Isoform Markers of Brain Aging

This project investigates whether isoform usage patterns can serve as molecular markers of human brain aging. Using bulk RNA-seq datasets from human brain regions, I developed isoform-level machine learning workflows to identify both tissue-agnostic and hippocampus-specific transcript usage markers associated with age.

The project was conducted in the Hicks Lab at Johns Hopkins University, Department of Biostatistics, and was presented as a poster at Biology of Genomes, Cold Spring Harbor Laboratory, 2026 (Poster PDF) (Chen et al., 2026).

Overview

Most transcriptomic aging studies focus on gene-level expression. However, genes can produce multiple transcript isoforms, and the relative usage of these isoforms may change across development and aging. This project asks whether isoform fractions provide useful and interpretable features for age prediction and aging marker discovery.

For each gene, I represented transcript usage using isoform fraction features, focusing on the relative abundance of major isoforms. I then trained and evaluated machine learning models to identify age-associated isoform usage patterns and compare their transferability across brain regions.

Key Findings

• Isoform fraction features captured age-associated transcript usage patterns beyond gene-level expression.
• Some isoform aging signals appeared transferable across brain regions, suggesting tissue-agnostic markers.
• Hippocampus-specific feature selection identified additional transcript usage markers that were not prioritized by models trained in other brain regions.
• Cross-region validation showed that aging-associated isoform usage can be partially shared but also region-specific.

My Contributions

• Built isoform-level machine learning models for human brain aging using bulk RNA-seq data.
• Designed feature engineering workflows based on isoform fraction features.
• Compared tissue-agnostic versus hippocampus-specific transcript usage markers.
• Performed feature selection, model evaluation, and cross-dataset validation.
• Generated publication-style figures and presented the project at Biology of Genomes 2026.

Methods

The analysis combined transcript quantification, isoform fraction construction, feature selection, and predictive modeling.

Key steps included:

1. Transcript-level quantification
   RNA-seq reads were quantified at the isoform level and summarized into transcript usage features.

2. Isoform fraction feature construction
   For selected genes, I calculated the relative usage of major isoforms to capture transcript switching behavior.

3. Feature selection
   I identified candidate aging-associated isoform markers using statistical and machine learning-based screening strategies.

4. Age prediction modeling
   I trained regularized regression models to evaluate whether isoform usage features could predict chronological age.

5. Cross-region validation
   I compared markers across brain regions to distinguish tissue-agnostic signals from hippocampus-specific aging patterns.

Poster

Selected Figures

Tools

Python · Jupyter · bulk RNA-seq · machine learning · elastic net regression · feature selection · transcriptomics

Status

This project is ongoing as part of my research in the Hicks Lab. Preprint coming soon in Fall 2026.