Deep Learning Implicitly Handles Tissue Specific Phenomena to Predict Tumor DNA Accessibility and Immune Activity

Area of Science:

• Genomics
• Cancer Biology
• Computational Biology

Background:

• DNA accessibility is crucial for gene regulation and tumor development.
• Predicting chromatin state from DNA sequence is challenging.
• Integrating gene expression data can enhance predictive models.

Purpose of the Study:

• To improve neural network models for predicting DNA accessibility using DNA sequence and RNA sequencing gene expression.
• To apply the improved model to analyze DNA accessibility landscapes in The Cancer Genome Atlas (TCGA).
• To investigate the relationship between DNA accessibility, immune pathways, and patient prognosis in lung adenocarcinoma and other cancers.

Main Methods:

• Enhanced neural network models to predict DNA accessibility from DNA sequence.
• Incorporated global RNA sequencing gene expression data into the models.
• Applied the expression-informed model to analyze promoter accessibility across TCGA data.
• Investigated correlations between DNA accessibility, immune pathways, and clinical outcomes.

Main Results:

• The expression-informed model achieved high accuracy in predicting DNA accessibility, extending applicability to new tissue types.
• Analysis of lung adenocarcinoma revealed an inverse correlation between DNA accessibility and immune pathways.
• DNA accessibility patterns from a single tumor type could predict immune inflammation across diverse cancers.
• Accessibility patterns showed a relationship with patient prognosis.

Conclusions:

• Improved neural network models integrating gene expression enhance DNA accessibility prediction.
• DNA accessibility analysis provides unique insights into immune responses in cancer.
• This approach can identify biomarkers for immune inflammation and patient prognosis across multiple cancer types.