Abstract
Drug combinations are essential to modern medicine, but their discovery remains slow and inefficient as experimental complexity expands rapidly with each additional drug tested. Although modern liquid handling systems enable complex and highly customizable experimental designs, a lack of strategies integrating these technologies with combination-specific analytical methods has limited throughput. Here we introduce Combocat, an open-source and streamlined framework that combines acoustic liquid handling protocols with machine learning-based inference to achieve ultrahigh-throughput drug combination screening. Using Combocat, we generate a reference dataset of over 800 unique combinations in a dense 10 × 10 matrix format across multiple cell types, and use this to train a predictive model that accurately infers drug combination effects from sparse data, drastically reducing the number of experimental measurements required. As proof of concept, we screened 9,045 combinations in a neuroblastoma cell line-the largest number of combinations tested in a single cell line to date-achieved using minimal resources. By integrating advanced drug dispensing technologies with predictive computational modeling, Combocat provides a scalable solution to accelerate the discovery of novel drug combinations.