Geometric deep learning for diploid de novo genome assembly
Professor Mile Sikic, Genome Institute of Singapore
Abstract:
De novo genome assembly is one of computational biology’s most fundamental challenges, particularly for diploid genomes, where maternal and paternal haplotypes must be correctly separated and reconstructed. Building upon recent advances in geometric deep learning and the growing availability of high-quality genome references essential for training machine learning models, this dissertation introduces DipGNNome, the first deep learning–based pipeline for diploid de novo genome assembly.
Additionally, we introduce grapHiC, a graph transformer trained with a contrastive learning objective on Hi-C data to estimate haplotype allocation probabilities. grapHiC is the first machine learning–based tool to phase de novo assemblies and the first tool to phase nodes directly on the unsimplified assembly graph, potentially reducing haplotype misjoins introduced by graph simplification or pathfinding.
Finally, we integrate these two approaches to enable diploid assembly without trio data, combining DipGNNome’s edge-quality predictions and beam-search–based assembly algorithm with grapHiC’s haplotype assignments. We show that our proposed tools can successfully assemble real genomes under both trio-based and Hi-C–based settings, achieving state-of-the-art performance, while employing fundamentally different, learning-driven methodologies.