Project Ex Vivo header - aerial view of lab tech working with a pipette

Project Ex Vivo

Modeling cancer outside the body specific to the patient

Project Ex Vivo is a joint cancer research collaboration between Microsoft and the Broad Institute of MIT and Harvard. The project spans multiple research areas, including machine learning, statistics, bioengineering, automation, molecular biology, and clinical application.

Cancer is understood primarily as a disease of DNA alterations and often treated using a reductionist approach, with each cancer distilled to a single metric such as histologic type or harboring a specific molecular abnormality. As a result, the current paradigm for precision oncology consists primarily of using genetic alterations to direct therapy (e.g., targeting tumor mutations). Strikingly, few patients receive prolonged benefit from this approach.

In Project Ex Vivo, we see cancers as complex (eco)systems, beyond just mutational variation, that necessitate systems-level understanding and intervention. Our ultimate objective is to more effectively model cancer ex vivo – outside the body – in a patient-specific manner. Doing so will unlock the ability to more effectively stratify patients and identify therapies that target diverse aspects of human cancers.

We use machine learning to develop and understand accurate representations of each tumor by integrating genetic markers, expression state, and tumor microenvironmental interactions. These representations help us precisely define and quantify the state and trajectory of each tumor in each patient. One of the benefits of this approach is that it gives us a clear measure of divergence between biological models (e.g., cell cultures, organoids) and the original tumor. We can directly target this divergence and minimize it using reinforcement learning and lab automation techniques to produce cellular models that are high-fidelity reproductions of the original tumor.

Our initial work in pancreatic ductal adenocarcinoma (opens in new tab) (PDAC) uncovered limitations in current cancer modeling pipelines and revealed fundamental insights into the therapeutic significance of cell state and its microenvironmental drivers.

Project Ex Vivo will codify new pipelines and workflows to establish non-genetic attributes (e.g., cell state) as targetable features in cancer, with an initial focus on pancreatic cancer and eventually extending to additional malignancies. We use machine learning to develop and understand accurate representations of each tumor and cellular model by integrating genetic markers, expression state, and tumor microenvironmental interactions. We will optimize model generation workflows for greater in vivo fidelity, nominate new candidates for therapeutic development, and establish a new paradigm integrating genetic and non-genetic attributes to guide therapy selection in precision oncology. We are developing rigorous experimental, computational, and automation pipelines that can be applied across a variety of cancer, non-cancer, and clinical contexts. By credentialing cell state and its relation to the local environment as a biomarker for both prognostication and therapy selection, the success of this project may improve patient stratification for diverse clinical trials (chemotherapy, immunotherapy, small molecules, etc.), thereby enhancing the likelihood of therapeutic efficacy, potentially lowering costs, and ultimately improving patient outcomes.