DeepKin: Predicting Relatedness From Low-Coverage Genomes and Palaeogenomes With Convolutional Neural Networks

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Estimating genetic relatedness is crucial in various fields.
• Traditional methods face challenges with limited or degraded genomic data, such as palaeogenomes and forensic samples.
• Convolutional neural networks (CNNs) offer potential for advanced genomic data analysis.

Purpose of the Study:

• To introduce DeepKin, a novel tool for predicting genetic relatedness using CNNs.
• To address the limitations of existing methods in handling sparse genomic data.
• To classify relatedness up to the third degree and identify close familial pairs.

Main Methods:

• Development of two CNN models trained exclusively on simulated genomic data.
• Input data format: PLINK's .map and .ped files.
• Benchmarking against the READv2 tool and validation on empirical palaeogenomic datasets.

Main Results:

• DeepKin achieves comparable or superior performance to READv2.
• Demonstrated robustness and adaptability across diverse genetic backgrounds.
• Achieved >90% accuracy with over 10,000 shared SNPs on palaeogenomic data.

Conclusions:

• DeepKin provides a novel methodological approach for relatedness estimation, especially with degraded samples.
• The tool is applicable to ancient DNA, forensic genetics, and conservation genetics.
• CNNs show promise for advancing relatedness prediction in challenging genomic scenarios.