GTM: a dual-branch local-global graph learning framework with path-aware optimization for genome assembly

Genome assembly aims to reconstruct complete genome sequences from sequencing reads and serves as the foundation for downstream analyses including gene annotation, structural variant detection, comparative genomics, and evolutionary studies. Although HiFi reads offer both long lengths and high accuracy, complex repetitive regions continue to introduce branches and ambiguous connections in assembly graphs, complicating the identification of correct paths. Recent graph neural network and Graph Transformer approaches have shown promise in edge scoring, yet most rely on a single modeling strategy and struggle to capture both local and global topological features. Issues such as class imbalance, intra-path inconsistency, and low-confidence predictions further limit performance. We propose GTM, a dual-branch local-global graph learning framework with path-aware optimization. GTM integrates GatedGCN for local topology, parallel Graph Transformer and Graph Mamba branches for global modeling, and a multi-objective loss that combines Asymmetric Loss, negative sample auxiliary loss, path consistency loss, and confidence regularization. Experiments on six real HiFi datasets (CHM13, A. thaliana , G. gallus , M. musculus , tomato, and maize) show that GTM generally produces relatively few contigs, achieves competitive NA50 and NGA50 values, and maintains favorable structural accuracy across the evaluated genomes. Controlled cross-training experiments further suggest a degree of cross-species transferability while also revealing a notable training-domain effect. The HG002-trained model retained some generalization capacity to plant genomes, whereas the rice-trained model generally showed lower contig counts and fewer misassemblies on the three primary plant benchmark datasets. A similar structural pattern was observed when both pretrained models were additionally evaluated on real Nipponbare HiFi reads, providing further support for the influence of the training domain on structural assembly performance. These findings suggest that GTM retains a degree of cross-species transferability while its assembly performance remains influenced by the training domain. The code is available from https://github.com/HpuBioinformatics/GTM . GTM offers a graph-learning framework for HiFi genome assembly that achieves encouraging performance across multiple genomes and exhibits a degree of cross-species transferability. At the same time, the observed differences between the HG002-trained and rice-trained models indicate that performance remains influenced by the training domain and suggest that rice-genome-derived simulated training data may provide additional benefits for structural assembly on plant genomes.