McGill's SIDISH AI Pinpoints High-Risk Cells Driving Aggressive Cancers

McGill University researchers have unveiled a groundbreaking artificial intelligence tool called SIDISH that promises to transform how we understand and combat aggressive cancers. Published in the prestigious journal Nature Communications, this preclinical study demonstrates SIDISH's ability to pinpoint rare, high-risk cancer cells responsible for rapid disease progression. By bridging the gap between detailed single-cell data and large-scale patient outcomes, the tool offers a new pathway for precision medicine, potentially accelerating the development of targeted therapies.

At the heart of this innovation is the recognition that tumors are not uniform masses but complex ecosystems where a small subset of cells drives malignancy. Traditional bulk RNA sequencing, while scalable for thousands of patients, averages signals across millions of cells, obscuring these dangerous outliers. Single-cell RNA sequencing (scRNA-seq), conversely, reveals intricate cellular heterogeneity but is limited by cost and small cohorts lacking survival data. SIDISH elegantly resolves this dichotomy, enabling scientists to identify the culprits behind poor prognoses in pancreatic ductal adenocarcinoma (PDAC), triple-negative breast cancer (TNBC), and lung adenocarcinoma (LUAD).

Canada's Cancer Burden: Why This Matters Now

Cancer remains a leading cause of death in Canada, with projections for 2026 estimating 254,100 new diagnoses and 87,900 fatalities. Lung, breast, prostate, and colorectal cancers will account for nearly half of cases, but aggressive subtypes like PDAC carry devastating mortality rates—over 80% of patients succumb within five years. Triple-negative breast cancer, lacking targeted therapies like HER2 inhibitors, and certain lung adenocarcinomas further exacerbate the challenge. These statistics underscore the urgency for tools like SIDISH, which could refine risk stratification and personalize treatments in a country where healthcare innovation is pivotal.

McGill University, nestled in Montreal's vibrant research ecosystem, stands at the forefront of this fight. Home to the Research Institute of the McGill University Health Centre (RI-MUHC), the institution has a storied history in oncology, now amplified by AI integration. This study exemplifies how Canadian higher education is leveraging computational biology to address national health crises, fostering collaborations that could lower incidence and improve survival rates.

The McGill Team Behind SIDISH

Leading the charge is Yasmin Jolasun, a PhD candidate in McGill's Department of Medicine, who serves as first author. Her work highlights the pivotal role of graduate students in cutting-edge research at Canadian universities. Senior author Jun Ding, an assistant professor in the same department and junior scientist at RI-MUHC, brings expertise in machine learning and quantitative life sciences. Collaborators include Kailu Song, Yumin Zheng, Jingtao Wang, Gregory J. Fonseca from Khalifa University, and David H. Eidelman, blending talents from McGill's Meakins-Christie Laboratories and beyond.

This multidisciplinary effort underscores McGill's strength in fostering environments where medicine, computer science, and AI converge. Ding's lab, affiliated with Mila-Quebec AI Institute, exemplifies how Quebec's AI hub is propelling health research forward.

Decoding Tumor Heterogeneity: The Core Challenge

Cancer's lethality often stems from intra-tumoral heterogeneity—diverse cell states within a single tumor. High-risk subpopulations, comprising just 5-10% of cells, evade therapies and fuel metastasis. Detecting them requires resolving single-cell granularity with clinical outcomes, a feat prior tools like Scissor or scAB struggled with due to linear assumptions or shared latent spaces diluting signals.

SIDISH addresses this through semi-supervised iterative deep learning. It unifies scRNA-seq's depth (revealing cell types, states) with bulk RNA-seq's breadth (thousands of patients' survival data from TCGA). This integration is crucial for Canadian researchers, where public datasets enable scalable validation without prohibitive costs.

How SIDISH Works: Step-by-Step Innovation

SIDISH operates in four iterative phases, refining predictions through bidirectional feedback:

• Phase 1: Feature Extraction - A variational autoencoder (VAE) processes scRNA-seq data, modeling dropout with zero-inflated negative binomial (ZINB) likelihood. For spatial transcriptomics, a graph convolutional network (GCN) incorporates neighborhood context.
• Phase 2: Survival Modeling - Transfer learning finetunes the VAE encoder into a deep Cox regression model on bulk data, predicting risk scores tied to patient survival.
• Phase 3: Risk Stratification - Scores fit a Weibull distribution; cells above a threshold (e.g., top 5%) are deemed high-risk.
• Phase 4: Iterative Refinement - SHAP values update gene/patient weights, prioritizing survival-relevant features, looping until convergence.

This process outperforms benchmarks, achieving superior concordance indices (C-Index) and survival stratification (log-rank P-values <10^{-15}). Code is available on GitHub for reproducibility, democratizing access for fellow researchers.

Photo by Faustina Okeke on Unsplash

Preclinical Breakthroughs in Pancreatic, Breast, and Lung Cancers

In PDAC (24 samples, 41,986 tumor cells + TCGA n=181), SIDISH flagged ~8.6% high-risk cells (mostly type 2 ductal), enriched in lipid metabolism and hypoxia pathways. Markers like KRT17 stratified independent cohorts (P=6.67×10^{-16}).

For TNBC (42,512 cells + TCGA n=1,194), 8.9% high-risk cancer epithelial cells upregulated ECM/angiogenesis genes; prognostic power confirmed (P=7.44×10^{-19}).

LUAD (4,102 cells + TCGA n=572) identified high-risk clusters with glycolytic signatures (LDHA, ENO1); validation P=5.40×10^{-19}.

Spatial analysis in PDAC (190,965 cells) localized 67.5% high-risk tumor cells, enhancing microenvironment insights.

Simulating Therapies: In Silico Perturbations for Drug Targets

SIDISH's perturbation module simulates gene knockouts using interaction networks (e.g., HIPPIE), ranking targets by high-risk cell reduction. Top hits: VEGFA (anti-angiogenic Bevacizumab), MAP2K1 (Trametinib) in LUAD; SPARC (Abraxane) in PDAC. Combinatorial effects (e.g., CDK1+BTG2) amplified outcomes, revealing patient-specific vulnerabilities.

This virtual screening could repurpose FDA-approved drugs swiftly, vital for Canada's oncology pipeline. For details on the study, visit the Nature Communications publication.

McGill's Growing Arsenal in AI Oncology

This builds on McGill's AI momentum: DOLPHIN (2025) detects exon/junction RNA markers beyond genes; earlier tools predict brain metastases via MRI. RI-MUHC's ecosystem positions McGill as Canada's AI-cancer leader, attracting talent and funding.

Broader Canadian efforts include U Toronto's AI for heart failure prediction and UBC's imaging AI, signaling a national push toward computational precision medicine.

Implications for Precision Medicine and Canadian Healthcare

"Our tool builds a bridge between both worlds," notes Jolasun. "It can identify which cells are most strongly associated with faster disease progression." Ding adds, "This could ease a major bottleneck in drug development."

In Canada, where pancreatic cancer claims lives rapidly, SIDISH could enable early intervention. Ethical AI use ensures bias mitigation, aligning with CIHR guidelines. For more on 2026 projections, see the CMAJ report.

Careers in AI-Cancer Research at Canadian Universities

This breakthrough highlights opportunities in computational oncology. McGill and peers seek experts in scRNA-seq, deep learning, and clinical translation. Roles span postdocs to faculty, with demand surging amid CIHR's AI-health investments.

Photo by Sakarie Mustafe Hidig on Unsplash

• Skills: Python, PyTorch, Scanpy, survival analysis.
• Impacts: Bridge academia-industry for startups like those from Mila.
• Outlook: Projected 20% growth in bioinformatics jobs by 2030.

Future Horizons: From Preclinical to Clinic

Ongoing refinements include multi-omics integration and real-time clinical deployment. Collaborations with pharma could fast-track trials. As spatial transcriptomics evolves, SIDISH may map high-risk niches in 3D, revolutionizing surgery and immunotherapy.

McGill's SIDISH exemplifies how Canadian higher education drives global health innovation, offering hope against aggressive cancers.