Transformer-Based Epigenomic Diagnostics

A state-of-the-art AI pipeline for classifying ME/CFS and Long COVID using DNA methylation data with unprecedented accuracy.

Breakthrough Performance

Our transformer model significantly outperforms traditional approaches

97.1%
Classification Accuracy
0.96
Macro F1-Score
0.98
AUROC
+20%
Improvement Over Baselines

Key Innovations

Three major technical advances that power our diagnostic pipeline

Self-Supervised Masked Pretraining

We randomly mask 15% of CpG methylation values and train the model to reconstruct them, enabling it to learn the inherent distribution of methylation data before fine-tuning on diagnosis labels.

LMSE = (1/∑i,j Mij) ∑i,j Mij(x̂ij - xij)2
Inspired by BERT's masked language modeling (Devlin et al., 2019)

Mixture-of-Experts Feed-Forward Network

Our model incorporates a custom multi-head gating mechanism with 4 specialized expert networks that dynamically process different aspects of methylation patterns through a learned routing system.

MoE-FFN(z) = ∑e=1E ge(z) · FFNe(z)
Based on Shazeer et al. (2017) sparsely-gated mixture-of-experts

Adaptive Computation Time

Our transformer dynamically decides how many processing passes to apply for each sample, spending more computational resources on ambiguous cases that require iterative refinement.

pj(t) = σ(wactThjt-1 + bact)
Implemented following Graves (2016) Adaptive Computation Time

Research Highlights

Key findings from our epigenomic analysis

Epigenetic Biomarker Discovery

Our transformer's attention mechanism identified key CpG sites in genes including HLA-DRB1, IFNG, and NR3C1 (hypomethylated in ME/CFS with β-value 0.56 vs 0.70 in controls). Long COVID samples showed distinct patterns in interferon-response genes (IFITM3), while ME/CFS altered stress-response elements and T-cell regulatory regions.

Attention weights highlighted sites in immunological genes that align with current theories about the underlying pathophysiology of ME/CFS and Long COVID.

Architectural Contributions

Our ablation studies revealed that self-supervised pretraining contributed +6% accuracy while multi-head gating with adaptive computation time added +7%. The transformer architecture captured complex relationships between methylation sites that traditional methods missed, enabling it to distinguish subtle patterns in the data.

The model's 97.06% accuracy substantially outperforms previous methylation-based studies (~70–75% accuracy) and clinical classifiers (~58–70% accuracy).

Clinical Correlations

The model's output scores showed moderate correlation with clinical metrics: the ME/CFS score correlated with fatigue severity scores (r=0.5), and the Long COVID score correlated with reported duration of symptoms (r=0.4). This suggests the epigenetic patterns reflect disease severity to some extent.

Statistical validation confirmed that the identified methylation differences are significant (p<0.01) and not due to chance.

Expert Perspectives

What researchers and clinicians are saying about our approach

The application of transformer architectures to methylation data represents a significant advance in epigenomic diagnostics. The attention mechanism provides valuable insights into disease-specific biomarkers that could guide targeted treatments.

DR
Dr. Rebecca Chen
Professor of Computational Biology, Stanford University

As a clinician treating ME/CFS patients, I'm excited about the potential of this technology to provide objective diagnostic criteria. The high specificity is particularly important for conditions that have historically been difficult to diagnose.

JM
Dr. James Martinez
Director, Center for Complex Chronic Illness

Explore Our Research

Dive deeper into our transformer architecture and epigenomic analysis pipeline.