The analysis of hematoxylin and eosin (HE)-stained bone marrow (BM) tissue sections in drug development toxicologic pathology studies is a key step in the in vivo safety assessment of new drug candidates. Routine histologic analysis provides critical information about the cellularity and tissue architecture of the BM but only limited insights into the cell lineages that comprise the hematopoietic tissue. The evaluation of BM cell types can be augmented by examining BM smear preparations or using immunohistochemical (IHC) labeling of histologic sections to identify lineages of interest; however, neither of these approaches is included in the standard assessment. In addition, manual evaluation is time-consuming, subject to inter-observer variability, and challenging due to the complexity of BM morphology and architecture. In this project, we developed a deep learning model to predict IHC labeling of BM cell lineages on HE-stained BM tissue sections. The model is trained on an immunohistochemistry-informed, HE-based ground truth for the sequential labeling of CD11b and myeloperoxidase markers and can predict cell segmentation mappings with over 69% agreement with the ground truth, using only HE slides as input. Furthermore, our method can discern cell population changes that reflect qualitative diagnoses identified by pathologists during routine slide interpretation. Our automated method holds promise for enhancing routine pathologist assessments of BM HE slides by extending the evaluation of hematopoietic cell lineages without the need to generate additional samples.

The diagnostic classification of digitized tissue images based on histopathologic lesions present in whole slide images (WSI) is a significant task that eludes modern image classification techniques. Even with advanced methods designed for digital histopathology, the domain of toxicologic pathology presents challenges in that histopathologic features may be at times complex, subtle, and/or rare. We propose an innovative weakly supervised learning method that leverages minimal annotations, a state-of-the-art self-supervised vision transformer for embedding extraction, and a novel guided attention mechanism that is better suited for heavily imbalanced datasets typical in toxicologic pathology. Our model demonstrates improvements in diagnostic classification and attention heatmap quality over the previously described clustering-constrained-attention multiple-instance learning method on several lesion classes in rat livers (38% improvement in AUC). We also demonstrate how an ensemble of binary classifiers improves interpretability and allows for multiclass classification and the classification of diagnostic regions of interest in each slide. The improved classification performance and higher contrast heatmaps better support toxicologic pathologists’ histopathology analysis and will enable more efficient workflows as they are further refined and integrated into routine use.

The histopathologic evaluation of regulatory toxicity studies using artificial intelligence (AI) has the potential to increase study efficiency. For example, AI could initially identify and exclude all organs without histopathologic lesions, allowing pathologists to focus solely on evaluating organs with identified lesions. In this study, whole slide images (WSIs) of liver sections from 58 different rat toxicity studies were collected, along with their corresponding histopathologic lesion diagnoses. Each WSI was labeled as either "lesion" or "no lesion" based on the presence or absence of reported histopathologic lesions. Multiple instance learning (MIL) approaches, including a transformer variant, were tested to predict lesions within a weakly supervised framework. Both methods achieved acceptable to excellent area under the receiver operating characteristic curve (AUROC) scores. Heatmap overlays were employed to visually assess the MIL model's effectiveness in detecting lesions, confirming the accuracy of targeted areas on the WSIs. In addition, using transfer learning principles, the MIL model initially developed for liver WSIs was adapted to kidney WSIs, demonstrating the model's versatility. This study showcases the application of weakly supervised learning for lesion detection in rat WSIs from toxicity studies, with the potential to significantly enhance the efficiency of the histopathologic evaluation process.

Toxicologic pathology is undergoing a digital transformation, with advances in imaging and computational methods enabling automation of traditionally manual workflows. Central to these digital workflows is the generation of high-quality whole slide images (WSIs), where one key determinant of image quality is focus sharpness. To address this, we have integrated a pair of productionalized computational models - 'MiQC' (Microscopic Quality Control) - into our routine image QC workflows. MiQC combines Local Binary Patterns (LBP) and DeepFocus-based deep learning algorithms to detect and quantify out-of-focus regions in WSIs. Subsequent to scanner-based focus metric assessment, MiQC further screens WSIs and supports technician review by generating heatmaps that highlight problematic areas. Even WSIs with scanner focus scores of 98-99% can contain unacceptable blur, which MiQC helps identify. Using this system, 85-95% of WSIs are approved without further intervention, and technician review time is reduced by nearly 50%. Compared to fully manual review, MiQC has doubled our throughput of QC'd slides per hour. This efficiency gain has accelerated the expansion of our high-quality WSI repository and provides a scalable, reproducible framework for enhancing image QC in toxicologic and broader digital pathology applications. MiQC supports higher throughput and integration of automated image analysis pipelines, laying the groundwork for robust downstream computational pathology workflows.

The virtual control group (VCG) concept provides a potential opportunity to reduce animal use in drug development by replacing concurrent control groups (CCGs) in nonclinical toxicity studies. This work investigated the feasibility and reliability of using VCGs in place of CCGs. A historical control database (HCD), constructed from Genentech Inc. rat toxicity study data, was reviewed to understand trends and sources of variability in control animals over time, and to identify data curation requirements for assembling VCGs, e.g., alignment of units of measurement. Several endpoints were investigated and stratified against different study design parameters. Sex, route of administration, fasting status, and body weight at study initiation were among the parameters that were indicated as key matching criteria. With a high-level understanding of potential sources of variability, a retrospective proof-of-concept (POC) study was designed, evaluating a historical rat pilot toxicity study for test article-related changes. A masked interpretation of the study was conducted using its CCG and two unique VCGs that were constructed from individual animal data pulled from our HCD. While the results of the microscopic pathology assessment and most endpoints were similar across the different control groups, the POC revealed the risk of using VCGs to interpret subtle test article- related changes in clinical pathology parameters. Within the context of our POC, it appears the use of a VCG is not completely equivalent to the CCG, especially with clinical pathology parameters. Additional work is needed to understand the potential utility, and thus, viability of VCGs in other contexts.

Atopic dermatitis (AD) is a chronic inflammatory skin disease with significant health/economic burdens. Existing therapies are not fully effective, necessitating development of new approaches for AD management. Here, we report that dietary grape powder (GP) mitigates AD-like symptoms in 2,4-dinitrofluorobenzene (DNFB)-induced AD in NC/NgaTndCrlj mice. Using prevention and intervention protocols, we tested the efficacy of 3% and 5% GP-fortified diet in a 13-weeks study. We found that GP feeding markedly inhibited development and progression of AD-like skin lesions, and caused reduction in i) epidermal thickness, mast cell infiltration, ulceration, excoriation and acanthosis in dorsal skin, ii) spleen weight, extramedullary hematopoiesis and lymph nodes sizes, and iii) ear weight and IgE levels. We also found significant modulations in 15 AD-associated serum cytokines/chemokines. Next, using quantitative global proteomics, we identified 714 proteins. Of these, 68 (normal control) and 21 (5% GP-prevention) were significantly modulated (≥2-fold) vs AD control (DNFB-treated) group, with many GP-modulated proteins reverting to normal levels. Ingenuity pathway analysis of GP-modulated proteins followed by validation using ProteinSimple identified changes in acute phase response signaling (FGA, FGB, FGG, HP, HPX, LRG1). Overall, GP supplementation inhibited DNFB-induced AD in NC/NgaTndCrlj mice in both prevention and intervention trials, and should be explored further.

PDF file - 54K, Panel A illustrates the colonoscopic surveillance protocol. Panel B shows the serial biopsy protocol.

Digital pathology workflows in toxicologic pathology rely on whole slide images (WSIs) from histopathology slides. Inconsistent color reproduction by WSI scanners of different models and from different manufacturers can result in different color representations and inter-scanner color variation in the WSIs. Although pathologists can accommodate a range of color variation during their evaluation of WSIs, color variability can degrade the performance of computational applications in digital pathology. In particular, color variability can compromise the generalization of artificial intelligence applications to large volumes of data from diverse sources. To address these challenges, we developed a process that includes two modules: (1) assessing the color reproducibility of our scanners and the color variation among them and (2) applying color correction to WSIs to minimize the color deviation and variation. Our process ensures consistent color reproduction across WSI scanners and enhances color homogeneity in WSIs, and its flexibility enables easy integration as a post-processing step following scanning by WSI scanners of different models and from different manufacturers.

PDF file - 28K, Differentially Expressed Gene Functional Categories between Adenomas and Intramucosal Carcinomas

PDF file - 583K, Mice expressing activated PI3K were scanned to identify those bearing tumors. After sacrifice, the intestine was removed to characterize tissue morphology, cell signaling and cell proliferation within the epithelium and the associated neoplasms

This file contains 5 supplemental figures and 1 supplemental table

Figure S1 shows that there are equivalent levels of CIN in p53-/- and CENP-E+/-;p53-/- lymphomas

PDF file - 21K, Characteristics of F1 ApcMin/+ mice treated with 4% dextran sodium sulfate.

Table S1 shows expected and observed birth frequencies.

Figure S2 shows the sex-specific survival advantage of p53 heterozygous male mice is not caused by differences in tumor spectrum.

Figure S5 shows that high CIN causes cell death due to excessive chromosome loss.

Supplementary Figures 1-2 from Inactivation of Apc in the Mouse Prostate Causes Prostate Carcinoma

PDF file - 30K, Relative gene expression is shown for three adenomas and three intramucosal carcinomas taken from the colons of DSS-treated F1 Min mice.

PDF file - 42K, Complete list of differentially expressed genes between intramucosal carcinomas and adenomas.

Figure S3 shows that reduction of CENP-E extends tumor latency only in tumors where it causes high CIN.