Scanning electron microscopy and machine learning reveal heterogeneity in capsular morphotypes of the human pathogen Cryptococcus spp.

Abstract

Phenotypic heterogeneity is an important trait for the development and survival of many microorganisms including the yeast Cryptococcus spp., a deadly pathogen spread worldwide. Here, we have applied scanning electron microscopy (SEM) to defne four Cryptococcus spp. capsule morphotypes, namely Regular, Spiky, Bald, and Phantom. These morphotypes were persistently observed in varying proportions among yeast isolates. To assess the distribution of such morphotypes we implemented an automated pipeline capable of (1) identifying potentially cell-associated objects in the SEM-derived images; (2) computing object-level features; and (3) classifying these objects into their corresponding classes. The machine learning approach used a Random Forest (RF) classifer whose overall accuracy reached 85% on the test dataset, with per-class specifcity above 90%, and sensitivity between 66 and 94%. Additionally, the RF model indicates that structural and texture features, e.g., object area, eccentricity, and contrast, are most relevant for classifcation. The RF results agree with the observed variation in these features, consistently also with visual inspection of SEM images. Finally, our work introduces morphological variants of Cryptococcus spp. capsule. These can be promptly identifed and characterized using computational models so that future work may unveil morphological associations with yeast virulence.