Automated Analysis and Classification of Histological Tissue Features by Multi-Dimensional Microscopic Molecular Profiling

RESEARCH ARTICLE Automated Analysis and Classification of Histological Tissue Features by MultiDimensional Microscopic Molecular Profiling Daniel P. Riordan1,2*, Sushama Varma3, Robert B. West3, Patrick O. Brown1,2 1 Department of Biochemistry, Stanford University School of Medicine, Stanford, California, United States of America, 2 Howard Hughes Medical Institute, Stanford University School of Medicine, Stanford, California, United States of America, 3 Department of Pathology, Stanford University School of Medicine, Stanford, California, United States of America * Abstract OPEN ACCESS Citation: Riordan DP, Varma S, West RB, Brown PO (2015) Automated Analysis and Classification of Histological Tissue Features by Multi-Dimensional Microscopic Molecular Profiling. PLoS ONE 10(7): e0128975. doi:10.1371/journal.pone.0128975 Editor: Arrate Muñoz-Barrutia, Universidad Carlos III of Madrid, SPAIN Received: October 10, 2014 Accepted: May 1, 2015 Published: July 15, 2015 Copyright: © 2015 Riordan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: The full set of MMMP images generated will be available online upon publication at the Stanford Tissue Microarray Database (http://tma.im). Funding: This work was supported by National Heart, Lung, and Blood Institute (NHLBI) U01HL099995 Progenitor Cell Biology Consortium Grant (POB, DPR, SV, RBW) and by the Howard Hughes Medical Institute (POB and DPR). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Characterization of the molecular attributes and spatial arrangements of cells and features within complex human tissues provides a critical basis for understanding processes involved in development and disease. Moreover, the ability to automate steps in the analysis and interpretation of histological images that currently require manual inspection by pathologists could revolutionize medical diagnostics. Toward this end, we developed a new imaging approach called multidimensional microscopic molecular profiling (MMMP) that can measure several independent molecular properties in situ at subcellular resolution for the same tissue specimen. MMMP involves repeated cycles of antibody or histochemical staining, imaging, and signal removal, which ultimately can generate information analogous to a multidimensional flow cytometry analysis on intact tissue sections. We performed a MMMP analysis on a tissue microarray containing a diverse set of 102 human tissues using a panel of 15 informative antibody and 5 histochemical stains plus DAPI. Large-scale unsupervised analysis of MMMP data, and visualization of the resulting classifications, identified molecular profiles that were associated with functional tissue features. We then directly annotated H&E images from this MMMP series such that canonical histological features of interest (e.g. blood vessels, epithelium, red blood cells) were individually labeled. By integrating image annotation data, we identified molecular signatures that were associated with specific histological annotations and we developed statistical models for automatically classifying these features. The classification accuracy for automated histology labeling was objectively evaluated using a cross-validation strategy, and significant accuracy (with a median per-pixel rate of 77% per feature from 15 annotated samples) for de novo feature prediction was obtained. These results suggest that high-dimensional profiling may advance the development of computer-based systems for automatically parsing relevant histological and cellular features from molecular imaging data of arbitrary human tissue samples, and can provide a framework and resource to spur the optimization of these technologies. PLOS ONE | DOI:10.1371/journal.pone.0128975 July 15, 2015 1 / 18 Automated Histology by Multi-Dimensional Molecular Profiling Competing Interests: The authors have declared that no competing interests exist. Introduction Microscopic examination of cellular morphology and structure is a classical approach that has provided an invaluable foundation for analyzing the function, development, and organization of complex tissues. Accordingly, a large number of biomedical research and diagnostic methods are based on the identification of architectural tissue features by histopathology [1–3]. At the same time, highly multiplexed interrogation of the molecular components of different samples has proven to be a tremendously rich complementary strategy for their characterization and classification. Large-scale molecular studies based on microarray analysis, high-throughput sequencing, and proteomic approaches have clearly demonstrated the advantages of quantitative multi-dimensional profiling for identifying functionally important subtypes of cancers and other cellular states with important clinical ramifications [4–6]. Nevertheless, these techniques often require physical disruption of the interrogated samples, which sacrifices critical spatial information related to the individual cells and their native positional arrangements and relationships within intact specimens. Therefore, technologies that enable the acquisition of high-dimensional molecular profiles while retaining the spatial integrity of the examined material offer great potential for advancing the detailed characterization of important biological samples. Accordingly, several different approaches for multiplex in situ profiling of tissue sections have been pursued. Traditional multi-color fluorescence microscopy enables the simultaneous monitoring of up to five spectrally resolvable dyes at once using standard optical filters, and up to seven fluorophores may be detected with multispectral approaches [7]. In order to overcome these limitations, several independent strategies based on serial staining and imaging have been developed which greatly expand the number of molecular characteristics that can be assayed from an individual sample [8–13]. Other novel strategies based on mass spectrometry imaging modalities for the simultaneous detection of up to 32 distinct markers have also been successfully applied to the dissection of cellular states from intact clinically relevant tissue samples, further demonstrating the exceptional power of such multiplex technologies in combination with advanced analytical techniques [14–18]. Here we describe a new approach, called multi-dimensional microscopic molecular profiling (MMMP), that can measure several independent molecular properties in situ at subcellular resolution for the same tissue specimen. The MMMP procedure was adapted to work with formalin-fixed paraffin-embedded tissue samples that are commonly used for clinical specimens, and therefore was c (...truncated)