Label-free Machine Learning Prediction of Chemotherapy on Tumor Spheroids using a Microfluidics Droplet Platform

Abstract

To rapidly evaluate the effects of anticancer treatments in 3D models, an integrated approach is proposed, combining a droplet-based microfluidic platform for spheroid formation and single-spheroid chemotherapy application, label-free morphological analysis, and machine learning to assess treatment response. Morphological features of spheroids, such as size and color intensity, are extracted and selected using the Multivariate Information-based Inductive Causation (MIIC) algorithm, and used to train a neural network for spheroid classification into viability classes, derived from metabolic assays performed within the same platform as a benchmark. The model was tested on Ewing sarcoma cell line and patient-derived xenograft (PDX) cells, demonstrating robust performance across datasets. It accurately predicts spheroid viability, used to generate dose-response curves and to determine half-maximal inhibitory concentration (IC50) values comparable to traditional biochemical assays. Notably, a model trained on cell line spheroids successfully classifies PDX spheroids, highlighting its adaptability. Compared to convolutional neural network-based approaches, this method works with smaller training datasets and provides greater interpretability by identifying key morphological features. The droplet platform further reduces cell requirements, while single spheroid confinement enhances classification quality. Overall, this label-free experimental and analytical platform is confirmed as a scalable, efficient, and dynamic tool for drug screening.