Sohrab Shah, PhD

Developing AI models to improve prediction of breast cancer recurrence risk and treatment response.

Impact

Artificial intelligence models trained on biological and medical data hold immense potential for assisting physicians in understanding and predicting patient outcomes. However, to ensure safety and utility, these AI models must be trained and deployed with the utmost care. Dr. Shah and his team have recently discovered that AI models trained on multimodal data –encompassing the tissue, cellular, and molecular levels—outperform models trained on a single data type. Dr. Shah aims to develop AI models using such multimodal data to predict the risk of recurrence after treatment in patients newly diagnosed with breast cancer and to assess the drivers of treatment resistance in metastatic breast cancer.

Progress Thus Far

Over the past year, Dr. Shah and his team made progress in using AI to improve how breast cancer is diagnosed and treated. They developed an AI model called Orpheus that can predict a patient’s risk of cancer recurrence by analyzing more than 4,000 tumor images from early breast cancer patients. Orpheus accurately identified high-risk patients across multiple hospitals and even performed better than a leading commercial test for identifying patients unlikely to develop metastatic disease. Unlike many AI tools, Orpheus is interpretable, meaning that clinicians can identify specific biological features, such as immune cell activity and tumor growth rates, that are driving the predictions. Building on this, the team trained more advanced AI models that combine tumor images with genetic and clinical data to predict whether chemotherapy given before surgery is likely to eliminate a patient’s cancer. These combined models outperformed models trained on any single data type, though image-based analysis alone gave the strongest individual result of the data types considered.

In parallel, the team created new computational tools to improve treatment for patients with HER2-positive metastatic breast cancer. One model uses tumor data to predict which tumors will respond best to a class of drugs called antibody-drug conjugates (ADCs), especially those with low HER2 levels who might otherwise not receive this therapy. They also studied why some tumors stop responding to ADCs by combining advanced imaging and gene expression data, uncovering patterns that reveal how cancers adapt to resist therapy. By mapping gene activity within the tumor’s architecture and comparing tissue before and after treatment, the team confirmed that resistant tumors often lose the very proteins these drugs are designed to target, while activating alternative survival pathways and immune-suppressing programs of their own.

What’s next

Dr. Shah and his team will expand their AI-based recurrence risk models by integrating images that reveal different features of a tumor for a richer, more complete analysis. This fusion approach is expected to improve the accuracy of stratifying early breast cancer patients by recurrence risk, enabling more precise identification of high-risk patients for tailored treatment or monitoring.

The team will also employ their computational tool, GRAFITI, to define detailed spatial signatures within metastatic breast tumors treated with ADCs, in order to more deeply understand the molecular mechanisms driving resistance. This work will help determine whether resistant cancer cells existed even before treatment began and simply expanded through a selection process, or whether resistance instead develops as cells adapt their gene activity over the course of treatment.