MFAID-Net: A Multi-modal Feature Fusion Deep Learning Network for Robust Adaptive Introgression Detection Across Diverse Evolutionary Scenarios

Abstract

Adaptive introgression (AI), the beneficial genetic transfer between species, is key to adaptation, yet its genomic identification is challenging. Existing methods are inconsistent and often require unavailable donor population data. We propose MFAID-Net, a novel multi-modal network for adaptive introgression detection. MFAID-Net employs a dual-encoder architecture, combining a Convolutional Neural Network for local genomic features with a Multi-Layer Perceptron for population genetic statistics and structure. A Transformer-like attention mechanism adaptively fuses modalities, dynamically weighting their importance. Multi-task learning differentiates AI from confounding signals like neutral introgression and selective sweeps. Comprehensive simulations demonstrate MFAID-Nets superior performance across challenging scenarios (e.g., strong/weak selection, ancient migration), often outperforming current methods. It is also robust to donor data availability. Ablation and attention analyses confirm component contributions and provide interpretability. MFAID-Net offers a more robust, accurate, and adaptable tool for discovering AI events.