Digital analysis of pathology whole-slide images is fast becoming a game changer in cancer diagnosis and treatment. Specifically, deep learning methods have shown great potential to support pathology analysis, with recent studies identifying molecular traits that were not previously recognized in pathology H&E whole-slide images. Simultaneous to these developments, it is becoming increasingly evident that tumor heterogeneity is an important determinant of cancer prognosis and susceptibility to treatment, and should therefore play a role in the evolving practices of matching treatment protocols to patients. State of the art diagnostic procedures, however, do not provide automated methods for characterizing and/or quantifying tumor heterogeneity, certainly not in a spatial context. Further, existing methods for analyzing pathology whole-slide images from bulk measurements require many training samples and complex pipelines. Our work addresses these two challenges. First, we train deep learning models to spatially resolve bulk mRNA and miRNA expression levels on pathology whole-slide images (WSIs). Our models reach up to 0.95 AUC on held-out test sets from two cancer cohorts using a simple training pipeline and a small number of training samples. Using the inferred gene expression levels, we further develop a method to spatially characterize tumor heterogeneity. Specifically, we produce tumor molecular cartographies and heterogeneity maps of WSIs and formulate a heterogeneity index (HTI) that quantifies the level of heterogeneity within these maps. Applying our methods to breast and lung cancer slides, we show a significant statistical link between heterogeneity and survival. Our methods potentially open a new and accessible approach to investigating tumor heterogeneity and other spatial molecular properties and their link to clinical characteristics, including treatment susceptibility and survival.

Controlling off-target editing activity is one of the central challenges in making CRISPR technology accurate and applicable in medical practice. Current algorithms for analyzing off-target activity do not provide statistical quantification, are not sufficiently sensitive in separating signal from noise in experiments with low editing rates, and do not address the detection of translocations. Here we present CRISPECTOR, a software tool that supports the detection and quantification of on- and off-target genome-editing activity from NGS data using paired treatment/control CRISPR experiments. In particular, CRISPECTOR facilitates the statistical analysis of NGS data from multiplex-PCR comparative experiments to detect and quantify adverse translocation events. We validate the observed results and show independent evidence of the occurrence of translocations in human cell lines, after genome editing. Our methodology is based on a statistical model comparison approach leading to better false-negative rates in sites with weak yet significant off-target activity.

AbstractDigital analysis of pathology whole-slide images is fast becoming a game changer in cancer diagnosis and treatment. Specifically, deep learning methods have shown great potential to support pathology analysis, with recent studies identifying molecular traits that were not previously recognized on pathology H&E whole-slide images. Simultaneous to these developments, it is becoming increasingly evident that tumor heterogeneity is an important determinant of cancer prognosis and susceptibility to treatment, and should therefore play a role in the evolving practices of matching treatment protocols to patients. State of the art diagnostic procedures, however, do not provide scalable methods for characterizing and/or quantifying tumor heterogeneity, certainly not in a spatial context. In this paper, we present a scalable approach that accurately and automatically spatially resolves mRNA and miRNA expression levels on pathology whole-slide images. This is the first demonstration of this type of inference from H&E images. We use this method to produce tumor molecular cartographies and to characterize certain aspects of tumor spatial transcriptomics. Specifically, we develop a heterogeneity index (HTI), derived from the molecular cartographies. Applying our methods to breast and lung cancer slides, we show a significant statistical link between heterogeneity and survival. Our results highlight the value of automated analysis of pathology whole slide images. Our methods potentially open a new approach to investigating tumor heterogeneity and other spatial molecular properties and their link to clinical characteristics, including treatment susceptibility and survival.

DNA methylation has been extensively linked to alterations in gene expression, playing a key role in the manifestation of multiple diseases, most notably cancer. For this reason, researchers have long been measuring DNA methylation in living organisms. The relationship between methylation and expression, and between methylation in different genomic regions is of great theoretical interest from a molecular biology perspective. Therefore, several models have been suggested to support the prediction of methylation status in samples. These models, however, have two main limitations: (a) they heavily rely on partially measured methylation levels as input, somewhat defeating the object as one is required to collect measurements from the sample of interest before applying the model; and (b) they are largely based on human mediated feature engineering, thus preventing the model from unveiling its own representations. To address these limitations we used deep learning, with an attention mechanism, to produce a general model that predicts DNA methylation for a given sample in any CpG position based solely on the sample's gene expression profile and the sequence surrounding the CpG.We show that our model is capable of generalizing to a completely separate test set of CpG positions and subjects. Depending on gene-CpG proximity conditions, our model can attain a Spearman correlation of up to 0.8 and MAE of 0.14 for thousands of CpG sites in the test data. We also identify and analyze several motifs and genes that our model suggests may be linked to methylation activity, such as Nodal and Hand1. Moreover, our approach, and most notably the use of attention mechanisms, offers a novel framework with which to extract valuable insights from gene expression data when combined with sequence information. The code and trained models are available at: https://github.com/YakhiniGroup/Methylation

No abstract