Abstract
Learning a small number of genetic variants associated with multiple complex genetic traits is of practical importance and remains challenging due to the high dimensional nature of data. In this paper, we proposed a two-graph guided multi-task Lasso to address this issue with an emphasis on estimating subnetwork-to-subnetwork associations in expression quantitative trait loci (eQTL) mapping. The proposed model can learn such subnetwork-to-subnetwork associations and therefore can be seen as a generalization of several state-of-the-art multi-task feature selection methods. Additionally, this model has a nice property of allowing flexible structured sparsity on both feature and label domains. Simulation study shows the improved performance of our model and a human eQTL data set is analyzed to further demonstrate the applications of the model.
Introduction
In genomics, expression quantitative trait loci (eQTL) mapping aims to identify the links between genetic variations, such as Single Nucleotide Polymorphisms (SNPs) or Copy Number Variants (CNVs), and gene expression levels. This task is inherently difficult due to the high dimensionality of the data, where the number of genetic markers ($p$) and gene traits ($K$) far exceeds the number of samples ($n$). Traditional methods often rely on simple pairwise correlations, assuming independence between genetic variants and gene expressions. However, biological processes are modular; genes typically function within pathways or subnetworks, and multiple genetic variants often jointly influence these groups. Existing Multi-task Lasso methods have attempted to group traits but frequently ignore the structural relationships within the genetic features themselves. This paper introduces a framework to identify complex, many-to-many relationships while maintaining structured sparsity.
Methodology: The MTLasso 2G Framework
The core contribution of this work is the Two-Graph Guided Multi-task Lasso (MTLasso 2G). This model extends the standard Lasso framework by incorporating prior structural knowledge represented as graphs for both genetic variants (features) and gene expressions (labels).
The model utilizes two specific penalty terms:
- Label Graph ($G_1$): Encodes known or estimated correlations between gene expressions.
- Feature Graph ($G_2$): Encodes correlations or physical proximity between genetic variants.
By combining a least-squares loss with an $L_1$ penalty for overall sparsity and two fused-lasso style penalties, the objective function encourages the model to assign similar association weights to nodes connected within the graphs. This approach effectively identifies "subnetwork-to-subnetwork" associations, reflecting the modular nature of biological pathways. To solve the resulting non-differentiable optimization problem, the authors employ a coordinate-descent algorithm within a Majorization-Minimization (MM) framework, transforming the complex objective into a series of smooth sub-problems.
Experimental Results
The authors validated MTLasso 2G using both synthetic simulations and real-world human eQTL data.
Simulation Study
Synthetic datasets were generated with varying numbers of traits ($K=10, 50$) and features ($J$ up to 500) to evaluate the model's ability to recover known association patterns under noise.
As shown in the performance plots, MTLasso 2G consistently achieved lower $L_1$ estimation errors and higher Area Under the Precision-Recall Curve (AUC) compared to standard Lasso and single-graph guided methods. The inclusion of both graphs allows the model to leverage domain knowledge from both the feature and label sides, leading to more robust feature selection.
Human eQTL Analysis
The model was applied to a human eQTL dataset from the 1000 Genomes Project, focusing on Chromosome 20 data from the Yoruba population. The analysis involved 139 CNVs and 379 gene expression profiles. MTLasso 2G identified 532 associations, filtering out many potential false positives found by single-graph methods. Notably, it uncovered a joint association between two CNVs and a co-expression subnetwork containing RBL1 and VAPB, genes known for their roles in cancer and motor neuron diseases, respectively.
Conclusion
The MTLasso 2G approach provides a flexible and powerful framework for eQTL mapping by incorporating structured sparsity on both feature and label domains. While the model's performance depends on the quality of the input graphs, it demonstrates significant improvements in accuracy and biological relevance over traditional methods. Future work may involve scaling the algorithm for whole-genome analysis and incorporating diverse prior knowledge, such as protein-protein interaction networks, into the graph structures.
Summary updated