Acute myeloid leukemia (AML) is a cancer of myeloid cells with limited treatment options. The BeatAML cohort collected and characterized samples of 805 patients over 10 years with the integration of ex vivo drug sensitivity, clinical annotations, and DNA and RNA sequencing. Through association analysis, strong cross-cohort concordance and features of drug response were previously identified. In this work, we propose PT-AML, which leverages the BeatAML cohort and their collection strategy to devise and validate machine learning techniques for predicting drug response of AML patients given their clinical, genomic and transcriptomic profiles. Our framework engineers vector representations for drugs, their relevant pathways, patient profiles and their cell states to accurately estimate drug response. The optimal PT-AML model attains a cross-validation mean-absolute-error (MAE) and spearman correlation (r) of \(21.02 \pm 1.5\) , \(0.85 \pm 0.02\) respectively and MAE of 39.04 and r of 0.66 on the unseen test set respectively highlighting the generalization capability of the model.