6525 Background: AML is a heterogeneous hematological malignancy with poor prognosis. Several treatments are approved for AML, but clinical trials have shown that current stratification approaches to determine patients’ eligibility produce false positives (treated patients that fail to respond) and negatives (patients not treated but could have responded). Venetoclax + azacitidine (VA) treatment is currently reserved for unfit patients, with younger patients stratified based on their FLT3status, and treated with either intensive chemotherapy (IC), or IC plus midostaurin (MIC). Here, we used phosphoproteomics to build a signature and algorithm that accurately predict which of these approved therapies may be more efficacious for a given patient. Methods: Routine bone marrow and peripheral blood diagnosis samples (s, n=182) were collected across the UK, Austria, Canada and the USA from 138 patients (p) subsequently treated with MIC (n=44/64 p/s), VA (n=40/48 p/s) or IC (n=54/70 p/s). Patients were grouped into Good Responders (GR) and Poor Responders (PR) based on treatment response. For VA, patients that achieved complete remission (CR) were considered GR, while refractory patients were considered PR. For MIC and IC, we considered patients that achieved CR without relapse within 6 months as GR, and those refractory or relapsed within 6 months as PR. Samples were processed for mass spectrometry-based phosphoproteomics. Phosphopeptide abundance data, generated with in-house PiQuant software, was used to identify phosphopeptides that distinguish GR and PR groups in each cohort. Statistical models based on these features were assessed via cross-validation. Results: We compared phosphoproteomes of 182 diagnosis samples from 138 AML patients, from three treatment cohorts (treated with IC, MIC or VA), with each cohort stratified by patients’ response to their respective treatment. Enrichment analysis in each cohort identified several phosphopeptides specific to one of the responder groups, mapping both to proteins with known roles in AML biology (e.g. DNMT3A or RUNX1) and proteins not yet implicated. Next, using machine learning, we identified phosphopeptides that could distinguish between PR and GR, and trained drug response prediction models based on the abundance of these phosphopeptides. In cross-validation, each model stratified patients with log rank p<0.001, HR<0.1 and more than 90% accuracy, greatly outperforming all currently-used stratification methods for first-line AML therapies. Conclusions: We built a suite of predictive models that accurately predict patient response to first-line AML treatment using phosphoproteomic data from routine diagnosis samples. Following validation in independent patient cohorts, this tool will be developed into a single test that predicts treatment response for AML patients, thus addressing an unmet clinical need in this disease.