Profiling DNA damage in 3D Histology Samples

Abstract

The morphology of individual cells can reveal much about the underlying states and mechanisms in biology. In tumor environments, the interplay among different cell morphologies in local neighborhoods can further improve this characterization. In this paper, we present an approach based on representation learning to capture similarities and subtle differences in cells positive for γH2AX, a common marker for DNA damage. We demonstrate that texture representations using GLCM and VAE-GAN enable profiling of cells in both singular and local neighborhood contexts. Additionally, we investigate a possible quantification of immune and DNA damage response interplay by enumerating CD8+ and γH2AX+ on different scales. Using our profiling approach, regions in treated tissues can be differentiated from control tissue regions, demonstrating its potential in aiding quantitative measurements of DNA damage and repair in tumor contexts.

Keywords: 3D histology; DNA damage; cell morphology; representation learning.

Fig. 1

A 3D image volume (A) from the dataset used in this work (CD8 in red, γH2AX in yellow, pan protein in green, DAPI in blue), z-slices of individual channels (B-D) and nuclei segmentation (E-G). Using the DAPI channel (B), Cellpose was used to generate segmentation of all nuclei (E). Using this segmentation and threshold masks on the γH2AX channel (C) and the CD8 channel (D), γH2AX+ cells and CD8+ are identified (F,G).

Fig. 2

The VAE-GAN architecture (A) used to construct the manifold of DNA damage in cells. The network is composed of three components: encoder, generator, and discriminator. The encoder (B) maps input images to 16-dimensional Gaussian distributions with diagonal covariance. The generator (C) produces a reconstruction from sampled points in the latent space. The discriminator (D) forces the generator to output images as similar to the input as possible. Latent encoding of the image volumes are then clustered to form pseudo-classes of DNA damage subtypes (E).

Fig. 3

Clustering on GLCM representations (A) and VAE-GAN encodings (B). Heatmaps, using pseudo-classes from GLCM (C) and VAE-GAN(D), show differences in γH2AX profiles in neighborhoods in control and treated samples. Proximity analysis shows a higher number of γH2AX+ cells near CD8+ cells in treated samples (E).