ӳ����ý

Ductal carcinoma in situ (DCIS), a pre-invasive tumor, accounts for about 25 percent of breast cancer diagnoses, a leading cause of cancer death. While doctors generally recommend treatment, they lack the appropriate evidence to reliably decide which tumor will remain benign and which might turn into a life-threatening invasive ductal carcinoma (IDC), resulting in high rates of overtreatment.

The current methods for understanding DCIS progression include manual assessment of nuclear morphology by pathologists, sequencing-based approaches, spatial transcriptomics, and highly multiplexed imaging. However, these methods face challenges due to cost, complexity, and limited information about the tissue microenvironment, which is necessary for accurate DCIS progression assessment.

In a new study published today in , researchers at the ӳ����ý of MIT and Harvard and the Paul Scherrer Institute at ETH Zürich in Switzerland have found a simple and effective method of predicting the disease stage of DCIS, which could ultimately lead to more informed recommendations for DCIS breast cancer treatment. Their analysis demonstrates that, without the need of multiple stains or sequencing-based technologies, chromatin imaging provides sufficient information about cell states and tissue organization to accurately predict tumor stages.

The study stems from a long-term collaboration combining AI and biology between Caroline Uhler, who directs the at the ӳ����ý, and is a professor in the Department of Electrical Engineering and Computer Science as well as the Institute for Data, Systems, and Society at MIT; and GV Shivashankar, professor of mechanogenomics and head of the Laboratory of Nanoscale Biology at the Paul Scherrer Institute.

Shivashankar’s lab is interested in understanding the underlying mechanisms for cell-state transitions and the association with disease states. They aim to improve early disease diagnostics by using multi-disciplinary approaches, such as single-cell imaging, functional genomics, and machine learning, to study the coupling between cell mechanics and genome organization in tissue contexts.

“Building on our previous studies with the Schmidt Center, we’re thrilled that we found a simple way to predict disease stage through the statistics of cell states, and we look forward to seeing how this can be applied to DCIS treatment,” said co-senior author Shivashankar.

The study aligns with the Schmidt Center’s goal of fostering a two-way street between biology and machine learning to advance biomedical discoveries and provide insights into how cells work in health and disease.

“As our research on DCIS shows, it’s important to create novel machine learning methods to analyze biomedical data,” said co-senior author Uhler. “Using machine learning to analyze data can lead to more accurate and simpler solutions for important biological questions, ultimately leading to better disease diagnosis and treatment.”

“Collaborating with Shivashankar’s lab, which I have done on several projects, provides me with a unique opportunity to develop computational methods for important biological problems,” said study first author Xinyi Zhang, a graduate student at MIT and the Schmidt Center. “I’m able to see what the real roadblocks and challenges are in the biomedical space and start thinking about what to develop next.”

Using unsupervised representation learning methods, the scientists analyzed 560 samples from 122 patients at 11 stages of DCIS progression from normal to cancerous breast tissues. They identified eight disease-relevant cell states based on nuclear morphology and chromatin organization, and found that all eight cell states exist in all disease stages, but with different abundances. 

Based on the learned representations, the researchers then arranged the cell types from healthy to cancerous, finding that the order matched the natural progression of the disease, even though the model wasn't trained directly on disease stages. The study also demonstrated that spatial organization of cells near breast ducts and the co-localization of cell states can better predict disease stage compared to cell state abundance alone. This approach highlighted distinct cell states, their relative abundances, and their spatial neighborhoods, indicating their potential as biomarkers for cancer staging.

Although follow-up clinical trials with longitudinal tracking of DCIS patients are needed, this study demonstrated that high-dimensional AI-inferred features based on simple and cheap chromatin images can provide valuable insights into tumor progression. Uhler noted that this study introduces a new approach to exploring disease progression within a tumor microenvironment, specifically by leveraging machine learning and computational methods to extract meaningful information from complex chromatin images, without the need for extensive staining or sequencing. 

By focusing on one of the Schmidt Center’s core missions – developing the foundations of machine learning to understand the programs of life – this study offers simple and cost-effective solutions for disease prognosis and treatment.

Adapted from .