GMIP-PLSR: A Nextflow Pipeline for GWAS and Multi-Omics Integration in Gene Prioritization Using PLSR

The authors present GMIP-PLSR, a Nextflow-based pipeline that integrates GWAS with multi-omics data using Partial Least Squares Regression to overcome multicollinearity issues in existing tools like PoPS, thereby significantly improving gene prioritization and biological insights for complex traits such as NAFLD.

The Gist

Imagine you are a detective trying to solve a massive mystery: Why do some people get sick with complex diseases like diabetes, heart disease, or fatty liver disease, while others don't?

For years, scientists have used a tool called GWAS (Genome-Wide Association Studies) to scan the entire human genetic code. Think of GWAS as a giant spotlight that scans a dark room and finds thousands of "glowing spots" (genetic variants) that are suspicious.

The Problem:
The problem is that the spotlight is too broad. It finds a whole neighborhood of glowing spots, but it can't tell you exactly which house (gene) is the culprit. It's like finding a street where a crime happened, but not knowing which specific door to knock on. Furthermore, many of these "suspicious" spots are just in the wrong part of the genome (non-coding regions), making them hard to understand.

The Old Solution (PoPS):
Scientists tried to solve this by bringing in other clues, like how genes talk to each other (networks) or how they behave in different tissues (RNA data). One popular tool called PoPS (Polygenic Priority Score) tried to rank these genes by asking: "Do these genes look like the ones we already know are bad?"

However, PoPS had a major flaw. It suffered from "Multicollinearity."

• The Analogy: Imagine you are trying to guess the price of a house. You ask a real estate agent for clues.

  • Clue 1: "It has a big kitchen."
  • Clue 2: "It has a large dining room."
  • Clue 3: "It has a spacious living area."
  • Clue 4: "It has an open floor plan."
  • Clue 5: "It has a huge kitchen."

  The agent is giving you the same information five different ways. If you try to add up all these clues, you get confused and overestimate the price. The clues are "too correlated." PoPS was getting confused by these overlapping clues, leading to shaky results.

The New Solution: GMIP-PLSR
The authors of this paper built a new, smarter detective framework called GMIP (GWAS & Multi-omics Integration Pipeline). But their real breakthrough is a specific upgrade called GMIP-PLSR.

Here is how they fixed the "confused clues" problem using a technique called PLSR (Partial Least Squares Regression):

1. The "Grouping" Strategy: Instead of treating every clue separately, PLSR looks at all the clues and groups the redundant ones together.
   • Analogy: Instead of counting the kitchen, dining room, and living room separately, PLSR says, "Okay, let's call this whole thing 'The Open Living Space'." It creates a single, powerful "super-clue" that captures the essence of all those overlapping features without the confusion.
2. The "Smart Filter": It filters out the noise and focuses only on the patterns that actually matter for the disease.
3. The Result: By cleaning up the data this way, GMIP-PLSR can point to the exact gene responsible for the disease much more accurately than the old methods.

The "Superpower" Case Study: NAFLD
To prove it works, they tested it on NAFLD (Non-Alcoholic Fatty Liver Disease).

• They used two types of clues:
  1. General Clues: Data from public databases (like a general encyclopedia).
  2. Specialized Clues: Data from a specific study of liver cells (like a specialized medical journal).
• The Outcome: The new system (GMIP-PLSR) combined these clues perfectly. It didn't just find any liver gene; it found the specific genes that drive fatty liver disease, identifying pathways that the old methods missed. It was like upgrading from a standard map to a GPS that knows exactly where the potholes are.

Why This Matters

• Better Drug Discovery: If we know the exact gene causing the disease, we can design drugs to target it specifically, rather than guessing.
• Personalized Medicine: It helps doctors understand why a specific patient might get sick, leading to better, tailored treatments.
• Efficiency: The tool is built on "Nextflow," which is like a robotic assembly line. It can run these complex analyses on a laptop or a supercomputer without breaking a sweat.

In a Nutshell:
The authors built a smart, modular pipeline that takes the messy, confusing data from genetic studies and cleans it up using a mathematical "grouping" trick (PLSR). This allows scientists to finally stop guessing which genes are causing complex diseases and start knowing for sure, paving the way for better cures.

Technical Summary

1. Problem Statement

Genome-wide association studies (GWAS) have successfully identified thousands of genetic loci associated with complex traits and diseases. However, a major bottleneck remains in gene prioritization:

• Causal Gene Identification: Most GWAS loci contain numerous variants in linkage disequilibrium (LD), making it difficult to pinpoint the specific causal gene.
• Non-coding Variants: A significant proportion of associated variants lie in non-coding regions, obscuring their functional targets.
• Limitations of Existing Tools: Current methods like PoPS (Polygenic Priority Score) integrate multi-omics data (e.g., gene expression, protein-protein interactions) but suffer from multicollinearity. When features (e.g., expression levels across tissues) are highly correlated, standard regression models (like Ridge regression used in PoPS) produce inflated standard errors and reduced interpretability, leading to suboptimal gene ranking.
• Lack of Standardization: There is no unified framework to systematically compare, combine, or optimize different gene prioritization strategies (e.g., NetWAS, NAGA, PoPS) across diverse datasets.

2. Methodology

The authors developed GMIP (GWAS & Multi-omics Integration Pipeline), a modular, scalable Nextflow pipeline, and introduced GMIP-PLSR, an extension that specifically addresses multicollinearity using Partial Least Squares Regression (PLSR).

A. Pipeline Architecture (GMIP)

The framework consists of four modular components:

1. SNP-to-Gene Mapping: Uses MAGMA to convert SNP-level GWAS summary statistics into gene-level Z-scores, accounting for LD using 1000 Genomes reference data.
2. Machine Learning Modeling: Integrates diverse gene features from multi-omics sources:
   • NetWAS: Tissue-specific networks derived from 987 genomic datasets.
   • NAGA (PCNet): A parsimonious composite network combining PPI, co-expression, and pathways.
   • PoPS Features: Bulk/scRNA-seq expression, curated pathways (KEGG, GO), and predicted PPI networks.
   • Custom Features: Disease-specific scRNA-seq data (e.g., NAFLD mouse data) processed via Seurat, PCA/ICA, and clustering.
3. Cross-Validation Strategy:
   • LOCO-CV (Leave-One-Chromosome-Out): The primary strategy used to prevent information leakage caused by chromosomal proximity between training and testing sets.
   • Stratified k-Fold: Included for comparison.
4. Evaluation:
   • Benchmarker: Uses Stratified LD Score Regression (S-LDSC) to calculate Normalized Tau ($\tau$) scores, measuring the enrichment of heritability in the top prioritized genes.
   • GSEA (Gene Set Enrichment Analysis): Evaluates if original GWAS significant genes are enriched at the top of the reprioritized list.

B. The GMIP-PLSR Innovation

To solve the multicollinearity issue in PoPS:

• Diagnosis: The authors calculated the Condition Index (CI) for feature sets, finding that >30% of features had CI > 30, indicating severe multicollinearity.
• Solution: Replaced the standard Ridge Regression in PoPS with Partial Least Squares Regression (PLSR).
  • PLSR constructs Latent Variables (LVs) that maximize the covariance between the predictor matrix (gene features) and the response matrix (GWAS Z-scores).
  • This simultaneously performs dimensionality reduction and regression, handling correlated predictors more effectively than Ridge regression while offering better biological interpretability of the latent components.
  • The optimal number of components was determined to be 3 (nc=3) across most datasets.

3. Key Contributions

1. Unified Framework (GMIP): The first Nextflow-based pipeline to modularize and standardize the comparison of multiple gene prioritization methods (NetWAS, NAGA, PoPS) and feature sets within a single workflow.
2. Multicollinearity Mitigation: The introduction of GMIP-PLSR, which demonstrates that PLSR significantly outperforms Ridge regression in the context of highly correlated multi-omics features.
3. Scalability and Reproducibility: Built on Nextflow, the pipeline is computationally efficient and adaptable to various environments (laptops to HPC clusters).
4. Disease-Specific Integration: Demonstrated the ability to integrate custom, disease-specific single-cell RNA-seq data (NAFLD) alongside general public features.

4. Key Results

• Performance on 8 Initial GWAS Traits:
  • NAGA performed well without cross-validation but suffered from overfitting under LOCO-CV.
  • PoPS maintained robustness under LOCO-CV but was limited by multicollinearity.
  • GMIP-PLSR (nc=3) consistently outperformed both standard PoPS and PCA+Ridge approaches. For example, in the RAD (Rheumatoid Arthritis) trait, the Normalized Tau score improved from 2.99 (PoPS) to 5.02 (GMIP-PLSR). Similar gains were observed for BMI and LDL.
• Large-Scale Evaluation (46 Traits):
  • Applied to 46 diverse GWAS traits with varying heritability.
  • 43 out of 46 traits were successfully reprioritized with significant enrichment ($p < 0.01$).
  • Heritability Threshold: Traits with observed heritability ($h^2$) > 0.05 generally yielded successful reprioritization.
  • Hyperparameter Optimization: Using the top 500 genes and 3 PLSR components yielded the best results.
  • Comparison: GMIP-PLSR outperformed PoPS in the vast majority of cases (points in the scatter plot lay above the line of equality).
• NAFLD Case Study:
  • Compared general PoPS features vs. NAFLD-specific scRNA-seq features.
  • PoPS features yielded higher heritability enrichment ($\tau = 2.96$) and broader pathway coverage (24 pathways).
  • scRNA-seq features provided focused insights into liver-specific pathology (4 pathways) but lower overall enrichment ($\tau = 1.59$).
  • The study confirmed that general features often capture broader biological signals, while disease-specific features offer targeted mechanistic insights.

5. Significance and Future Directions

• Biological Impact: GMIP-PLSR provides a more reliable method for identifying causal genes, which is critical for drug target discovery and understanding disease mechanisms.
• Methodological Advancement: By addressing multicollinearity via PLSR, the study sets a new standard for integrating high-dimensional, correlated multi-omics data in post-GWAS analysis.
• Future Perspectives:
  • Integration of locus-based fine-mapping (e.g., FINEMAP, PAINTOR) to refine causal variant identification.
  • Utilization of NAGA network features within the PLSR framework to capture indirect network effects.
  • Incorporation of RNA-seq foundational models (e.g., large-scale single-cell models) to extract latent features for gene ranking.
  • Application to drug discovery pipelines to identify druggable targets.

In conclusion, GMIP-PLSR represents a significant step forward in computational genomics, offering a robust, scalable, and statistically superior framework for translating GWAS findings into actionable biological insights.