A self-supervised machine learning pipeline for extracting information from live cell images at multiple doses and timepoints

Abstract

Abstract Live cells are complex information-processing systems that continuously sense their environment and respond dynamically. However, conventional endpoint assays typically require fixation or cell destruction and fail to capture complex temporal changes. Live brightfield imaging offers a scalable, label-free solution that remains underutilized due to particularly low contrast, acute technical batch sensitivity, and the limited availability of robust computational methods for this modality. Leveraging recent self-supervised learning developments, we introduce Live Cell Dynamics (LCD), a novel end-to-end transformer-based pipeline, using novel plane-agnostic augmentation (treating different focal planes as views of the same state) and incorporating cross-batch sampling. LCD addresses brightfield modality challenges and extracts subtle dose- and time-dependent live cell states. Through systematic ablation we evaluate each self-supervised training innovation on a single cell line, measuring phenotypic activity (mean Average Precision) and Mechanism of Action (MoA) classification (F1-score), with 189 compounds in pre-training and 81 in holdout spanning ten MoAs. Our approach outperforms ablated baselines across all doses and timepoints for activity and MoA classification, enables compound polypharmacology detection from multi-dose/timepoint profiles, and supports unsupervised nuclei detection and counting. It leads to training foundation models from continuous live brightfield imaging to detect subtle live cell state changes, enabling scalable, cost-effective drug development.