Harnessing deep learning for SNP-based disease prediction in genomics.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Single nucleotide polymorphism (SNP) data is crucial for understanding genetic variations and disease associations.
• Predictive modeling using genomic data holds significant potential for personalized medicine and disease risk assessment.
• Evaluating diverse deep learning (DL) architectures requires standardized methodologies for reliable comparisons.

Purpose of the Study:

• To investigate the efficacy of various deep learning models in predicting disease status from SNP data.
• To establish a standardized pipeline for preprocessing, feature selection, and model training in genomic data analysis.
• To compare the performance of different DL architectures, including feedforward networks, autoencoders, CNNs, and RNNs, on multiple datasets.

Main Methods:

• Utilized eight Gene Expression Omnibus (GEO) datasets for analysis.
• Implemented a consistent data processing pipeline: genotype encoding, data cleaning, and feature selection.
• Trained and evaluated multiple deep learning architectures: feedforward networks, autoencoders, Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs).

Main Results:

• Feedforward networks and autoencoders consistently outperformed other models across most datasets.
• The standardized pipeline facilitated a direct and fair comparison of model performances.
• All evaluated DL models showed varying degrees of success in predicting disease status from SNP data.

Conclusions:

• Deep learning models, particularly feedforward networks and autoencoders, are effective for disease prediction using SNP data.
• A standardized approach to data preprocessing and model training is essential for reliable genomic deep learning applications.
• This study provides a practical framework for applying deep learning in genomics, with potential for future advancements.