690\) and MAE = \(37.712\) on the held-out test set. Random cross-validation overestimates performance by \(\approx13\%\) relative to patient-stratified cross-validation, the clinically appropriate benchmark. External validation on LeeAML ( \(r_{pc}=0.597\) ) and FIMM-AML ( \(|r_{pc}|=0.557\) ) confirms cross-institutional generalizability, outperforming ElasticNet ( \(r_{pc}\) = \(0.360\) ) and MDREAM (Spearman correlation, \(r_{sc}=0.680\) ) baselines. Ablation identifies gene expression as the dominant predictor; somatic mutations with cell-state enrichment alone match full multi-omics performance. SHAP interpretability reveals KD-Embed dimensions as the primary drug-identity signal, with monocyte-like enrichment, BCL6 expression, and immunogenic cell death pathway activity as recurrent multi-drug sensitivity determinants. PREDICT-AML provides an interpretable, externally validated tool for individualized therapeutic prioritization in AML.