[Paper Review] Analysis of Extremely Obese Individuals Using Deep Learning Stacked Autoencoders and Genome-Wide Genetic Data

Casimiro Adays Curbelo Montañez, Paul Fergus+2 more|arXiv (Cornell University)|Apr 16, 2018

Genetic Associations and Epidemiology4 citations

TL;DR

This study proposes a deep learning approach using stacked autoencoders (SAE) to analyze genome-wide genetic data in extremely obese individuals, overcoming limitations of traditional GWAS by capturing nonlinear SNP-SNP interactions. The method achieved an AUC of 0.975 and Gini score of 0.950 in classifying extreme obesity, significantly outperforming linear GWAS models.

ABSTRACT

The aetiology of polygenic obesity is multifactorial, which indicates that life-style and environmental factors may influence multiples genes to aggravate this disorder. Several low-risk single nucleotide polymorphisms (SNPs) have been associated with BMI. However, identified loci only explain a small proportion of the variation ob-served for this phenotype. The linear nature of genome wide association studies (GWAS) used to identify associations between genetic variants and the phenotype have had limited success in explaining the heritability variation of BMI and shown low predictive capacity in classification studies. GWAS ignores the epistatic interactions that less significant variants have on the phenotypic outcome. In this paper we utilise a novel deep learning-based methodology to reduce the high dimensional space in GWAS and find epistatic interactions between SNPs for classification purposes. SNPs were filtered based on the effects associations have with BMI. Since Bonferroni adjustment for multiple testing is highly conservative, an important proportion of SNPs involved in SNP-SNP interactions are ignored. Therefore, only SNPs with p-values < 1x10-2 were considered for subsequent epistasis analysis using stacked auto encoders (SAE). This allows the nonlinearity present in SNP-SNP interactions to be discovered through progressively smaller hidden layer units and to initialise a multi-layer feedforward artificial neural network (ANN) classifier. The classifier is fine-tuned to classify extremely obese and non-obese individuals. The best results were obtained with 2000 compressed units (SE=0.949153, SP=0.933014, Gini=0.949936, Lo-gloss=0.1956, AUC=0.97497 and MSE=0.054057). Using 50 compressed units it was possible to achieve (SE=0.785311, SP=0.799043, Gini=0.703566, Logloss=0.476864, AUC=0.85178 and MSE=0.156315).

Motivation & Objective

• To address the limited predictive power of traditional GWAS in explaining BMI heritability and classifying extreme obesity.
• To identify nonlinear epistatic interactions among SNPs that are missed by linear GWAS due to conservative multiple testing corrections.
• To reduce high-dimensional genetic data using stacked autoencoders for improved classification performance.
• To develop a deep learning pipeline that initializes and fine-tunes a multi-layer feedforward neural network for obesity classification.
• To evaluate the performance of the model using standard metrics such as AUC, Gini, and log-loss on a dataset filtered for SNPs with p < 1×10⁻².

Proposed method

• Filtered SNPs based on p-values < 1×10⁻² to reduce dimensionality while preserving potentially relevant variants.
• Applied stacked autoencoders (SAE) to learn hierarchical, nonlinear representations of SNP data through successive encoding and decoding layers.
• Used compressed representations from the SAE (2000 and 50 units) as input to a multi-layer feedforward artificial neural network (ANN) classifier.
• Fine-tuned the ANN classifier using backpropagation to optimize for classification of extremely obese vs. non-obese individuals.
• Employed a two-stage process: pre-training via SAE to initialize weights, followed by end-to-end fine-tuning for improved performance.
• Evaluated model performance using standard metrics: AUC, Gini, sensitivity (SE), specificity (SP), log-loss, and mean squared error (MSE).

Experimental results

Research questions

• RQ1 Can deep learning-based dimensionality reduction via stacked autoencoders improve the classification of extreme obesity beyond traditional GWAS?
• RQ2 To what extent do nonlinear SNP-SNP interactions contribute to the heritability of BMI in extremely obese individuals?
• RQ3 How does filtering SNPs using a less conservative p-value threshold (p < 1×10⁻²) affect the detection of epistatic interactions?
• RQ4 What is the optimal compressed representation size (number of units) in the SAE for maximizing classification accuracy in extreme obesity prediction?
• RQ5 How do performance metrics such as AUC, Gini, and log-loss compare between SAE-based models and standard GWAS approaches?

Key findings

• The model with 2000 compressed units achieved an AUC of 0.97497, indicating excellent discrimination between extremely obese and non-obese individuals.
• The same model reached a Gini score of 0.949936, reflecting strong model performance in ranking predictions.
• Sensitivity (SE) was 0.949153 and specificity (SP) was 0.933014, showing high true positive and true negative rates.
• Log-loss was minimized to 0.1956, indicating high confidence and accuracy in predicted probabilities.
• With only 50 compressed units, the model still achieved an AUC of 0.85178 and Gini of 0.703566, demonstrating robustness at lower dimensionality.
• The mean squared error (MSE) was 0.054057 for the 2000-unit model, indicating low prediction error.