Research Overview
Our research addresses the challenging diagnosis of Myalgic Encephalomyelitis/Chronic Fatigue Syndrome (ME/CFS) and Long COVID using blood DNA methylation patterns. While these conditions affect millions worldwide, current clinical methods lack reliable molecular biomarkers.
ME/CFS & Long COVID
Complex chronic conditions affecting millions globally with overlapping symptoms and limited diagnostic tools.
Epigenomic Approach
Leveraging DNA methylation patterns as biomarkers for disease detection and classification.
AI-Powered Diagnostics
Using advanced deep learning to detect subtle, yet genome-wide, epigenetic changes.
The Challenge
ME/CFS and Long COVID present significant diagnostic challenges for clinicians:
- Diagnosis relies heavily on symptom reporting and exclusion of other conditions
- No established laboratory tests or biomarkers for objective diagnosis
- Significant overlap in symptoms between these conditions and other disorders
- Heterogeneous patient populations with varying symptom presentations
ME/CFS affects an estimated 2.5 million Americans and 17-24 million people worldwide. Long COVID may affect up to 30% of individuals who contract COVID-19, representing a significant global health burden.
Our Approach
Our work leverages advanced deep learning—in particular, a transformer-based pipeline—to detect subtle, yet genome-wide, epigenetic changes. Key innovations include:
Self-Supervised Masked Pretraining
We pretrain our model on unlabeled methylation data by masking random CpG sites and training the network to reconstruct these missing values, enabling it to learn the inherent structure of methylation patterns.
Mixture-of-Experts (MoE) Feed-Forward Network
Our architecture implements specialized "expert" networks that dynamically process different aspects of methylation data, allowing the model to develop specialized pathways for different epigenetic patterns.
Adaptive Computation Time (ACT)
We incorporate an ACT mechanism that allows the model to iteratively refine predictions for ambiguous cases by performing additional transformer passes, improving accuracy for difficult-to-classify samples.
Impact & Significance
This study not only achieves superior classification performance compared to traditional models but also provides interpretable insights into immune, metabolic, and stress-response pathways, supporting future targeted treatments.
By developing an objective, molecular-based diagnostic tool, we aim to:
- Reduce diagnostic delays and improve patient outcomes
- Provide clinicians with reliable biomarkers for ME/CFS and Long COVID
- Enable stratification of patient subgroups for personalized treatment approaches
- Advance understanding of the biological mechanisms underlying these conditions