For newly diagnosed breast cancer, estrogen receptor status (ERS) is a key molecular marker used for prognosis and treatment decisions. During clinical management, ERS is determined by pathologists from immunohistochemistry (IHC) staining of biopsied tissue for the targeted receptor, which highlights the presence of cellular surface antigens. This is an expensive, time-consuming process which introduces discordance in results due to variability in IHC preparation and pathologist subjectivity. In contrast, hematoxylin and eosin (H&E) staining—which highlights cellular morphology—is quick, less expensive, and less variable in preparation. Here we show that machine learning can determine molecular marker status, as assessed by hormone receptors, directly from cellular morphology. We develop a multiple instance learning-based deep neural network that determines ERS from H&E-stained whole slide images (WSI). Our algorithm—trained strictly with WSI-level annotations—is accurate on a varied, multi-country dataset of 3,474 patients, achieving an area under the curve (AUC) of 0.92 for sensitivity and specificity. Our approach has the potential to augment clinicians’ capabilities in cancer prognosis and theragnosis by harnessing biological signals imperceptible to the human eye.
Because histologic types are subjective and difficult to reproduce between pathologists, tissue morphology often takes a back seat to molecular testing for the selection of breast cancer treatments. This work explores whether a deep-learning algorithm can learn objective histologic H&E features that predict the clinical subtypes of breast cancer, as assessed by immunostaining for estrogen, progesterone, and Her2 receptors (ER/PR/Her2). Translating deep learning to this and related problems in histopathology presents a challenge due to the lack of large, well-annotated data sets, which are typically required for the algorithms to learn statistically significant discriminatory patterns. To overcome this limitation, we introduce the concept of “tissue fingerprints,” which leverages large, unannotated datasets in a label-free manner to learn H&E features that can distinguish one patient from another. The hypothesis is that training the algorithm to learn the morphological differences between patients will implicitly teach it about the biologic variation between them. Following this training internship, we used the features the network learned, which we call “fingerprints,” to predict ER, PR, and Her2 status in two datasets. Despite the discovery dataset being relatively small by the standards of the machine learning community (n = 939), fingerprints enabled the determination of ER, PR, and Her2 status from whole slide H&E images with 0.89 AUC (ER), 0.81 AUC (PR), and 0.79 AUC (Her2) on a large, independent test set (n = 2531). Tissue fingerprints are concise but meaningful histopathologic image representations that capture biological information and may enable machine learning algorithms that go beyond the traditional ER/PR/Her2 clinical groupings by directly predicting theragnosis.
In this pilot study, we introduce a machine learning framework to identify relationships between cancer tissue morphology and hormone receptor pathway activation in breast cancer pathology hematoxylin and eosin (H&E)-stained samples. As a proof-of-concept, we focus on predicting clinical estrogen receptor (ER) status—defined as greater than one percent of cells positive for estrogen receptor by immunohistochemistry staining—from spatial arrangement of nuclear features. Our learning pipeline segments nuclei from H&E images, extracts their position, shape and orientation descriptors, and then passes them to a deep neural network to predict ER status. After training on 57 tissue cores of invasive ductal carcinoma (IDC), our pipeline predicted ER status in an independent test set of patient samples (AUC ROC = 0.72, 95%CI = 0.55–0.89, n = 56). This proof of concept shows that machine-derived descriptors of morphologic histology patterns can be correlated to signaling pathway status. Unlike other deep learning approaches to pathology, our system uses deep neural networks to learn spatial relationships between pre-defined biological features, which improves the interpretability of the system and sheds light on the features the neural network uses to predict ER status. Future studies will correlate morphometry to quantitative measures of estrogen receptor status and, ultimately response to hormonal therapy.
Cell-free DNA (cfDNA) is an attractive cancer biomarker, as it is thought to reflect a component of the underlying genetic makeup of the tumor and is readily accessible in serial fashion. Because chemotherapy regimens are expected to act rapidly on cancer and cfDNA is cleared from the blood within minutes, we hypothesized that cfDNA would reflect immediate effects of treatment. Here, we developed a method for monitoring long cfDNA fragments, and report dynamic changes in response to cytotoxic chemotherapy. Peripheral blood was obtained from 15 patients with metastatic castration-resistant prostate cancer (CRPC) immediately before and after cytotoxic chemotherapy infusion. cfDNA was extracted and quantified for long interspersed nuclear elements (LINE1; 297 bp) using qPCR. Targeted deep sequencing was performed to quantify the frequency of mutations in exon 8 of the androgen receptor (AR), a mutational hotspot region in CRPC. Single nucleotide mutations in AR exon 8 were found in 6 subjects (6/15 = 40%). Analytical variability was minimized by pooling independent PCR reactions for each library. In 5 patients, tumor-derived long cfDNA levels were found to change immediately after infusion. Detailed analysis of one subject suggests that cytotoxic chemotherapy can produce rapidly observable effects on cfDNA.
Live cell imaging has improved our ability to measure phenotypic heterogeneity. However, bottlenecks in imaging and image processing often make it difficult to differentiate interesting biological behavior from technical artifact. Thus there is a need for new methods that improve data quality without sacrificing throughput. Here we present a 3-step workflow to improve dynamic phenotype measurements of heterogeneous cell populations. We provide guidelines for image acquisition, phenotype tracking, and data filtering to remove erroneous cell tracks using the novel Tracking Aberration Measure (TrAM). Our workflow is broadly applicable across imaging platforms and analysis software. By applying this workflow to cancer cell assays, we reduced aberrant cell track prevalence from 17% to 2%. The cost of this improvement was removing 15% of the well-tracked cells. This enabled detection of significant motility differences between cell lines. Similarly, we avoided detecting a false change in translocation kinetics by eliminating the true cause: varied proportions of unresponsive cells. Finally, by systematically seeking heterogeneous behaviors, we detected subpopulations that otherwise could have been missed, including early apoptotic events and pre-mitotic cells. We provide optimized protocols for specific applications and step-by-step guidelines for adapting them to a variety of biological systems.
The genetic origins of chemotherapy resistance are well established; however, the role of epigenetics in drug resistance is less well understood. To investigate mechanisms of drug resistance, we performed systematic genetic, epigenetic, and transcriptomic analyses of an alkylating agent-sensitive murine lymphoma cell line and a series of resistant lines derived by drug dose escalation. Dose escalation of the alkylating agent mafosfamide was used to create a series of increasingly drug-resistant mouse Burkitt’s lymphoma cell lines. Whole genome sequencing, DNA microarrays, reduced representation bisulfite sequencing, and chromatin immunoprecipitation sequencing were used to identify alterations in DNA sequence, mRNA expression, CpG methylation, and H3K27me3 occupancy, respectively, that were associated with increased resistance. Our data suggest that acquired resistance cannot be explained by genetic alterations. Based on integration of transcriptional profiles with transcription factor binding data, we hypothesize that resistance is driven by epigenetic plasticity. We observed that the resistant cells had H3K27me3 and DNA methylation profiles distinct from those of the parental lines. Moreover, we observed DNA methylation changes in the promoters of genes regulated by E2a and members of the polycomb repressor complex 2 (PRC2) and differentially expressed genes were enriched for targets of E2a. The integrative analysis considering H3K27me3 further supported a role for PRC2 in mediating resistance. By integrating our results with data from the Immunological Genome Project (Immgen.org), we showed that these transcriptional changes track the B-cell maturation axis. Our data suggest a novel mechanism of drug resistance in which E2a and PRC2 drive changes in the B-cell epigenome; these alterations attenuate alkylating agent treatment-induced apoptosis.