Mechanistic Deep Learning Framework on Cell Traits Derived from Single-Cell Mass Cytometry Data

Germline genetic variations can alter cellular differentiation, potentially impacting the response of immune cells to inflammatory challenges. Current variant- and gene-based methods in mouse and human models have established associations with disease phenotypes; however, the underlying mechanisms at the cellular level are less well-understood. Immunophenotyping by multi-parameter flow cytometry, and more recently mass cytometry, has allowed high-resolution identification and characterization of hematopoietic cells. The obtained characterization yields increased dimensionality; however, conventional analysis workflows have been inefficient, incomplete, or unreliable. In this work, we develop a comprehensive machine learning framework – MDL4Cyto – that is tailored to the analysis of mass cytometry data, incorporating statistical, unsupervised learning, and supervised learning models. The statistical modeling can be used to illuminate cell fate decision and cell-type dynamics. The unsupervised learning models along with complementary marker enrichment analyses highlight genetic perturbations that are significantly associated with alterations in cell populations in the hematopoietic system. Furthermore, our supervised learning models, including deep learning and tree-based algorithms, address the bottleneck to data pre-processing that characterizes conventional workflows and generate inferences (e.g., on marker/cell-type interactions) from raw experimental characterization. Notably, we reveal a close relationship among network design, prediction performance, and the underlying biological context. We show that the network architecture extracted from the differentiation cascade of the investigated biological system yields enhanced prediction performance. The presented methodology will enable new insights into hematopoietic differentiation at baseline and following perturbation. Highlights Analysis pipeline on mass cytometry data with high-performance implementation of statistical, unsupervised learning, and supervised learning models Concordance of machine learning results with biological contexts Biologically-informed neural network designs enhance prediction performance Graphical Abstract