A systematic assessment of machine learning for structural variant filtering.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate structural variant (SV) identification from long-read sequencing is challenging.
• Existing machine learning (ML) methods lack systematic performance comparison.
• Deep learning and foundation models are increasingly applied to SV analysis.

Purpose of the Study:

• To benchmark diverse ML paradigms for SV filtering.
• To evaluate classical, deep learning, and foundation models on standardized data.
• To provide a framework for selecting SV detection methods.

Main Methods:

• Comprehensive benchmark of five ML paradigms: Random Forest, computer vision models (ResNet/VICReg), diffusion-based anomaly detection, sparse autoencoders (SAEs) on Evo2-7B, and ensembles.
• Utilized standardized Genome in a Bottle (GIAB) data for HG002 and HG005 samples.
• Evaluated models based on accuracy, efficiency, and interpretability.

Main Results:

• A simple Random Forest classifier achieved a peak F1-score of 95.7%, comparable to complex models like ResNet50 (95.9%) and diffusion models (95.8%).
• Diffusion models and SAEs showed promise in representation learning and feature interpretability but did not exceed the performance ceiling.
• Ensemble methods provided no performance advantage in this study.

Conclusions:

• Simpler, interpretable ML models offer the best balance of accuracy, speed, and transparency for germline SV filtering.
• Increased model complexity should only be pursued if it addresses unmet biological needs.
• This benchmark facilitates pragmatic method selection for SV analysis.