A Network-Guided Modular Framework for drug response prediction in acute myeloid leukemia

Drug response prediction in acute myeloid leukemia (AML) is challenged by sample heterogeneity and high-dimensional RNA sequencing profiles. We develop NGM-AML, a modular model predicting ex vivo drug response from BeatAML2 RNA-seq and sensitivity data. After filtering, 306 Waves 1+2 samples (28055 records) serve as training and 173 Waves 3+4 samples (16123 records) as the test set across 111 drugs. Drug targets and AML prior genes are mapped to a PPI network; random walk with restart and community detection construct 45 modules. Per drug, module scores are partitioned into sensitivity and resistance components by their association with the area under the dose-response curve. The results show that NGM-AML achieves mean Pearson and Spearman correlations of 0.324 and 0.326 across drugs, with an MAE of 37.396. Pooling all test records yields a Pearson correlation of 0.703 between predicted and observed AUC. For representative drugs, Pearson correlations reach 0.780 for Venetoclax and 0.631 for Trametinib. Within patients, median Spearman correlation and NDCG@5 are 0.75 and 0.96 for drug ranking. Runtime decreases from 87125.4 s for the raw RNA-seq model to 1102.5 s for NGM-AML. Enriched processes include extracellular matrix adhesion, integrin signaling, and RTK/MAPK pathways, consistent with known AML survival and drug-resistance mechanisms.