Related Experiment Video
Updated: Jun 25, 2026
09:53
Quantifying the Brain Metastatic Tumor Micro-Environment using an Organ-On-A Chip 3D Model, Machine Learning, and Confocal Tomography
Published on: August 16, 2020
Deep Unsupervised Domain Adaptation for Translating Cancer Dependency Maps From Cell Lines to Breast Cancer Tumor
Yu Shi1, Wei Xu1,2, Pingzhao Hu1,3,4,5,6
1Biostatistics Division, Dalla Lana School of Public Health, University of Toronto, Toronto, Ontario, Canada.
Genetic Epidemiology
|June 24, 2026
Summary
We developed a deep unsupervised domain adaptation algorithm to predict cancer dependencies in patient tumors. This approach improves the translation of preclinical findings into personalized cancer treatments, identifying potential drug targets for breast cancer.
Area of Science:
- Computational biology
- Genomics
- Precision medicine
Background:
- Cancer dependency maps (DepMap) identify genetic vulnerabilities using loss-of-function screens.
- Discrepancies between cancer cell line models and patient tumors hinder clinical translation.
- Artificial intelligence, specifically domain adaptation, can bridge this gap by aligning molecular data.
Purpose of the Study:
- To develop and validate a deep unsupervised domain adaptation (UDA) algorithm for aligning cancer cell line and patient tumor data.
- To predict cancer dependency maps for breast cancer (BC) patients using The Cancer Genome Atlas (TCGA) data.
- To assess the utility of predicted dependency maps for subtype classification and synthetic lethality (SL) discovery.
Main Methods:
- Trained a deep UDA algorithm on labeled cancer cell line data (source domain) and unlabeled patient data (target domain).
- Applied the trained model to predict BC dependency maps from TCGA data.
- Validated the model by classifying ER+/HER2+ BC subtypes and identifying SL gene pairs.
Main Results:
- The UDA model accurately predicted cancer dependency maps for patient-derived tumors.
- The predicted maps achieved high accuracy (AUC-ROC 0.99) in classifying ER+/HER2+ BC subtypes.
- Identified two potential synthetic lethality gene pairs (PBRM1-NF2 and PBRM1-CTNND2) for precision therapy development.
Conclusions:
- Deep unsupervised domain adaptation is a powerful approach for transferring biological knowledge between cancer models.
- This method enhances the translation of preclinical findings into patient-specific treatment strategies.
- The identified SL pairs offer potential therapeutic targets for ER+/HER2+ breast cancer.
