Understanding how different cells contribute to disease is one of the biggest challenges in modern biology. Single-cell RNA sequencing gives us a close look at how individual cells behave, but it is still expensive and usually limited to small groups of patients. This means that many projects using single-cell techniques do not have the large, diverse clinical datasets needed to connect cellular behavior to real-world patient outcomes.

Researchers from the Research Institute of McGill University Health Centre, developed SIDISH, a computational framework designed to bridge this gap. SIDISH combines the detail of single-cell data with the scale of bulk RNA sequencing, which is far more affordable and widely available.

Methods Overview

SIDISH (Semi-supervised Iterative Deep Learning for Identifying Single-cell High-Risk Populations) overcomes these limitations by integrating scRNA-seq, bulk RNA-seq, and spatial transcriptomics through an advanced deep learning framework. By iteratively refining High-Risk cell predictions using Variational Autoencoders (VAE), Deep Cox Regression, and SHAP-based feature selection, SIDISH uncovers cellular subpopulations linked to poor survival while enabling robust patient-level risk assessment. In addition to identifying High-Risk cells, SIDISH extends to spatial data, allowing the discovery and localization of malignant subpopulations within their tissue context. SIDISH also employs in silico perturbation to simulate gene knockouts, ranking potential therapeutic targets based on their impact on disease progression. This combined ability—disease risk assessment, spatial mapping, and therapeutic prioritization—positions SIDISH as a transformative tool in precision medicine, biomarker discovery, and drug development.

SIDISH uses a set of advanced neural network tools, including a variational autoencoder and deep Cox regression, to learn which cell types are linked to poor outcomes. It can analyze bulk RNA data from large patient cohorts and still make predictions that capture single-cell level biology. This means scientists can identify high-risk cellular populations without needing to generate expensive single-cell datasets for every patient group.

One of SIDISH’s most powerful features is its ability to run virtual experiments known as in silico perturbations. By simulating how removing or modifying certain high-risk cells might affect a disease, SIDISH highlights promising therapeutic targets for future drug development.

The framework also works on spatial transcriptomics data, which shows not only what genes cells express but where those cells sit within a tissue. This allows SIDISH to pinpoint high-risk cells and map them in their natural environment, giving deeper insight into how disease spreads or persists.

Because SIDISH works across many diseases and data types, it offers a scalable path toward precision medicine. By linking cellular behavior to clinical outcomes, it brings scientists closer to finding meaningful biomarkers and identifying treatments tailored to individual patients.

Availability – Code for the models and results reproduction is publicly available on GitHub: https://github.com/mcgilldinglab/SIDISH.