RANKOR: A Machine Learning Model for Drug Repurposing on Individual-Level Single Cell Data

<p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" class="first" dir="auto" id="d11262133e95">Precision oncology aspires to align cancer patients with therapies that most effectively target their individual molecular profiles. A major challenge in this field is the identification of drugs capable of reversing disease-specific transcriptional states in a cell-type–resolved manner. To address this, we present RANKOR, a machine learning framework that integrates patient-derived single-cell transcriptomic data with chemical structure representations to prioritize candidate drugs for repurposing. Unlike existing methods that depend on predefined gene sets, RANKOR leverages contrastive learning to jointly embed drug-induced gene expression profiles (from LINCS Phase 5 perturbation datasets) and chemical structures (encoded from SMILES via Morgan fingerprints) into a shared latent space. This approach enables generalization to unseen compounds while preserving mechanistic interpretability by clustering drugs with similar mechanisms of action (MoAs). <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d11262133e97">The model is trained to predict drug- or MoA-specific transcriptional shifts, and drug ranking is performed by assessing the similarity between patient disease-state signatures and predicted drug-induced expression reversals. Importantly, RANKOR supports explainability: gene-level feature attribution highlights molecular drivers of the ranking decisions, which we validate through gene set enrichment analyses that recover pathways consistent with known drug targets. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d11262133e99">We applied RANKOR to single-cell RNA sequencing (scRNA-seq) data from acute lymphoblastic leukemia (ALL) patients, capturing intratumoral heterogeneity and cell-type–specific responses. Comparative evaluation against recent repurposing pipeline, demonstrated that RANKOR more consistently prioritized leukemia-relevant candidates, including FDA-approved therapies for ALL and related hematological malignancies, as well as compounds targeting pathways central to leukemogenesis. Furthermore, the model’s ranking behavior was biologically grounded: combinations with weak or absent transcriptional responses were deprioritized, reflecting signal dependence. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d11262133e101">Our results highlight RANKOR’s potential to support clinically actionable drug repurposing in oncology, by offering (i) scalability to novel compounds, (ii) interpretability through mechanistic gene prioritization, and (iii) relevance at the single-cell resolution necessary for addressing cancer heterogeneity. This work provides a framework for precision medicine that bridges molecular data and chemical space, advancing the search for effective, personalized therapeutic strategies.