Leveraging AI to automate detection and quantification of extrachromosomal DNA to decode drug responses

Abstract

Introduction: Traditional drug discovery efforts primarily target rapid, reversible protein-mediated adaptations to counteract cancer cell resistance. However, cancer cells also utilize DNA-based strategies, often perceived as slow, irreversible changes like point mutations or drug-resistant clone selection. Extrachromosomal DNA (ecDNA), in contrast, represents a rapid, reversible, and predictable DNA alteration critical for cancer's adaptive response.

Methods: In this study, we developed a novel post-processing pipeline for automated detection and quantification of ecDNA in metaphase Fluorescence in situ Hybridization (FISH) images, leveraging the Microscopy Image Analyzer (MIA) tool. This pipeline is tailored to monitor ecDNA dynamics during drug treatment.

Results: Our approach effectively quantified ecDNA changes, providing a robust framework for analyzing the adaptive responses of cancer cells under therapeutic pressure.

Discussion: The pipeline not only serves as a valuable resource for automating ecDNA detection in metaphase FISH images but also highlights the role of ecDNA in facilitating swift and reversible adaptation to epigenetic remodeling agents such as JQ1.

Keywords: computer vision; cytogenetics; deep neural networks; double minute chromosomes; ecDNA; extrachromosomal DNA; fluorescence in situ hybridization; machine learning.

FIGURE 1

Scaling Annotations of ecDNA in Metaphase FISH images using AI (A). Microscopy methods, such as scanning electron microscopy (SEM) and DNA Fluorescence in situ Hybridization (FISH) are gold standard approaches for visualizing ecDNA. DNA FISH uses fluorescent probes that bind to specific DNA sequences that indicate which regions of the genome are amplified by ecDNA. Multiple species can be observed depending on which fluorescent probe is detected. (B) Under stress, ecDNAs can re-enter chromosomes to form homogeneous staining regions (HSRs). (C) FISH is the gold standard method for detecting and analyzing ecDNA in cell nuclei. It consists of several steps which include cell cycle synchronization, fixation, and hybridization. With advances in AI, time-intensive manual labeling and counting of ecDNA is accelerated and can be scaled up. (D) Asymmetric division of ecDNA molecules into daughter cells during replication and division leads to a heterogeneous population of cells with various ecDNA counts and species. (E) Convolutional Neural Networks is a computer vision approach that is well suited to image segmentation. (F) A computer vision approach applied to detecting and counting ecDNAs in metaphase FISH images. (G) Using ground truth (i.e., manually labeled metaphase FISH images), we can assess the accuracy of the computer vision method at detecting and counting ecDNAs per cell nucleus. (H) Region of Interest (ROI), defined as the human-annotated boundary representing the nuclear spread of a single cell. This is the area within the image where ecDNA is predicted and counted.

FIGURE 2

AI Models and Assessment of their Accuracies (A). We created different models to test the highest possible accuracy that we could expect our model to reach. (B) The breakdown of images used to train the Convolutional Neural Networks (CNNs) by cell line model system. (C) CNN-predicted chromosomes, ecDNA, and in-tact nuclei from a metaphase FISH image. (D) A zoomed-in image of ecDNA near chromosomes, highlighting the challenges in detecting individual ecDNAs when they cluster close together. (E) Accuracy metrics for the full (n = 1,164) model, showing the distribution of error (the difference in ecDNAs predicted versus manually counted) across images. (F) Existing algorithms struggle to predict ecDNAs in close proximity to chromosomes. (G) Annotated results from MIA, showing instances of true positives/negatives and false positives/negatives as well as cases where MIA finds ecDNA that were not correctly annotated. (H) Contour detection of conjoined segmentations. The figure illustrates examples of closely clustered ecDNA entities that are identified as a single object by the algorithm. Contour detection enables the separation of these conjoined objects into distinct segments, improving the accuracy of ecDNA boundary identification and differentiation.

FIGURE 3

Scaling Data Analytics to Probe ecDNA-Mediated Drug Response Mechanisms (A). Metaphase FISH image with dual probe labeling of NCI-H2170 cells that have four or more ecDNA species present in most of their cell nuclei. (B) Having multiple species of ecDNA present across cell nuclei can generate widespread cell-to-cell heterogeneity. As ecDNA segregates unevenly into daughter cells, each of these species will segregate unevenly. This generates extreme population genetic heterogeneity in terms of copy number variation differences across cells. This heterogeneity could lead to accelerated adaptation and drug resistance. (C) JQ1 is a molecule that was recently discovered to impact ecDNA higher-order clustering. It inhibits BRD4, impacting MYC transcriptional activity. Because MYC is often amplified on ecDNA, JQ1 may globally influence cells that harbor ecDNA-based MYC amplications. (D) Experimental results and AI-predicted ecDNA counts for MIA and ecSeg comparing ecDNA counts in control (untreated) and drug-treated (JQ1) NCI-H2170 cells. Significance was determined using a Mann-Whitney-Wilcoxon test with p-values < 0.0001 (****) and < 0.001 (***). (E) Under treatment, cells eliminate their ecDNA rapidly and reintegrate them into chromosomes, forming homogeneous staining regions (HSRs). (F) Drug treatment increases the number of multimerized ecDNAs that are seen in metaphase FISH images, which may be a precursor to chromosomal reintegration. (G) FISH data showing different structural views of ecDNA during drug treatment, including double minutes, multimerized ecDNA or “hubs” and HSRs. (H) A model predicting ecDNA reintegration rate during drug treatment. (I) The fraction of cells that were observed to have HSRs during metaphase FISH image analysis. (J) Cells with lower ecDNA counts are more likely to be observed to have HSRs compared to cells with high ecDNA levels. Significance was determined with a Chi-Square test with p-value < 0.05 (*). Similarly, cells with lower ecDNA levels are more likely to have multimerized structures compared to cells with high levels of ecDNA.