AstraZeneca

Jason Hipp

The gold standard of cancer diagnosis is performed by pathologists reviewing thin slices of chemically stained (Hematoxylin & Eosin based, H&E) tumor tissues. This technique has changed very little in the past 100+ years. The quality and speed of Whole Slide Imaging (WSI) scanners have recently reached the ability to be used clinically. These scanners convert a glass digital slide into a giga-pixel image and have recently obtained FDA clearance.

With the recent advancements of Artificial Intelligence (AI) algorithms, more specifically referred to as Machine Learning (ML) and Deep Learning (DL) algorithms, offer the ability to extract pixel level information from these gigapixel images that can not be otherwise observed by the visual interpretation by human pathologists. While this has significant promise in improving the accuracy and consistency of diagnoses in the clinical practice of pathology, this new technology has an even great potential impact in identifying predictive image based biomarkers for Companion Diagnostic (CDx) use and to identify drug mechanisms of action by identifying and quantifying perturbations in the cellular Tumor MicroEnvironment (TME); as opposed to traditionally developed molecular biomarkers and assays.  Although there are greater than 50 articles in a recent pubmed search of pathology and artificial intelligence and images, there is very little available in the burgeoning field of applying it to drug discovery.

While a whole slide image of a patient’s tumor represents the morphological (size, shape, and texture) features of tumor cells, it also serves as a window into the genetic and molecules being made by the tumor cells. For example, recent papers have shown that from a H&E image, the mutational status of the tumor can be predicted with ML. Thus, one can logically hypothesize that the tumor morphology is induced by the tumors DNA/RNA/Protein expression. Furthermore, other studies are showing that specific protein expression can be predicted from H&E images alone, such as PDL1. If such encoded information is in the H&E images of tumors, one can also imagine other variables, such as clinical and treatment related response, that can be similarly predicted from H&E images.

When thinking of the future, one can find even more data with a higher likelihood of finding more predictive variables of drug response by integrating digital pathology images with orthogonally related radiology images.

Radiological images of tumors also obtain key tumor related data by its examination at a different length scale than in pathology. For example, current studies have shown the ML can predict chemo response and tumor infiltrating lymphocytes in CT images.

Similarly, this goes to say that the integration of -Omics data derived from the same tumor as the imaging data could provide an additional dimension of encoded data for interpretation. Obtaining these 3 orthogonally related types of data from tumors is relatively easy to obtain with current technologies. However, current practice in drug development is limited by analyzing these data sets independently, and thus we end up sub-sampling the tumor data and are limited by the type and amount of knowledge we can obtain.

What needs to be developed are multi-modal AI/ML/DL tools that can analyze these orthogonally related data sets simultaneously. I refer to this burgeoning field as pathology data science. While pathologist already integrate molecular data into their pathological visual interpretation of the tumor; pathologists of the future— pathology data scientists, will begin to leverage AI/ML/DL algorithms to synthesize and interpret the pixel level data in a pathology image with the pixel level data of a radiology image with -Omics data as one potential diagnostic source as a CDx, rather than having separate types of analyses.  In addition, this new approach will place demands on pathology labs for slide staining quality that is fitted for digital and computational analysis rather than visual assessment by pathologists.

“When thinking of the future, one can find even more data with a higher likelihood of finding more predictive variables of drug response by integrating digital pathology images with orthogonally related radiology images.”

In summary, I believe the future role of imaging and the pathology data scientist in drug discovery will be in the ability to analyze and synthesize the pixels of tumor images at all length scales (pathologically, radiologically) and its synthesis with high throughput molecular data derived from the same tumor, quantify perturbations in the TME and potentially predict various types of clinical variables and drug response as a CDx that is currently not possible. Pathology data scientists will play a critical role in developing and applying this technology as they sit at the intersection of these data sets and are the only discipline (not radiology or molecular alone) to diagnose cancer alone.