DNAI
Back to Pipeline Overview
Treatment Designv1.0

Methylation Decoder

Reconstructs epigenetic patterns and detects silenced tumor suppressors

Architecture
MLP Decoder with Sigmoid Output
Reconstruction R²
0.519

Variance explained in CpG beta-value reconstruction

Target: > 0.45
PCA Ceiling
0.706

Theoretical max R² from 48d PCA (decoder captures 73.5%)

Target: Reference
High-R² Probes
559/1000

Probes with individual R² > 0.5

Target: > 500
TSG Panel
19 genes

Known tumor suppressors screened (MGMT, CDKN2A, MLH1, BRCA1, etc.)

Target: Curated

Overview

Decodes the VAE's 48-dimensional methylation latent back into 1,000 CpG probe beta-values, enabling detection of epigenetically silenced tumor suppressor genes. By cross-referencing decoded methylation with RNA expression, it identifies genes that are hypermethylated (promoter silencing) but not expressed — a hallmark of epigenetic tumor suppressor inactivation. Reports CIMP (CpG Island Methylator Phenotype) classification and screens 19 known tumor suppressor genes including MGMT, CDKN2A, MLH1, and BRCA1.

Inputs

2 inputs
z_meth48

Methylation latent from VAE (dims 248-296)

Source: vae
RNA expressionPer TSG gene

Gene expression for silencing cross-reference

Source: RNA-seq

Outputs

3 outputs
CpG beta-values1,000

Reconstructed methylation beta-values [0, 1]

Silenced TSGsUp to 19

Tumor suppressors with high methylation + low expression

CIMP statusClassification

CpG Island Methylator Phenotype (CIMP-H / CIMP-L / Non-CIMP)

Mathematical Formulation

Decoder

3-layer MLP with sigmoid output for beta-values in [0, 1]

Silencing Detection

Gene is silenced when hypermethylated AND underexpressed

Key Features

  • 48d → 1,000 CpG reconstruction via 3-layer MLP with sigmoid output
  • RNA cross-reference for functional silencing confirmation
  • 19 curated tumor suppressor genes screened
  • CIMP classification (CpG Island Methylator Phenotype)
  • Per-probe R² quality metrics for confidence assessment

Key Innovations

  • 1Latent-to-epigenome reconstruction enables methylation inference from VAE
  • 2Cross-modal validation (methylation + expression) for silencing detection
  • 3Captures 73.5% of theoretical PCA ceiling in 48 dimensions
  • 4Clinically relevant TSG panel includes actionable targets (MGMT for temozolomide)

Hyperparameters

Hidden Layers
[256, 128]
Activation
ReLU (hidden), Sigmoid (output)
Training Samples
2,357 TCGA (80/20 split)
Epochs
200
Training Time
12.8s (MPS)
Optimizer
Adam (lr=1e-3)

Training Details

Trained on 2,357 TCGA samples with matched methylation data. 80/20 train/val split, 200 epochs, MSE loss on beta-values. PCA-48d ceiling is R²=0.706, decoder achieves 0.519 (73.5% of ceiling). 559/1000 probes exceed R²>0.5, 18 probes exceed R²>0.8.

Pipeline Position

Multi-modal VAE
Methylation Decoder
Final Predictions