DeepSweep

DeepSweep applies deep learning to detect loci under recent positive selection in whole-genome sequence (WGS) data of malaria parasites, with the aim of identifying genetic signatures relevant to adaptation and drug resistance.

Key Features:

• Supervised deep learning: Trains a supervised model on haplotypic images from genomic regions known to have undergone selective sweeps.
• High predictive accuracy: Achieves areas under the ROC curve exceeding 0.95 when tested on simulated genomic data.
• Application to Plasmodium datasets: Applied to WGS data from Plasmodium falciparum (n=1125) and Plasmodium vivax (n=368) for genome-wide sweep detection.
• Comparison with haplotype methods: Shows 60–75% overlap of hits at P<0.0001 with within-population iHS and across-population Rsb extended haplotype homozygosity metrics.
• Detection of drug-resistance loci: Identifies regions proximal to known drug resistance loci including pfcrt, pfdhps, pfmdr1, and pvmrp1.
• Generalizability: Methodology can be trained on other organisms and selection pressures beyond malaria parasites.

Scientific Applications:

• Genome-based surveillance: Surveillance of parasite populations to detect recent positive selection and emerging adaptive variants.
• Drug-resistance monitoring: Identification and localization of signals near known drug-resistance genes to inform studies of resistance evolution.
• Method comparison and validation: Complementary validation of selection signals alongside extended haplotype homozygosity metrics iHS and Rsb.
• Large-scale WGS analysis: Screening of large WGS datasets for loci under recent selective sweeps.

Methodology:

Train a supervised deep learning model on haplotypic images from known selective sweep regions, evaluate performance on simulated genomic data using ROC AUC, apply the model to WGS datasets of P. falciparum (n=1125) and P. vivax (n=368), and compare detected hits with within-population iHS and across-population Rsb while identifying regions proximal to pfcrt, pfdhps, pfmdr1, and pvmrp1.