Scalable CNN-based classification of selective sweeps using derived allele frequencies.

Area of Science:

• Genetics
• Computational Biology
• Machine Learning

Background:

• Selective sweeps are crucial for understanding adaptation but are challenging to detect due to confounding factors like population bottlenecks.
• Existing methods using single nucleotide polymorphisms (SNPs) are sensitive to these factors and computationally intensive.
• Machine learning, particularly convolutional neural networks (CNNs), offers potential for robust detection but faces scalability issues with large genomic datasets.

Purpose of the Study:

• To develop a computationally efficient and scalable deep learning model for accurate selective sweep detection.
• To create a model robust to confounding evolutionary factors without requiring data preprocessing.
• To enable practical application of CNNs for large-scale genomic data analysis.

Main Methods:

• Proposed a 1D CNN-based model named FAST-NN.
• Utilized derived allele frequencies, avoiding summary statistics or raw SNP data.
• Performed neural architecture search based on the SweepNet architecture and evaluated data fusion approaches.

Main Results:

• FAST-NN achieved up to 12% higher inference accuracy than the reference architecture across challenging scenarios.
• The model demonstrates sample-size invariance and scalability.
• FAST-NN is significantly faster than existing methods: 30x-259x on CPU and 2.0x-6.2x on GPU for 128-1000 samples.

Conclusions:

• FAST-NN provides a highly accurate, computationally efficient, and scalable solution for selective sweep detection.
• The model's robustness to confounding factors and lack of preprocessing requirement make it practical for large-scale genomic studies.
• This work facilitates the widespread adoption of deep learning in population genetics research.