CAP TODAY Pathology/Laboratory Medicine/Laboratory Management
Editors: Donna E. Hansel, MD, PhD, chair of pathology, Oregon Health and Science University, Portland; Richard D. Press, MD, PhD, professor and director of molecular pathology, OHSU; James Solomon, MD, PhD, assistant professor, Department of Pathology and Laboratory Medicine, Weill Cornell Medicine, New York; Sounak Gupta, MBBS, PhD, senior associate consultant, Mayo Clinic, Rochester, Minn.; Tauangtham Anekpuritanang, MD, molecular pathology fellow, Department of Pathology, OHSU; Hassan Ghani, MD, molecular genetic pathology fellow, Department of Pathology, OHSU; and Fei Yang, MD, assistant professor, Department of Pathology, OHSU.
DNA microscopy: optics-free spatio-genetic imaging by a standalone chemical reaction
September 2019—The pathologist’s ability to interpret the complex spatial organization within and between cells and intercellular matrices is the basic underlying principle of morphologic pathology. Even in the genomic era, molecular genetic information is not clinically useful without tissue context. Modern spatial capturing methods, either by low-fidelity light microscopy or high-fidelity electron microscopy, cannot concomitantly interrogate a nucleic acid sequence. Analogously, modern DNA sequencing technology cannot capture architectural detail from the nucleic acid “soup.” In situ hybridization is a current method that allows visualization of limited genomic information in the tissue context. Other crude methods, such as tissue microdissection, may also provide indirect low-fidelity spatial detail to complement the downstream molecular result. Another more modern approach is digital spatial profiling technology (NanoString), which can integrate visual data with genomic data embedded in composite hybridization-based images. To address this important technical gap, the authors developed and validated a novel DNA microscopy system. Instead of combining visual (imaging) and nucleic acid (chemical) information, they took advantage of the direct physico-chemical diffusive properties of in situ synthesized DNA molecules to directly create virtual images that informed sequence and spatial information. The underlying principle of this DNA microscope is that newly synthesized uniquely barcoded cDNA molecules derived in situ from unique sequence-specific RNA transcripts in the cell will diffuse from their initial in situ position to form a spherical diffusion cloud around their spatial point of origin. Multiple diffusion clouds originating from unique transcript molecules will interact with each other, and the intersecting zones will contain a mixture of molecules of different sequences from different sources. Higher counts of intersecting diffusion clouds of transcript A and transcript B would then infer a relatively closer spatial orientation of these two molecules. By analyzing the diffusion clouds originating from each transcript molecule, an algorithm was devised to calculate the relative point of origin of each sequence-specific diffusion cloud and construct a virtual image of the interrogated RNA molecules in either two- or three-dimensional space. Each diffusion cloud can be identified by unique, randomly generated, sequence-specific molecular identifiers. A major benefit of this DNA microscopy assay is that the reactions can be performed using standard (easily fabricated) laboratory equipment. The weakness is that, at the cellular level, these molecular diffusion clouds are small compared to the size of the cells. Therefore, the computational algorithm needs to deal with a lot of “empty space,” which is a manageable future software-improvement issue. In conclusion, the authors demonstrated the high sequencing accuracy (compared to the known reference sequence) of DNA microscopy and showed that the ultimate computer-generated virtual image of the cell-specific localization of specific targeted transcripts recapitulates the “real” in situ image taken by epifluorescence optical microscopy before the DNA microscopy reaction. Although this method is still in the early stages of development, it is an interesting technology that may become a useful tool for the pathologist of the future.
Weinstein JA, Regev A, Zhang F. DNA microscopy: optics-free spatio-genetic imaging by a stand-alone chemical reaction. Cell. 2019;178:229–241.
Correspondence: Dr. Joshua A. Weinstein at jwein@broadinstitute.org
Deep learning to predict microsatellite instability from histology in GI cancer
Artificial intelligence in medicine, particularly in the pattern-recognition specialties of diagnostic pathology and radiology, has been a hot topic in recent years. Several published reports claim that deep learning and neural network systems can outperform pathologists and radiologists in specific diagnostic tasks. The promise of artificial intelligence (AI) in medicine is the potential for a low-cost, rapid, accurate, universally available diagnostic modality that can improve patient care. Microsatellite instability (MSI) in colorectal cancer, for example, is an important biomarker for predicting a tumor’s response to immunotherapy. However, its identification has historically required specialized IHC or molecular methods that are expensive and not routinely available in many pathology labs. The authors hypothesized that AI can be used to algorithmically interpret routine H&E-stained gastrointestinal cancer histology slides to identify MSI solely based on morphologic features. The research team used standard methods to train their AI system for image-based identification of MSI. In particular, the team prepared a training set consisting of histologic images from tumors with known MSI status and used it to inform a learning neural network algorithm (Resnet18). The training set consisted of a large set of gastric adenocarcinoma, colorectal adenocarcinoma, and endometrial carcinoma (including H&E-stained, formalin-fixed, paraffin-embedded samples and cryostat sections) images from participating institutions and The Cancer Genome Atlas database. The slide images were tessellated into tiles, and each tile was scored based on the likelihood of the image being associated with an MSI-h (high) tumor. The trained predictive models were then used to predict the MSI status from a validating set of unknown images consisting of gastrointestinal cancer cases from Germany, the United States, and Japan. The AI models described predicted MSI status when the training set and validation cohort were from the same tumor type and patient ethnicity (area under the curve, 0.75–0.84). The current algorithm outperformed other analogous models for predicting molecular features from histology. However, the performance dropped when the validating set and the training set were from different tumor types or patient ethnicities, or both. Larger cohorts with greater diversity in tissue type and patient ethnicity may be required to further improve the performance of the AI models. The researchers also correlated the quantitative “MSIness” score to transcriptomic and immunohistochemical data and showed that algorithmic MSIness was correlated, as expected, to the lymphocyte gene-expression signature in gastric adenocarcinoma and the PD-L1 expression and IFN-gamma signature in colorectal adenocarcinoma. Although improving, AI is not ready to replace standard histopathology and molecular diagnostic methods. The obvious value of these prediction models will be their cost-effectiveness and universal applicability across heterogeneous health care settings. Molecular diagnostic and IHC testing for MSI, although the gold standard, are expensive and time-consuming. Because the performance of these algorithms inevitably will improve, pathologists should be prepared to adopt these technologies to stay relevant in the coming age of artificial intelligence.