CLIP-AML: Contrastive Learning Framework for AML treatment response prediction

Background: Acute myeloid leukemia (AML) is a biologically heterogeneous cancer of myeloid cells in which patients exhibit widely varying responses, making reliable treatment selection a persistent challenge. Predictive models that can leverage multi-omic patient profiles to guide individualized therapy selection are therefore urgently needed. The BeatAML cohort collected and characterized patient samples over 10 years, integrating ex-vivo drug sensitivity, clinical annotations, DNA and RNA sequencing.Methods: We propose CLIP-AML which leverages the BeatAML cohort to devise a contrastive learning-based deep learning framework for predicting ex vivo drug response in AML patients. Our framework engineers vector representations for drugs, patient genomic profiles, cell state, and pathway activities. The novel contrastive learning (CLIP)-based objective jointly learns to align embeddings between patient multi-omic profiles with drug representations bringing drugs with higher sensitivity closer to patient profiles while pushing away resistant drugs in the embedding space.Results: The optimal CLIP-AML model achieves a validation mean absolute error (MAE) of 34.11 ± 1.36 and Pearson correlation (r_pc) of 0.679 ± 0.028, with an MAE of 39.170 and r_pc of 0.657 on unseen test set, competitive against end-to-end deep learning models by 2- across multiple evaluation metrics. This demonstrates the generalization capability of CLIP-AML for out-of-box patient drug sensitivity prediction.