Osteosarcoma microenvironment composition is a major feature to identify hard-to-treat patient tumors at diagnosis. Using machine learning to capture this microenvironment from bulk RNA-seq, we define the biological pathway and potential actionable targets associated with these tumors and validate a minimal set of biomarkers to detect hard-to-treat patients by RNA-seq and Nanostring in 178 additional tumors. In line with these results and to improve our understanding of the crosstalk between microenvironment and tumor cells leading to disease relapse, we initiated a spatio-temporal investigation at near single-cell level of Osteosarcoma in PDX and patients tumors.