Researchers have developed an artificial intelligence (AI) tool that can identify small groups of cells most responsible for driving aggressive cancers. The tool, called SIDISH, offers scientists a clearer path to designing targeted therapies by showing which cells inside a tumor are most strongly linked with poor patient outcomes—rather than treating all cancer cells as if they behave the same way.
SIDISH stands for semi-supervised iterative deep learning for identifying single-cell high-risk populations.
In a preclinical study published by Jolasun et al in Nature Communications, investigators showed SIDISH successfully identified “high-risk” cells across pancreatic, breast, and lung cancers using tumor samples collected from patients and analyzed in the lab.
How the Tool Works
SIDISH’s key innovation is that it connects what happens inside individual cells with patient outcomes, a long-standing challenge in cancer research.
“Single-cell data is very detailed, but it usually comes from only a few patients and rarely includes how those patients actually fared. Patient data, on the other hand, often at the bulk level, includes survival information but averages signals from millions of cells, hiding the rare but dangerous ones that drive disease,” explained first author Yasmin Jolasun, a PhD student in McGill University’s Department of Medicine.
Existing computational tools have struggled to meaningfully bring these two types of data together.
“Our tool builds a bridge between both worlds. It can identify which cells are most strongly associated with faster disease progression and patient survival outcomes,” said Ms. Jolasun.
While SIDISH was tested first in cancers, the same approach could be applied to other complex diseases where cell-to-cell differences play a major role, she added.
Beyond identifying the problem, SIDISH can also simulate how high-risk cells respond when specific genes are turned on or off, helping predict which genes might be promising drug targets.
“This could ease a major bottleneck in drug development, where finding the right targets often requires years of trial-and-error testing,” said senior author Jun Ding, PhD, Assistant Professor in McGill’s Department of Medicine and a junior scientist at the Research Institute of the McGill University Health Centre.
For example, he said, a patient’s tumor could be analyzed with single cell-sequencing; SIDISH would then identify the cells driving that tumor and simulate how they respond to different drugs, generating a short list of treatments most likely to be effective.
“In the short term, SIDISH could help us repurpose existing U.S. Food and Drug Administration–approved drugs using public datasets. In the long term, it has the potential to fundamentally change how new drugs are discovered,” said Dr. Ding.
The work remains in development and is not yet used in clinical care. The research team is now applying SIDISH to additional diseases and collaborating with industry partners to further refine the approach.
DISCLOSURE: For full disclosures of the study authors, visit nature.com.
