AI Driven Spatial Pathology: The Next NGS

The job of the pathologist has changed since the emergence of precision oncology. Prior to targeted therapies, the main customer of the pathologist was the surgeon. Surgery was the lynchpin of cancer therapy along with radiation and conventional chemotherapy. The pathologist’s attention was focused on detailed assessment of tumor characteristics, such as tumor grade and differentiation, the completeness of surgical resection, the presence of lymphatic and/or vascular invasion and lymph node involvement. The benefits from targeted therapy changed that. Today, pathologists are asked to provide detailed information on a variety of biomarkers that are used to select the best therapy for patients.

Biomarkers are essential for precision oncology; put another way, the incorrect assessment of a biomarker can result in a patient receiving ineffective therapy. In this new world, the oncologist is now the main customer of the pathologist – and the details of the pathology report matter more than ever. This approach places new and greater responsibility on pathologists, and they have responded by developing and implementing new guidelines, technologies and proficiency tests to ensure accurate and reproducible biomarker results.

The College of American Pathologists and the American Society of Clinical Oncology have published guidelines for the proper handling of tissue specimens, including fixation, processing, staining and reporting. Proficiency testing has become mandatory for laboratory accreditation giving clinicians more confidence in reported results, and new technologies such as next generation sequencing (NGS) have become an essential tool to reliably identify the growing number of molecular alterations that drive therapy decisions. Though targeted therapies are a significant advancement in cancer treatment, increasing response rates and progression-free survival in most patients whose tumor express the target, they demonstrate limited durability and improvement in overall survival.

One new weapon in precision oncology is immunotherapy, which has resulted in durable responses and even complete remissions in a subset of patients. A variety of biomarkers have been developed to aid in the identification of patients most likely to respond to a common type of immunotherapy called checkpoint inhibition. These include PD-L1 expression, tumor mutational burden, gene expression profiling, and tumor infiltrating lymphocytes. Unfortunately, even with these tests, it is difficult to reliably predict which patients will benefit from therapy.

Immunotherapy therapy works by restoring the immune system’s ability to recognize and attack cancer cells rather than killing the tumor directly. To understand how to restore proper immune function, it is important to learn how the immune system recognizes and eliminates tumor cells. Cancers are the result of an accumulation of genetic alterations that produce aberrant proteins, some of which are recognized as abnormal by a person’s immune system. These are known as neoantigens. As neoantigens are released by the tumor, they are captured by dendritic cells for processing. Dendritic cells present the captured antigens to T cells in the lymph node, resulting in activation of CD8+ cytotoxic T cells and CD4+ helper cells. This process primes T cells to recognize and target cancer cells expressing these neoantigens. Activated T cells leave the lymph node and travel through the bloodstream to the tumor microenvironment where they infiltrate the tumor. Once there, activated T cells identify and bind to cancer cells that present the specific neoantigen they are primed against. Finally, CD8+ T cells release cytotoxic molecules resulting in cancer cell death.

Unfortunately, cancers employ multiple strategies to evade immune surveillance. These include the ability of cancer cells to downregulate antigen presentation, create an immunosuppressive microenvironment (by recruiting regulatory T cells or expressing inhibitory molecules like PD-L1 or cytotoxic T-lymphocyte-associated protein-4), produce immunosuppressive factors like tumor growth factor-β or indoleamine-pyrrole 2,3-dioxygenase, activate survival pathways, and recruit myeloid-derived tumor suppressor cells, tumor associated macrophages, and cancer associated fibroblasts. Additionally, prolonged exposure to neoantigens can result in the progressive loss of effector T cell function, known as T cell exhaustion. These cells demonstrate decreased effector molecules, such as perforin and granzyme B, and show upregulation of inhibitory receptors, such as PD-1, T-cell immunoglobulin, mucin domain-3, and lymphocyte activation gene 3.

Just as NGS interrogates the many potential molecular alterations that serve as biomarkers for targeted therapy, new technologies are needed to identify better biomarkers that allow the selection of patients who will benefit from immunotherapy. But cancers seem “smart,” using every mechanism at their disposal to avoid immunosurveillance and thrive. To personalize immunotherapy, new spatial tools must identify the different types, states, amounts, and location of cells that make up the tumor and its microenvironment. Spatial biology tools are now being commonly employed in research, but they are slow, expensive, and require specialized equipment, which limits their adoption, especially for clinical use. 

But spatial pathology platforms, driven by deep learning algorithms, are now emerging to overcome these limitations. Eventually, these platforms will integrate into existing digital pathology workflows, allowing pathologists to dissect the complex interactions taking place within a tumor and its microenvironment – crucially, in a rapid and cost-effective manner. To do this, two main types of deep learning models are used. The first model takes a whole slide image as an input and uses various neural networks to predict tumor features, such as gene signatures, genomic abnormalities, and instability. The second model uses deep learning to classify the components of whole slide images into regions, like tumor and stroma, and then segment and classify every individual cell on the slide. 

Deep learning models have traditionally been trained using pathologist’s annotations but pathologists are limited in their ability to reliably classify cells. Next generation cell classification methods will use multiplexed immunofluorescence or transcriptomic profiling to train classifiers capable of reliably predicting cell types and cell states from H&E slides.

Immunotherapy and other emerging therapies, such as antibody-drug conjugates, could transform oncology care and improve patient outcomes – but only if we can identify biomarkers that can reliably select patients who derive benefit. Thankfully, advancements in digital pathology, high plex profiling, and deep learning have arrived. Together, they will provide pathologists with the platform necessary to classify every cell on a slide and unravel their relationships and interactions to produce the next generation of tumor biomarkers, fulfilling the true promise of precision oncology.