Search

Abstract Cancers occur across species. Understanding what is consistent and varies across species can provide new insights into cancer initiation and evolution, with significant implications for animal welfare and wildlife conservation. We build a pan-species cancer digital pathology atlas (panspecies.ai) and conduct a pan-species study of computational comparative pathology using a supervised convolutional neural network algorithm trained on human samples. The artificial intelligence algorithm achieves high accuracy in measuring immune response through single-cell classification for two transmissible cancers (canine transmissible venereal tumour, 0.94; Tasmanian devil facial tumour disease, 0.88). In 18 other vertebrate species (mammalia = 11, reptilia = 4, aves = 2, and amphibia = 1), accuracy (range 0.57–0.94) is influenced by cell morphological similarity preserved across different taxonomic groups, tumour sites, and variations in the immune compartment. Furthermore, a spatial immune score based on artificial intelligence and spatial statistics is associated with prognosis in canine melanoma and prostate tumours. A metric, named morphospace overlap, is developed to guide veterinary pathologists towards rational deployment of this technology on new samples. This study provides the foundation and guidelines for transferring artificial intelligence technologies to veterinary pathology based on understanding of morphological conservation, which could vastly accelerate developments in veterinary medicine and comparative oncology.

Summary: Background: Efficient biomarker discovery and clinical translation depend on the fast and accurate analytical output from crucial technologies such as multiplex imaging. However, reliable cell classification often requires extensive annotations. Label-efficient strategies are urgently needed to reveal diverse cell distribution and spatial interactions in large-scale multiplex datasets. Methods: This study proposed Self-supervised Learning for Antigen Detection (SANDI) for accurate cell phenotyping while mitigating the annotation burden. The model first learns intrinsic pairwise similarities in unlabelled cell images, followed by a classification step to map learnt features to cell labels using a small set of annotated references. We acquired four multiplex immunohistochemistry datasets and one imaging mass cytometry dataset, comprising 2825 to 15,258 single-cell images to train and test the model. Findings: With 1% annotations (18–114 cells), SANDI achieved weighted F1-scores ranging from 0.82 to 0.98 across the five datasets, which was comparable to the fully supervised classifier trained on 1828–11,459 annotated cells (−0.002 to −0.053 of averaged weighted F1-score, Wilcoxon rank-sum test, P = 0.31). Leveraging the immune checkpoint markers stained in ovarian cancer slides, SANDI-based cell identification reveals spatial expulsion between PD1-expressing T helper cells and T regulatory cells, suggesting an interplay between PD1 expression and T regulatory cell-mediated immunosuppression. Interpretation: By striking a fine balance between minimal expert guidance and the power of deep learning to learn similarity within abundant data, SANDI presents new opportunities for efficient, large-scale learning for histology multiplex imaging data. Funding: This study was funded by the Royal Marsden/ICR National Institute of Health Research Biomedical Research Centre.

Abstract Hypoxia promotes aggressive tumor phenotypes and mediates the recruitment of suppressive T cells in invasive breast carcinomas. We investigated the role of hypoxia in relation to T-cell regulation in ductal carcinoma in situ (DCIS). We designed a deep learning system tailored for the tissue architecture complexity of DCIS, and compared pure DCIS cases with the synchronous DCIS and invasive components within invasive ductal carcinoma cases. Single-cell classification was applied in tandem with a new method for DCIS ductal segmentation in dual-stained CA9 and FOXP3, whole-tumor section digital pathology images. Pure DCIS typically has an intermediate level of colocalization of FOXP3+ and CA9+ cells, but in invasive carcinoma cases, the FOXP3+ (T-regulatory) cells may have relocated from the DCIS and into the invasive parts of the tumor, leading to high levels of colocalization in the invasive parts but low levels in the synchronous DCIS component. This may be due to invasive, hypoxic tumors evolving to recruit T-regulatory cells in order to evade immune predation. Our data support the notion that hypoxia promotes immune tolerance through recruitment of T-regulatory cells, and furthermore indicate a spatial pattern of relocalization of T-regulatory cells from DCIS to hypoxic tumor cells. Spatial colocalization of hypoxic and T-regulatory cells may be a key event and useful marker of DCIS progression.

Summary: Glioblastoma (GBM) recurrence originates from invasive margin cells that escape surgical debulking, but to what extent these cells resemble their bulk counterparts remains unclear. Here, we generated three immunocompetent somatic GBM mouse models, driven by subtype-associated mutations, to compare matched bulk and margin cells. We find that, regardless of mutations, tumors converge on common sets of neural-like cellular states. However, bulk and margin have distinct biology. Injury-like programs associated with immune infiltration dominate in the bulk, leading to the generation of lowly proliferative injured neural progenitor-like cells (iNPCs). iNPCs account for a significant proportion of dormant GBM cells and are induced by interferon signaling within T cell niches. In contrast, developmental-like trajectories are favored within the immune-cold margin microenvironment resulting in differentiation toward invasive astrocyte-like cells. These findings suggest that the regional tumor microenvironment dominantly controls GBM cell fate and biological vulnerabilities identified in the bulk may not extend to the margin residuum.

Abstract Despite increasing evidence supporting the clinical relevance of tumour infiltrating lymphocytes (TILs) in invasive breast cancer, TIL spatial variability within ductal carcinoma in situ (DCIS) samples and its association with progression are not well understood. To characterise tissue spatial architecture and the microenvironment of DCIS, we designed and validated a new deep learning pipeline, UNMaSk. Following automated detection of individual DCIS ducts using a new method IM-Net, we applied spatial tessellation to create virtual boundaries for each duct. To study local TIL infiltration for each duct, DRDIN was developed for mapping the distribution of TILs. In a dataset comprising grade 2–3 pure DCIS and DCIS adjacent to invasive cancer (adjacent DCIS), we found that pure DCIS cases had more TILs compared to adjacent DCIS. However, the colocalisation of TILs with DCIS ducts was significantly lower in pure DCIS compared to adjacent DCIS, which may suggest a more inflamed tissue ecology local to DCIS ducts in adjacent DCIS cases. Our study demonstrates that technological developments in deep convolutional neural networks and digital pathology can enable an automated morphological and microenvironmental analysis of DCIS, providing a new way to study differential immune ecology for individual ducts and identify new markers of progression.

Background: Despite treatment advances, there remains a significant risk of recurrence in ovarian cancer, at which stage it is usually incurable. Consequently, there is a clear need for improved patient stratification. However, at present clinical prognosticators remain largely unchanged due to the lack of reproducible methods to identify high-risk patients. Methods: In high-grade serous ovarian cancer patients with advanced disease, we spatially define a tumour ecological balance of stromal resource and immune hazard using high-throughput image and spatial analysis of routine histology slides. On this basis an EcoScore is developed to classify tumours by a shift in this balance towards cancer-favouring or inhibiting conditions. Findings: The EcoScore provides prognostic value stronger than, and independent of, known risk factors. Crucially, the clinical relevance of mutational burden and genomic instability differ under different stromal resource conditions, suggesting that the selective advantage of these cancer hallmarks is dependent on the context of stromal spatial structure. Under a high resource condition defined by a high level of geographical intermixing of cancer and stromal cells, selection appears to be driven by point mutations; whereas, in low resource tumours featured with high hypoxia and low cancer-immune co-localization, selection is fuelled by aneuploidy. Interpretation: Our study offers empirical evidence that cancer fitness depends on tumour spatial constraints, and presents a biological basis for developing better assessments of tumour adaptive strategies in overcoming ecological constraints including immune surveillance and hypoxia. Keywords: Tumour ecology, Tumour spatial heterogeneity, Cancer evolution, Ovarian cancer, Histological image analysis

High computational cost associated with digital pathology image analysis approaches is a challenge towards their translation in routine pathology clinic. Here, we propose a computationally efficient framework (SuperHistopath), designed to map global context features reflecting the rich tumor morphological heterogeneity. SuperHistopath efficiently combines i) a segmentation approach using the linear iterative clustering (SLIC) superpixels algorithm applied directly on the whole-slide images at low resolution (5x magnification) to adhere to region boundaries and form homogeneous spatial units at tissue-level, followed by ii) classification of superpixels using a convolution neural network (CNN). To demonstrate how versatile SuperHistopath was in accomplishing histopathology tasks, we classified tumor tissue, stroma, necrosis, lymphocytes clusters, differentiating regions, fat, hemorrhage and normal tissue, in 127 melanomas, 23 triple-negative breast cancers, and 73 samples from transgenic mouse models of high-risk childhood neuroblastoma with high accuracy (98.8%, 93.1% and 98.3% respectively). Furthermore, SuperHistopath enabled discovery of significant differences in tumor phenotype of neuroblastoma mouse models emulating genomic variants of high-risk disease, and stratification of melanoma patients (high ratio of lymphocyte-to-tumor superpixels (p = 0.015) and low stroma-to-tumor ratio (p = 0.028) were associated with a favorable prognosis). Finally, SuperHistopath is efficient for annotation of ground-truth datasets (as there is no need of boundary delineation), training and application (~5 min for classifying a whole-slide image and as low as ~30 min for network training). These attributes make SuperHistopath particularly attractive for research in rich datasets and could also facilitate its adoption in the clinic to accelerate pathologist workflow with the quantification of phenotypes, predictive/prognosis markers.

Cancer cells can actively engage in overcoming microenvironmental constraints such as tissue stiffness through adapting their shapes; however it is unclear how microenvironmental cells shape cancer nuclear morphology in human tumors in situ. Here the authors merge machine learning, digital pathology and spatial statistics to study this issue; furthermore the authors identify decreased immune infiltration in the surrounding of diversified cancer cells in a subset of ovarian tumors.

Computational pathology-based cell classification algorithms are revolutionizing the study of the tumor microenvironment and can provide novel predictive/prognosis biomarkers crucial for the delivery of precision oncology. Current algorithms used on hematoxylin and eosin slides are based on individual cell nuclei morphology with limited local context features. Here, we propose a novel multi-resolution hierarchical framework (SuperCRF) inspired by the way pathologists perceive regional tissue architecture to improve cell classification and demonstrate its clinical applications. We develop SuperCRF by training a state-of-art deep learning spatially constrained- convolution neural network (SC-CNN) to detect and classify cells from 105 high-resolution (20×) H&E-stained slides of The Cancer Genome Atlas melanoma dataset and subsequently, a conditional random field (CRF) by combining cellular neighborhood with tumor regional classification from lower resolution images (5, 1.25×) given by a superpixel-based machine learning framework. SuperCRF led to an 11.85% overall improvement in the accuracy of the state-of-art deep learning SC-CNN cell classifier. Consistent with a stroma-mediated immune suppressive microenvironment, SuperCRF demonstrated that (i) a high ratio of lymphocytes to all lymphocytes within the stromal compartment (p = 0.026) and (ii) a high ratio of stromal cells to all cells (p < 0.0001 compared to p = 0.039 for SC-CNN only) are associated with poor survival in patients with melanoma. SuperCRF improves cell classification by introducing global and local context-based information and can be implemented in combination with any single-cell classifier. SuperCRF provides valuable tools to study the tumor microenvironment and identify predictors of survival and response to therapy.

Diffusion-weighted magnetic resonance imaging (DWI) enables non-invasive, quantitative staging of prostate cancer via measurement of the apparent diffusion coefficient (ADC) of water within tissues. In cancer, more advanced disease is often characterized by higher cellular density (cellularity), which is generally accepted to correspond to a lower measured ADC. A quantitative relationship between tissue structure and in vivo measurements of ADC has yet to be determined for prostate cancer. In this study, we establish a theoretical framework for relating ADC measurements with tissue cellularity and the proportion of space occupied by prostate lumina, both of which are estimated through automatic image processing of whole-slide digital histology samples taken from a cohort of six healthy mice and nine transgenic adenocarcinoma of the mouse prostate (TRAMP) mice. We demonstrate that a significant inverse relationship exists between ADC and tissue cellularity that is well characterized by our model, and that a decrease of the luminal space within the prostate is associated with a decrease in ADC and more aggressive tumor subtype. The parameters estimated from our model in this mouse cohort predict the diffusion coefficient of water within the prostate-tissue to be 2.18 × 10−3 mm2/s (95% CI: 1.90, 2.55). This value is significantly lower than the diffusion coefficient of free water at body temperature suggesting that the presence of organelles and macromolecules within tissues can drastically hinder the random motion of water molecules within prostate tissue. We validate the assumptions made by our model using novel in silico analysis of whole-slide histology to provide the simulated ADC (sADC); this is demonstrated to have a significant positive correlation with in vivo measured ADC (r2 = 0.55) in our mouse population. The estimation of the structural properties of prostate tissue is vital for predicting and staging cancer aggressiveness, but prostate tissue biopsies are painful, invasive, and are prone to complications such as sepsis. The developments made in this study provide the possibility of estimating the structural properties of prostate tissue via non-invasive virtual biopsies from MRI, minimizing the need for multiple tissue biopsies and allowing sequential measurements to be made for prostate cancer monitoring.

To develop an image analysis technique that distinguishes pseudoprogression from true progression by analyzing tumour heterogeneity in T2-weighted images using topological descriptors of image heterogeneity called Minkowski functionals (MFs).Using a retrospective patient cohort (n = 50), and blinded to treatment response outcome, unsupervised feature estimation was performed to investigate MFs for the presence of outliers, potential confounders, and sensitivity to treatment response. The progression and pseudoprogression groups were then unblinded and supervised feature selection was performed using MFs, size and signal intensity features. A support vector machine model was obtained and evaluated using a prospective test cohort.The model gave a classification accuracy, using a combination of MFs and size features, of more than 85% in both retrospective and prospective datasets. A different feature selection method (Random Forest) and classifier (Lasso) gave the same results. Although not apparent to the reporting radiologist, the T2-weighted hyperintensity phenotype of those patients with progression was heterogeneous, large and frond-like when compared to those with pseudoprogression.Analysis of heterogeneity, in T2-weighted MR images, which are acquired routinely in the clinic, has the potential to detect an earlier treatment response allowing an early change in treatment strategy. Prospective validation of this technique in larger datasets is required.

BackgroundThe intra-tumor diversity of cancer cells is under intense investigation; however, little is known about the heterogeneity of the tumor microenvironment that is key to cancer progression and evolution. We aimed to assess the degree of microenvironmental heterogeneity in breast cancer and correlate this with genomic and clinical parameters.Methods and findingsWe developed a quantitative measure of microenvironmental heterogeneity along three spatial dimensions (3-D) in solid tumors, termed the tumor ecosystem diversity index (EDI), using fully automated histology image analysis coupled with statistical measures commonly used in ecology. This measure was compared with disease-specific survival, key mutations, genome-wide copy number, and expression profiling data in a retrospective study of 510 breast cancer patients as a test set and 516 breast cancer patients as an independent validation set. In high-grade (grade 3) breast cancers, we uncovered a striking link between high microenvironmental heterogeneity measured by EDI and a poor prognosis that cannot be explained by tumor size, genomics, or any other data types. However, this association was not observed in low-grade (grade 1 and 2) breast cancers. The prognostic value of EDI was superior to known prognostic factors and was enhanced with the addition of TP53 mutation status (multivariate analysis test set, p = 9 × 10-4, hazard ratio = 1.47, 95% CI 1.17-1.84; validation set, p = 0.0011, hazard ratio = 1.78, 95% CI 1.26-2.52). Integration with genome-wide profiling data identified losses of specific genes on 4p14 and 5q13 that were enriched in grade 3 tumors with high microenvironmental diversity that also substratified patients into poor prognostic groups. Limitations of this study include the number of cell types included in the model, that EDI has prognostic value only in grade 3 tumors, and that our spatial heterogeneity measure was dependent on spatial scale and tumor size.ConclusionsTo our knowledge, this is the first study to couple unbiased measures of microenvironmental heterogeneity with genomic alterations to predict breast cancer clinical outcome. We propose a clinically relevant role of microenvironmental heterogeneity for advanced breast tumors, and highlight that ecological statistics can be translated into medical advances for identifying a new type of biomarker and, furthermore, for understanding the synergistic interplay of microenvironmental heterogeneity with genomic alterations in cancer cells.

Predicting response to treatment and disease-specific deaths are key tasks in cancer research yet there is a lack of methodologies to achieve these. Large-scale ’omics and digital pathology technologies have led to the need for effective statistical methods for data fusion to extract the most useful patterns from these diverse data types. We present FusionGP, a method for combining heterogeneous data types designed specifically for predicting outcome of treatment and disease. FusionGP is a Gaussian process model that includes a generalization of feature selection for biomarker discovery, allowing for simultaneous, sparse feature selection across multiple data types. Importantly, it can accommodate highly nonlinear structure in the data, and automatically infers the optimal contribution from each input data type. FusionGP compares favourably to several popular classification methods, including the Random Forest classifier, a stepwise logistic regression model and the Support Vector Machine on single data types. By combining gene expression, copy number alteration and digital pathology image data in 119 estrogen receptor (ER)-negative and 345 ER-positive breast tumours, we aim to predict two important clinical outcomes: death and chemoinsensitivity. While gene expression data give the best predictive performance in the majority of cases, the digital pathology data are much better for predicting death in ER cases. Thus, FusionGP is a new tool for selecting informative features from heterogeneous data types and predicting treatment response and prognosis.

Different data types can offer complementary perspectives on the same biological phenomenon. In cancer studies, for example, data on copy number alterations indicate losses and amplifications of genomic regions in tumours, while transcriptomic data point to the impact of genomic and environmental events on the internal wiring of the cell. Fusing different data provides a more comprehensive model of the cancer cell than that offered by any single type. However, biological signals in different patients exhibit diverse degrees of concordance due to cancer heterogeneity and inherent noise in the measurements. This is a particularly important issue in cancer subtype discovery, where personalised strategies to guide therapy are of vital importance. We present a nonparametric Bayesian model for discovering prognostic cancer subtypes by integrating gene expression and copy number variation data. Our model is constructed from a hierarchy of Dirichlet Processes and addresses three key challenges in data fusion: (i) To separate concordant from discordant signals, (ii) to select informative features, (iii) to estimate the number of disease subtypes. Concordance of signals is assessed individually for each patient, giving us an additional level of insight into the underlying disease structure. We exemplify the power of our model in prostate cancer and breast cancer and show that it outperforms competing methods. In the prostate cancer data, we identify an entirely new subtype with extremely poor survival outcome and show how other analyses fail to detect it. In the breast cancer data, we find subtypes with superior prognostic value by using the concordant results. These discoveries were crucially dependent on our model's ability to distinguish concordant and discordant signals within each patient sample, and would otherwise have been missed. We therefore demonstrate the importance of taking a patient-specific approach, using highly-flexible nonparametric Bayesian methods.

Inferring regulatory relationships among many genes based on their temporal variation in transcript abundance has been a popular research topic. Due to the nature of microarray experiments, classical tools for time series analysis lose power since the number of variables far exceeds the number of the samples. In this paper, we describe some of the existing multivariate inference techniques that are applicable to hundreds of variables and show the potential challenges for small-sample, large-scale data. We propose a directed partial correlation (DPC) method as an efficient and effective solution to regulatory network inference using these data. Specifically for genomic data, the proposed method is designed to deal with large-scale datasets. It combines the efficiency of partial correlation for setting up network topology by testing conditional independence, and the concept of Granger causality to assess topology change with induced interruptions. The idea is that when a transcription factor is induced artificially within a gene network, the disruption of the network by the induction signifies a genes role in transcriptional regulation. The benchmarking results using GeneNetWeaver, the simulator for the DREAM challenges, provide strong evidence of the outstanding performance of the proposed DPC method. When applied to real biological data, the inferred starch metabolism network in Arabidopsis reveals many biologically meaningful network modules worthy of further investigation. These results collectively suggest DPC is a versatile tool for genomics research. The R package DPC is available for download (http://code.google.com/p/dpcnet/).