TrajOmics: An Omni-Purpose Trajectory Mining Pipeline for Multi-Omics Longitudinal Analysis

Overview

TrajOmics is an AI-driven framework for analyzing disease trajectories using longitudinal multi-omics data. It combines non-linear dimensionality reduction with trajectory mining to uncover complex disease progression patterns. TrajOmics is designed to handle high-dimensional, heterogeneous data while preserving key molecular drivers, offering a novel approach to disease subtyping and progression modeling.

---

Architecture

---

Key Features

• High-Quality Dimensionality Reduction: Utilizes Locally Linear Embedding (LLE) for robust, non-linear dimensionality reduction.
• Feature Selection & Hyperparameter Optimization: Integrates statistical feature selection with hyperparameter tuning for optimal model performance.
• Network-Based Filtering: Leverages Sparse Generalized Tensor Canonical Correlation Analysis Network (SGTCCA-Net) to capture complex feature relationships.
• Trajectory Clustering: Identifies distinct disease progression patterns, enabling personalized disease management.
• Cross-Validation and Robustness Testing: Validates the stability of the generated manifold across multiple visits and sampling conditions.

---

Motivation

Longitudinal multi-omics datasets present unique challenges:

• High Dimensionality and Heterogeneity: Risk of overfitting and poor generalization with traditional methods.
• Capturing Molecular Complexity: Existing models often overlook key molecular drivers.
• Scalability and Computational Challenges: Difficulty in scaling to large datasets.
• Dynamic Temporal Complexity: Challenges in modeling disease progression over time.

TrajOmics addresses these challenges by combining advanced dimensionality reduction with robust feature selection and trajectory mining, enabling more accurate disease subtyping and progression modeling.

---

Methodology

1. Dimensionality Reduction

• Non-Linear Manifold Learning: Locally Linear Embedding (LLE) is used to capture complex, non-linear relationships.
• Hyperparameter Optimization: Fine-tuning the number of neighbors and components for optimal manifold construction.

2. Feature Selection and Network-Based Filtering

• Statistical Feature Selection:
  • Variance Threshold
  • Random Forest
  • Forward Feature Selection
  • LASSO
  • Information Gain
  • Fischer Score
• Network Inference: Sparse Generalized Tensor Canonical Correlation Analysis Network (SGTCCA-Net) to enhance feature relevance.

3. Trajectory Clustering

• Identifies distinct disease progression patterns.
• Captures temporal dynamics to enhance disease modeling and personalized healthcare.

---

Dataset

Chronic Obstructive Pulmonary Disease (COPD) dataset from the Subpopulations and Intermediate Markers in COPD Study (SPIROMICS):

• Longitudinal multi-omics data
• Captures disease progression over multiple visits
• Ideal for trajectory mining and subtyping

---

Installation

git clone https://github.com/<your-username>/TrajOmics.git
cd TrajOmics
pip install -r requirements.txt