Accelerated Pheno-Tree (APT) for large-scale, label-free of image-based single-cell analysis

Abstract

Recent advances in imaging cytometry enable high-resolution analysis of single-cell phenotypes (both physical and biochemical) at high throughput with the overall aim of revealing the phenotypic variability within an enormous and heterogeneous population of cells. However, analysis of large-scale high dimensional single-cell image data is computationally intensive and soon becomes unscaleable from both a memory and run time perspective. To address this challenge, we develop Accelerated Pheno-Tree (APT) – an unsupervised clustering algorithm tailored for analyzing large-scale high dimensional single-cell image-based data. As a proof-of-concept demonstration, we adopt APT in time-stretch quantitative phase imaging (TS-QPI) – an ultrahigh-throughput label-free imaging technique that allows large-scale single-cell biophysical phenotyping. APT allows fast unbiased clustering and visualization of high-dimensional datasets of above 1 million single cell TS-QPI - bypassing the need for prior knowledge of the data as well as data down-sampling which are common in the existing clustering methods. Integrating two key computational steps, i.e. accelerated non-linear dimension reduction (LargeVis) of the TS-QPI data followed by the graph-based and data-driven agglomerative clustering (based on accelerated minimum spanning tree construction), APT successfully distinguishes multiple cell types (e.g. 7 lung cancer cell lines, and sub-types of PBMC cells) entirely based on their intrinsic biophysical phenotypes (up to 30 dimensions) quantified from label-free TS-QPI (total cell count: 1.1 million cells). We anticipate that APT could be particularly useful in ultralarge-scale single-cell analysis and facilitates exploration of the heterogeneity within cell populations based on single-cell biophysical features with high accuracy.