Deep learning methods may not outperform other machine learning methods on analyzing genomic studies | JoVE Visualize

Area of Science:

Genomics
Biomedical Data Science
Machine Learning

Background:

Deep learning (DL) shows success in biomedical image processing but faces challenges with genomic data due to small sample sizes and lack of common structural patterns.
Genomic datasets often lack the typical structures (e.g., convolutional layers) that benefit pre-trained DL models.
Over-reliance on DL for genomic analysis necessitates a comparative evaluation against established non-deep machine learning (ML) methods.

Purpose of the Study:

To benchmark the performance of deep learning (DL) versus non-deep machine learning (ML) methods for analyzing genomic data across various sample sizes.
To assess the predictive accuracy of different models for three lung diseases (asthma, COPD, lung cancer) using UK Biobank data.
To provide guidance on the appropriate use of DL in genomic studies, considering sample size limitations.

Main Methods:

A benchmark study was conducted using UK Biobank data, including comprehensive genomic information (millions of Single-Nucleotide Polymorphisms) and patient characteristics.
Five prediction models were evaluated: three non-deep ML (Elastic Net, XGBoost, SVM) and two DL (DNN, LSTM).
Performance was assessed using standard metrics like F1-score and the hit curve for rare event prediction.

Main Results:

Deep learning (DL) methods frequently underperformed non-deep machine learning (ML) models in analyzing genomic data, even with datasets exceeding 200,000 samples.
Performance differences between DL and non-deep ML diminished as sample size increased.
The study suggests that DL methods may offer benefits only with genomic datasets significantly larger than those analyzed.

Conclusions:

Non-deep machine learning (ML) methods are often more suitable and perform comparably or better than deep learning (DL) for current genomic data analysis.
The findings caution against the overuse of DL in genomic studies, even with biobank-scale data.
Increasing sample sizes may enhance DL performance in genomics, but current evidence favors established ML approaches for typical genomic datasets.