RubricOE: a learning framework for genetic epidemiology

Subrata Saha, Aldo Guzmán-Sáenz, Aritra Bose, Filippo Utro, Daniel E. Platt, Laxmi ParidamedRxiv 2021 · DOI

Abstract

Genetic epidemiology is a growing area of interest in the past years due to the availability of genetic data with the decreasing cost of sequencing. Machine learning (ML) algorithms can be a very useful tool to study the genetic factors on disease incidence or on different traits characterizing a population. There are many challenges that plagues the field of genetic epidemiology including the unbalanced case-control data sets, fallibility of standard genome wide association studies with single marker analysis, heavily underdetermined systems with millions of markers in contrast of a few thousands of samples, to name a few. Ensemble ML methods can be a very useful tool to tackle many of these challenges and thus we propose RubricOE, a pipeline of ML algorithms with error bar computations to obtain interpretable genetic and non-genetic features from genomic or transcriptomic data combined with clinical factors in the form of electronic health records. RubricOE is shown to be robust in simulation studies, detecting true associations with traits of interest in arbitrarily structured multi-ethnic populations.

Introduction

In the world of genomics, researchers face a daunting mathematical hurdle known as the "curse of dimensionality." A typical study might involve millions of genetic markers (SNPs) but only a few thousand participants—the classic "small $n$, large $p$" problem. Standard Genome-Wide Association Studies (GWAS) usually look at one marker at a time, often missing complex interactions between genes (epistasis) and remaining highly sensitive to "population structure." These genetic differences between ethnic groups can lead to false-positive associations. Furthermore, many datasets are "unbalanced," with far more healthy controls than diseased cases. RubricOE (a Rubric for Omics and Epidemiology) is introduced to provide a more robust, multi-marker framework that can handle these imbalances and structured populations while remaining computationally efficient.

Methodology

RubricOE is an ensemble pipeline consisting of three main stages:

  1. Quality Control (QC): Standard cleaning of input genomic and clinical data.
  2. Iterative Feature Selection (IFS): This stage uses Ridge Regression to rank SNPs based on their contribution to a trait. Unlike standard regression, Ridge handles correlated variables (Linkage Disequilibrium) well. The pipeline then uses Linear Support Vector Machines (LSVM) to score these features using the Youden’s J statistic, a measure of informedness that balances sensitivity and specificity.
  3. Stable Set Construction (SSC): To ensure results are not artifacts of a specific data split, the process is repeated multiple times on different subsets. Only the SNPs that consistently appear across these iterations are included in the final "Stable Set."

A key mathematical contribution is the adaptation of error bar computations and variance partitioning specifically for the Ridge Regression framework, allowing the model to provide interpretable scores for each genetic feature.

Experimental Setup and Datasets

The authors evaluated RubricOE using six simulated datasets, each containing 1,000 individuals and 10,000 SNPs. They used three distinct population models to test the framework's versatility:

  • Balding-Nichols (BN): Three distinct, non-mixing populations.
  • Pritchard-Stephens-Donnelly (PSD): Admixed populations where individuals share alleles across groups.
  • 1000 Genomes Project (TGP): A model based on real-world human genetic variation, representing the most complex and realistic scenario.
PCA plots of simulated population structures: (A) Balding-Nichols, (B) Pritchard-Stephens-Donnelly, and (C) 1000 Genomes Project.

Key Results

RubricOE demonstrated high accuracy and robustness across varying genetic effect sizes. In the PSD model with a 70% genetic effect, RubricOE correctly identified 90% of the true causal SNPs. Even in the most challenging TGP scenario with a low (10%) genetic effect, the framework still detected 30% of causal SNPs, whereas standard methods often fail to find any signal in such noisy conditions.

Performance metrics of RubricOE across different population models and genetic effect scenarios.

The "Stable Set" approach resulted in very few false positives, even as the genetic effect size increased. Furthermore, the authors demonstrated that peak classification accuracy is often reached using only the first few hundred ranked features, meaning the pipeline can be "short-circuited" to save time when dealing with millions of markers.

Discussion and Future Work

RubricOE's reliance on the persistence of features across multiple subsampled models acts as a powerful filter against the noise inherent in high-dimensional genomic data. While the simulation results are promising, the next step for this framework is validation against massive, real-world biobanks like the UK Biobank. Future work could involve integrating Deep Learning into the pipeline to capture even more complex, non-linear gene-environment interactions and further exploring the integration of transcriptomic data and Electronic Health Records (EHR) for a truly multi-omic view of disease.

Summary updated

Abstract

Abstract Genetic epidemiology is a growing area of interest in the past years due to the availability of genetic data with the decreasing cost of sequencing. Machine learning (ML) algorithms can be a very useful tool to study the genetic factors on disease incidence or on different traits characterizing a population. There are many challenges that plagues the field of genetic epidemiology including the unbalanced case-control data sets, fallibility of standard genome wide association studies with single marker analysis, heavily underdetermined systems with millions of markers in contrast of a few thousands of samples, to name a few. Ensemble ML methods can be a very useful tool to tackle many of these challenges and thus we propose RubricOE, a pipeline of ML algorithms with error bar computations to obtain interpretable genetic and non-genetic features from genomic or transcriptomic data combined with clinical factors in the form of electronic health records. RubricOE is shown to be robust in simulation studies, detecting true associations with traits of interest in arbitrarily structured multi-ethnic populations.
BibTeX
@misc{saha2021rubricoe,
  author = {Subrata Saha and Aldo Guzmán-Sáenz and Aritra Bose and Filippo Utro and Daniel E. Platt and Laxmi Parida},
  title = {{RubricOE: a learning framework for genetic epidemiology}},
  year = {2021},
  doi = {10.1101/2021.03.09.21253105},
  howpublished = {medRxiv},
}
Figure