Deep learning-based analysis reveals patient-level cancer therapy trajectories using single-cell PBMC chromatin images

Abstract Peripheral blood mononuclear cells (PBMCs) profiled in different data modalities have emerged as diagnostic and prognostic biomarkers for cancer. Here we demonstrate that PBMC chromatin images contain sufficient information to track patients’ trajectories during and after radiation therapy. We collected images from 150 patients across a variety of cancer types including Central Nervous System (CNS) cancer, Head & Neck cancer, Chordoma and Chondrosarcoma, Sarcoma, and Lymphoma, and adapted a multiple instance learning-based framework, which effectively enhanced the patient-level classification and interpretation of images from diverse samples compared to simpler cell-level neural network-based classifiers. We found that patients whose PBMCs returned to a state more similar to healthy after the therapy had more favorable outcomes. Moreover, our framework could indicate whether Head & Neck cancer patients’ PBMCs would likely return to a state similar to healthy after radiation therapy using solely chromatin images from before the treatment. These findings motivate further study of PBMC chromatin images to be used as a non-invasive, easily obtained, and inexpensive biomarker for cancer detection, treatment monitoring, and treatment response prediction.