Machine learning based methodology to identify cell shape phenotypes associated with microenvironmental cues.

Cell morphology has been identified as a potential indicator of stem cell response to biomaterials. However, determination of cell shape phenotype in biomaterials is complicated by heterogeneous cell populations, microenvironment heterogeneity, and multi-parametric definitions of cell morphology. To associate cell morphology with cell-material interactions, we developed a shape phenotyping framework based on support vector machines. A feature selection procedure was implemented to select the most significant combination of cell shape metrics to build classifiers with both accuracy and stability to identify and predict microenvironment-driven morphological differences in heterogeneous cell populations. The analysis was conducted at a multi-cell level, where a "supercell" method used average shape measurements of small groups of single cells to account for heterogeneous populations and microenvironment. A subsampling validation algorithm revealed the range of supercell sizes and sample sizes needed for classifier stability and generalization capability. As an example, the responses of human bone marrow stromal cells (hBMSCs) to fibrous vs flat microenvironments were compared on day 1. Our analysis showed that 57 cells (grouped into supercells of size 4) are the minimum needed for phenotyping. The analysis identified that a combination of minor axis length, solidity, and mean negative curvature were the strongest early shape-based indicator of hBMSCs response to fibrous microenvironment.