ML-Based RNA-Seq Classification for Autoimmune Disease Biomarker Discovery

Executive Summary

We developed a machine learning pipeline for classifying RNA-Seq gene expression data into three categories: Healthy, Rheumatoid Arthritis (RA), and Systemic Lupus Erythematosus (SLE). Our best-performing model, SGDClassifier with Elastic Net regularization, achieved 93.30% accuracy, outperforming more complex non-linear approaches. Key genes identified as potential biomarkers include ANP32D, IFI27, and MIR378F, offering insights into molecular mechanisms underlying these autoimmune conditions.

Introduction

This project aims to:

1. Classify individuals as healthy or having RA or SLE based on RNA-Seq gene expression profiles
2. Identify key genes that could serve as potential biomarkers for these conditions

RNA-Seq quantifies mRNA molecules in biological samples, reflecting gene expression levels and offering insights into disease mechanisms.

Methodology

Data Preprocessing

• Feature Alignment: Identified common genes across all datasets to create a consistent feature space
• Transformation: Applied log2(x+1) transformation to stabilize variance and reduce skewness
• Class Imbalance: Addressed imbalanced class distribution using weighted learning approaches

Model Development

Linear Approach

• SGDClassifier: Tested multiple regularization strategies:
  • No penalty (baseline)
  • L1 penalty (feature selection)
  • L2 penalty (coefficient shrinkage)
  • Elastic Net penalty (combined L1 and L2)
• Evaluation: 5-fold stratified cross-validation with class weighting

Non-Linear Approach

• XGBoost: Implemented to capture potential non-linear relationships
• Feature Selection: Tested model with all features vs. top 500 genes
• Interpretability: Used SHAP values to identify gene contributions

Results

Model Performance

Model	Accuracy	Precision (Macro)	Recall (Macro)	F1-Score (Macro)
SGDClassifier (Elastic Net)	93.30%	0.89	0.96	0.92
XGBoost (All Features)	74.75%	0.86	0.84	0.81
XGBoost (Top 500 Features)	75.42%	0.87	0.84	0.81

The Elastic Net model demonstrated superior performance across all metrics, with particularly strong class-specific results:

Class	Precision	Recall	F1-Score
Healthy	0.76	1.00	0.87
RA	0.92	1.00	0.96
SLE	1.00	0.87	0.93

Biomarker Discovery

Key Genes by Condition

Healthy Biomarkers:

• Upregulated: TRIM69, GTF2IP12, DES, SMPD5, EME2
• Downregulated: PINK1-AS, SNX32, GTF2IP1, HPR, CHRNB2

RA Biomarkers:

• Upregulated: TPI1P3, MIR378F, SLC23A3, CXCL9, AGMO
• Downregulated: ANP32D, TCP11X2, TCP11X1, CPEB1, ACADL

SLE Biomarkers:

• Upregulated: ANP32D, IFI27, SLC25A18, MIR6724-1, SLCO5A1
• Downregulated: AGMO, IGLV3-30, MIR378F, CLRN1, SLC26A3

Notable Findings:

• ANP32D: Strongly downregulated in RA, upregulated in SLE
• MIR378F: Upregulated in RA, downregulated in SLE
• IFI27: Significantly upregulated in SLE, potential key biomarker

Analysis & Insights

Model Performance Analysis

1. Linear vs. Non-Linear: Despite XGBoost's theoretical advantage in capturing complex relationships, the linear Elastic Net model performed substantially better, suggesting:

   • Linear relationships may sufficiently capture gene expression patterns for classification
   • Regularization is more critical than model complexity for high-dimensional, sparse RNA-Seq data

2. Regularization Benefits: Elastic Net's combined L1/L2 approach effectively:

   • Managed the high feature-to-sample ratio
   • Reduced overfitting through sparse feature selection
   • Handled multicollinearity among gene expression levels

Challenges & Limitations

• Dimensionality: High gene count relative to sample size presents statistical challenges
• Data Sparsity: Zero-inflated gene expression data requires specialized handling
• Disease Similarity: RA and SLE share autoimmune mechanisms, complicating differentiation
• Validation Need: Computational biomarkers require experimental validation

Conclusions & Future Directions

Key Takeaways

• Model Selection: For RNA-Seq classification, regularized linear models can outperform complex non-linear alternatives
• Potential Biomarkers: ANP32D, IFI27, and MIR378F warrant further investigation as disease biomarkers
• Disease Differentiation: Distinct gene expression patterns can effectively differentiate between RA and SLE

Future Work

• External Validation: Validate identified biomarkers in independent cohorts
• Biological Pathway Analysis: Explore biological significance of key genes
• Feature Engineering: Incorporate prior biological knowledge into model development
• Advanced Models: Test deep learning approaches with regulatory prior information
• Multi-omics Integration: Combine RNA-Seq with other data types (DNA methylation, proteomics) for improved classification