[Paper Review] Utilising Deep Learning and Genome Wide Association Studies for Epistatic-Driven Preterm Birth Classification in African-American Women

Paul Fergus, Casimiro Curbelo Montañez+3 more|arXiv (Cornell University)|Jan 6, 2018

Preterm Birth and Chorioamnionitis27 references3 citations

TL;DR

This study proposes a stacked autoencoder-based deep learning framework to detect higher-order epistatic interactions among SNPs for preterm birth classification in African-American women using GWAS data. By compressing 4,666 SNPs into latent representations, the model achieves high AUC (0.9825) and F1-like performance, demonstrating deep learning's potential to uncover complex genetic interactions missed by traditional GWAS.

ABSTRACT

Genome Wide Association Studies (GWAS) are used to identify statistically significant genetic variants in case-control studies. GWAS typically use a p-value threshold of 5 x 10-8 to identify highly ranked single nucleotide polymorphisms (SNPs). However, evidence has shown that many of these are, in fact, false positives. Using lower p-values it is possible to to investigate the joint epistatic interactions between SNPs and provide better insights into phenotype expression. However, computational complexity is increased exponentially as a function of higher-order combinations. In this paper, we propose a novel framework, based on nonlinear transformations of combinatorically large SNP data, using stacked autoencoders, to identify higher-order SNP interactions. We focus on the challenging problem of classifying preterm births. Evidence suggests that this complex condition has a strong genetic component with unexplained heritability reportedly between 20%-40%. This claim is substantiated using a GWAS data set, obtained from dbGap, which contains predominantly urban low-income African-American women who had normal deliveries (between 37 and 42 weeks of gestation) and preterm deliveries (less than 37 weeks of gestation). Latent representations from original SNP sequences are used to initialize a deep learning classifier before it is fine-tuned for classification tasks (term and preterm births). The complete network models the epistatic effects of major and minor SNP perturbations. All models are evaluated using standard binary classifier performance metrics. The findings show that important information pertaining to SNPs and epistasis can be extracted from 4666 raw SNPs generated using logistic regression (p-value=5 x 10-3) and used to fit a deep learning model and obtain results (Sen=0.9289, Spec=0.9591, Gini=0.9651, Logloss=0.3080, AUC=0.9825, MSE=0.0942) using 500 hidden nodes.

Motivation & Objective

• To address the unexplained heritability of preterm birth (PTB), which is estimated at 20–40%, by identifying higher-order epistatic interactions among SNPs.
• To overcome the limitations of traditional GWAS, which rely on single-SNP p-values (e.g., 5×10⁻⁸) and often miss true associations due to low power and high false positive rates.
• To develop a scalable, nonlinear method for modeling complex, combinatorial SNP interactions that are computationally infeasible with traditional approaches like multifactor dimensionality reduction (MDR).
• To create a robust, interpretable classification system using deep learning that can serve as an early screening tool for PTB risk in high-risk populations.

Proposed method

• A stacked autoencoder is used to learn nonlinear, compressed representations from 4,666 SNPs selected via logistic regression (p < 5×10⁻³) from a GWAS dataset.
• The learned latent features from the autoencoder are used to initialize a deep neural network classifier for binary classification of term vs. preterm birth.
• The model is fine-tuned using backpropagation to minimize mean squared error (MSE) and cross-entropy loss (logloss), with performance evaluated using standard metrics.
• Two architectures are tested: one with 50 hidden nodes and another with 500 hidden nodes to assess scalability and generalization.
• The framework models epistatic effects by capturing nonlinear interactions between SNPs through hierarchical feature learning in deep layers.
• Model training uses early stopping and validation monitoring to prevent overfitting, with no significant overfitting observed between training and validation sets.

Experimental results

Research questions

• RQ1 Can deep learning models effectively detect higher-order epistatic interactions among SNPs that are missed by conventional GWAS?
• RQ2 Does using a stacked autoencoder to compress SNP data improve classification accuracy for preterm birth prediction in African-American women?
• RQ3 Can a deep learning framework achieve high performance in PTB classification while maintaining generalization and avoiding overfitting?
• RQ4 To what extent do latent representations from autoencoders capture biologically meaningful genetic interactions relevant to PTB?

Key findings

• The model achieved an AUC of 0.9745 with 50 hidden nodes (Sensitivity: 0.9562, Specificity: 0.8780, Gini: 0.9490, Logloss: 0.5901, MSE: 0.2010).
• With 500 hidden nodes, the model achieved a higher AUC of 0.9825, sensitivity of 0.9289, specificity of 0.9591, Gini score of 0.9651, logloss of 0.3080, and MSE of 0.0942.
• No significant overfitting was observed, as training and validation performance remained consistent across both architectures.
• The results demonstrate that deep learning can extract meaningful epistatic patterns from high-dimensional SNP data, even when individual SNPs show weak marginal effects.
• The framework shows promise as a scalable alternative to traditional epistasis detection methods, which suffer from exponential computational complexity.
• The study provides a foundation for future work integrating interpretable models (e.g., random forests) to trace SNP contributions in the learned representations.