Researchers at McGill University have developed an AI tool that can identify rare groups of cells responsible for driving aggressive cancer progression. The tool, called SIDISH, offers a new way to distinguish which cells within a tumour are most strongly linked to poor patient outcomes. Thuis is an important step toward more precise and effective treatments.
The findings demonstrate how SIDISH successfully identified high-risk cell populations in pancreatic, breast and lung cancer samples. These samples were derived from patients and analysed in a preclinical setting. Unlike conventional approaches that treat tumours as relatively uniform, SIDISH focuses on the biological diversity within tumours, where even small populations of cells can significantly influence disease progression.
Challenge
A central challenge in cancer research has been linking highly detailed single-cell data with real-world patient outcomes. Single-cell datasets typically provide deep biological insights but are often limited in scale and lack clinical follow-up data. Conversely, patient-level datasets include survival outcomes but average signals across millions of cells, masking the impact of rare but dangerous cell populations.
“By connecting these two data types, we can better understand which cells are truly driving disease progression,” said Yasmin Jolasun, first author of the study. SIDISH uses a semi-supervised deep learning approach to bridge this gap, enabling researchers to link cellular behaviour directly to clinical outcomes. The research was published in Nature Communications.
Targeted therapies
Beyond identifying high-risk cells, SIDISH also has the capability to simulate how these cells respond to genetic changes, such as switching specific genes on or off. This functionality allows researchers to predict which genes may serve as effective drug targets, thus potentially accelerating the development of new therapies.
According to senior author Jun Ding, this could help address one of the major bottlenecks in drug development: identifying viable targets. “Instead of relying solely on years of experimental testing, we can now prioritise the most promising options computationally,” he explained.
In practice, this approach could support a more personalised treatment pathway. For example, tumour samples from a patient could be analysed using single-cell sequencing, after which SIDISH would identify the most aggressive cell populations. The tool could then simulate how these cells respond to different therapeutic strategies, generating a shortlist of treatments most likely to be effective.
Beyond oncology
Although SIDISH is still in development and not yet used in clinical care, its potential applications extend beyond oncology. The same methodology could be applied to other complex diseases where variability between cells plays a crucial role, such as neurological or immune-related disorders.
In the short term, the researchers suggest the tool could support the repurposing of existing, approved drugs by analysing publicly available datasets. Over the longer term, SIDISH may contribute to a fundamental shift in how new drugs are discovered and developed.
The research team is currently expanding the use of SIDISH to additional disease areas and collaborating with industry partners to further refine the technology. As AI continues to reshape biomedical research, tools like SIDISH highlight the growing potential of data-driven approaches to unlock more targeted and effective healthcare solutions.
AI in cancer drug development
Last year, researchers at Institute of Cancer Research developed an AI technology that predicts how cancer cells respond to drugs by analysing changes in their 3D shape. Using deep learning and nearly 100,000 3D images of melanoma cells, the system identified drug responses with 99.3% accuracy, outperforming traditional 2D imaging methods.
The approach enables faster and more precise drug testing, potentially accelerating development timelines by up to six years and reducing the preclinical phase from three years to just three months. Beyond melanoma, the technology has shown potential across other cell types, suggesting broad applicability. It could also improve targeted therapies and make clinical trials more efficient. The research, published in Cell Systems, is being advanced through collaborations and the spin-off Sentinal4D.
