Representing immune profiles with self-supervised set representation learning 4557

Abstract Description The immune system is a complex network of cells, orchestrating coordinated responses throughout the human body over a person’s lifetime. Cytometry enables profiling of this network, but current analytical approaches focus on enumerating and phenotyping immune cells, failing to quantify the system as a whole. We present MAESTRO, an AI model designed to generate vector representations of immune profiles from cytometry data. Unlike studies that learn cell-level representations, MAESTRO captures subject-level representations by learning from all cells in a sample. MAESTRO uses attention mechanisms to handle sets of variable number of cells and ensure permutation invariance, coupled with a self-distillation framework to handle the large number of cells. Our model operates without labeled data for training, taking .fcs files as input. It is capable of reconstructing immune profiles (cells) even when 90% are hidden as its training objective. After training, MAESTRO’s learned immune representation of 1,514 samples across 11 phenotypes enables robust prediction of diagnosis, age, gated cell-type distribution, response, and treatment. Our model demonstrates superior performance compared to gating, clustering, and machine learning methods for set representation, serving as a complement when gating alone is insufficient. MAESTRO’s holistic representation of the immune system is essential for predicting outcomes, identifying health trajectories, and advancing precision medicine. Funding Sources This work was partially supported by NIH AI082630, QuantumLeap Healthcare Collaborative fund, the University of Pennsylvania Perelman School of Medicine COVID Fund. Topic Categories Computational and Systems Immunology (COMP)