A deep-learning model for characterizing tumor heterogeneity using patient-derived organoids.

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Cancer is characterized by genotypic and phenotypic diversity, leading to complex inter- and intratumor heterogeneity.
• Analyzing this heterogeneity is crucial for precision medicine and effective cancer treatment strategies.
• Current analytical approaches often struggle with the complexity introduced by tumor heterogeneity.

Purpose of the Study:

• To develop and validate a sequential deep learning model for analyzing cancer heterogeneity.
• To link patient-specific tumor morphological features to gene expression profiles.
• To identify gene subsets relevant for distinguishing between different tumors and inform personalized cancer care.

Main Methods:

• Utilized patient-derived organoids (PDOs) to model tumor heterogeneity.
• Employed a sequential deep learning approach with preprocessing to manage data complexity.
• Characterized morphological heterogeneity using microscopy images.
• Integrated morphological data with PDO gene expression data.

Main Results:

• The sequential deep learning model successfully organized complex heterogeneous data.
• Identified specific morphological features of PDOs that correlate with their origin.
• Extracted gene subsets strongly associated with intertumor differences by linking morphology and gene expression.
• Validated the relevance of selected genes for potential clinical applications.

Conclusions:

• The developed sequential deep learning model offers an efficient method for analyzing cancer heterogeneity.
• Linking morphological features to gene expression in PDOs is a viable strategy for identifying key cancer-related genes.
• Findings support the application of this approach in preclinical cancer studies and personalized clinical care.