Early Prediction of Single-Cell Derived Sphere Formation Rate Using Convolutional Neural Network Image Analysis.

Abstract

Functional identification of cancer stem-like cells (CSCs) is an established method to identify and study this cancer subpopulation critical for cancer progression and metastasis. The method is based on the unique capability of single CSCs to survive and grow to tumorspheres in harsh suspension culture environment. Recent advances in microfluidic technology have enabled isolating and culturing thousands of single cells on a chip. However, tumorsphere assay takes a relatively long period of time, limiting the throughput of this assay. In this work, we incorporated machine learning with single-cell analysis to expedite tumorsphere assay. We collected 1,710 single-cell events as the database and trained a convolutional neural network model that predicts whether a single cell could grow to a tumorsphere on Day 14 based on its Day 4 image. With this future-telling model, we precisely estimated the sphere formation rate of SUM159 breast cancer cells to be 17.8% based on Day 4 images. The estimation was close to the ground truth of 17.6% on Day 14. The preliminary work demonstrates not only the feasibility to significantly accelerate tumorsphere assay but also a synergistic combination between single-cell analysis with machine learning, which can be applied to many other biomedical applications.