One stretch of DNA, four evolutionary forces: multiscale haplotype images lift classification accuracy to 90.85%
Synopsis
Telling neutrality, selective sweeps, balancing selection and adaptive introgression apart is hard because the haplotype footprints these processes leave overlap. SHADES renders haplotype frequencies, differences between frequency distributions, and sequence differences between ranked haplotypes as three image types, trains convolutional networks at window sizes of 11, 21 and 51 SNPs, and stacks the three scales for four-class classification. On simulated test replicates, the deeper SHADES[M2] reached 90.85% overall accuracy, against 67.83% for SURFDAWave, 68.30% for diploS/HIC, 65.20% for TrIdent, 63.20% for FAST-NN and 79.43% for FASTER-NN on the same test set.
Interpretation
A single genome scan can now return a four-way call on evolutionary process instead of requiring a separate detector for each kind of selection signal. Most existing detectors were built for one phenomenon in isolation, so they could not place several evolutionary processes in one discriminative framework, even though neutral, sweep, balancing-selection and introgression footprints overlap heavily. On 1,000 simulated test replicates per class, SHADES[M2] reached 90.85% overall accuracy, while SURFDAWave reached 67.83%, diploS/HIC 68.30%, TrIdent 65.20%, FAST-NN 63.20% and FASTER-NN 79.43% on the same test set.
The model turns a genomic stretch into three images recording haplotype frequencies, differences in frequency distributions between windows, and sequence differences between frequency-ranked neighbouring haplotypes. These statistics are imaged at window widths of 11, 21 and 51 SNPs; a shallow M1 and a deeper Inception-ResNet-v2-based M2 are each trained at all three scales, and a stacking layer concatenates the three penultimate feature vectors before a fully connected classifier. Training and test data come from SLiM forward simulations under European CEU demographic history, with 2 Mb sequences, 198 sampled haplotypes and 6,000 replicates per class, of which 5,000 train and 1,000 test.
On real population data the model's high-confidence candidates coincide with known targets of selection. Those loci had previously been reported by separate statistical tests or targeted scans, and now emerge together from one four-class scan. Scanning the 22 autosomes of 1000 Genomes European samples in 3,115-SNP windows with a 21-SNP stride gave 183,558 windows; the LCT region carried sweep support (LCT confidence 0.808, MCM6 0.878), the MHC region a dense balancing-selection cluster, and BNC2 adaptive introgression support of 0.63.
Perspective
The demonstrated setting is 2 Mb sequences simulated with SLiM under European CEU demographic history with 198 sampled haplotypes, plus an autosome-wide scan of 1000 Genomes European samples; the discussion names high recombination rate as the main difficulty for the method.
The authors report SHADES[M2] accuracy falling to 66.33% under the high recombination rate setting and note that elevated recombination rapidly erodes haplotype footprints of past selection; adding gene flow from an unsampled population lowered SHADES[M2] accuracy by 7.72% relative to the CEU model. The specific pathogens and selected phenotypes behind the novel candidates remain unresolved.