Cell response heterogeneity upon treatment is a main obstacle in preclinical development of efficacious cancer drugs, due to the emergence of drug-tolerant cells. We have previously developed a single-cell workflow, Fate-Seq (1,2), to profile drug-tolerant persisters. Fate-Seq is based on the prediction of the drug response of each cells that are individually profile at the molecular level.To achieve this goal, Fate-Seq couples 3 single-cell techniques: first the prediction of the cell response phenotype (resistant or sensitive) for clonal cancer cells treated with a chosen drug, then the isolation of the predicted resistant cells from the predicted sensitive ones by laser-capture, and finally the RNA sequencing of each single-cell (sc-RNA-seq). These sc-RNA-seq dataset are then analyzed using random walks with restart, to prioritize the genes according to the drug-sensitive state of each cell and to identify the genes causing for cell drug-resistance (as opposed to the genes associated with drug-resistance, 3). To automatize and increase the prediction throughput, we present 3 new developments in our workflow using machine learning models to classify cell drug response, from the cell signaling dynamics observed with fate-seq, and to determine the molecular factors defining the drug efficacy of the cell type tested. Theses molecular factors represent good candidates to be targeted during a co-treatment, in combination with the first drug analyzed with our pipeline. We then introduce our eDRUGs (early Drug Response UpGraded) classifier, that combines mechanistic modeling of apoptosis (cell death) through cell signaling pathway, and machine learning classification models to predict cell drug response within an hour, using the fluorescent time-trajectories as input. This new method is twice as accurate as our previous prediction method (4). Finally, we will also propose a novel analysis method of sc-RNA-seq data obtained with Fate-Seq. This method consists in training binary classifiers on the scRNAseq expression data obtained .from the pipeline, using a range of models and explainable AI techniques such as DeepLift (5), in addition to clustering techniques, to obtain attribution scores for each gene. These scores are expected to reveal a reduced gene set, more manageable for drug combinations design.