Multi-Omics Integration with Autoencoders and Graph Neural Networks

Overview

Open-source computational framework for ranking candidate biosynthetic-pathway genes by integrating transcriptomic and metabolomic data through deep learning, with very few samples (n = 21). The system combines statistical correlation, autoencoder and graph-attention importance, and gene-family homology into one multi-evidence ranking to reduce false positives among candidates.

Key Idea: combine several kinds of evidence — correlation, autoencoder importance, graph-attention importance and gene-family evidence — into one candidate ranking, rather than relying on a single method.

Problem Statement

Discovering genes responsible for metabolite biosynthesis is challenging due to:

• Ultra-high dimensionality: Genomic (100K+ features) and metabolomic (10K+ features) data vastly exceed sample sizes
• Small sample regime: Typical cohorts have n=20-100 samples due to cost constraints
• False positive risk: Single-method approaches lack cross-validation across evidence types
• Nonlinear relationships: Gene-metabolite associations involve complex regulatory cascades

Pipeline Architecture (5 Stages)

Stage 1 — Data Preparation

• Log2 transformation of FPKM values + z-score standardization
• Metabolite intensity normalization
• Quality control visualizations and sample metadata generation

Stage 2 — Model Training

• Independent gene and metabolite autoencoder training → 64-dimensional latent representations
• Graph Attention Network (GAT) training on concatenated latent vectors with auxiliary tissue/geographic classification tasks

Stage 3 — Analysis & Visualization

• Dimensionality reduction: t-SNE, UMAP, PCA on latent embeddings
• Gene-metabolite correlation matrices across all sample conditions

Stage 4 — Feature Importance Extraction

Three complementary importance metrics:

• Reconstruction importance: Per-gene MSE contribution from autoencoder
• Gradient-based importance: Backpropagation through encoder layers
• GNN-to-gene mapping: Projects latent-space importance back to individual genes via correlation weights

Stage 5 — Multi-Evidence Ranking

Weighted rank normalization combining all evidence:

• 0.4 × Correlation score
• 0.3 × Autoencoder importance
• 0.2 × GNN importance
• 0.1 × BLAST/HMM homology bonus

Produces ranked gene lists with supporting documentation per gene family.

Methodology

Dual Autoencoder Architecture

Gene Autoencoder (545M parameters):

• Encoder: ~132K → 2,048 → 512 → 128 → 64 (symmetric decoder)
• Regularization: latent L2 penalty, early stopping (patience 15) on a single 80/20 split

Metabolite Autoencoder (30.9M parameters):

• Encoder: ~7K → 2,048 → 512 → 128 → 64 (symmetric decoder)
• Regularization: latent L2 penalty, early stopping (patience 15) on a single 80/20 split

Graph Attention Network

• Operates on concatenated 128-dim latent vectors (64 gene + 64 metabolite)
• Multi-task classification for tissue/geographic metadata
• Classification accuracy is training accuracy on a transductive graph, so it does not by itself validate the latent representations

Multi-Evidence Scoring

• Rank-based normalization ensures fair contribution from evidence types at different scales
• Latent-correlation mapping preserves interpretability back to biological features
• Gene family integration supports post-hoc BLAST/HMM result incorporation

Key Features

• Modular pipeline: Run individual stages or complete workflows
• Hardware abstraction: Automatic detection of MPS (Apple Silicon), CUDA, or CPU
• Gene family integration: Post-hoc BLAST/HMM incorporation with batch processing
• Configurable weights: Customizable evidence combination via YAML config
• Checkpoint system: Intermediate model state preservation
• Publication-quality output: 300 DPI figures, CSV tables, Excel-formatted rankings

Results

Model Performance

• Gene AE compresses ~132K genes to a 64-d latent under a ~6.6K:1 feature-to-sample ratio
• Metabolite AE compresses ~7K metabolites to a 64-d latent
• Sample-level GAT uses the latents for auxiliary tissue / origin classification

Design Rationale

1. Large-scale autoencoders with a latent L2 penalty and early stopping for an extreme feature-to-sample ratio (~6.6K:1)
2. Multi-evidence triangulation to reduce false positives relative to any single method
3. Rank-based fusion to combine heterogeneous evidence types

Technical Stack

Privacy & Ethics

• No biological identifiers: Sample names, patient IDs removed
• Feature anonymization: Specific gene/metabolite names withheld in public repository
• Aggregated reporting: Only statistical summaries and framework architecture shared
• Collaborative approval: Data provider explicitly approved open-source release

Limitations & Future Work

Current Limitations

• Small sample size (n=21): Limits generalizability; recommend n≥100 for production
• Specific biological context: Associations may not generalize across tissues/conditions
• Computational cost: 545M-parameter models require GPU acceleration

Future Directions

1. Scaling to larger cohorts (n≥100)
2. Transfer learning with public datasets (TCGA, GTEx)
3. Attention mechanisms for relation-specific interpretability
4. Multi-modal extension (proteomics, epigenomics, clinical variables)
5. Federated learning for multi-institution collaboration

Code Availability

GitHub Repository: omics-ae-gnn

Note: Specific biological data not included to protect collaborator privacy. Framework designed for generalizability across datasets.

Project Status: Completed (December 2025 - February 2026) Technologies: PyTorch, GAT, Python Category: AI for Health, Computational Biology, Biosynthetic Pathway Gene Ranking