DipGNNome

Reconstructing a genome from sequencing reads without a reference is hard enough for a haploid organism; for a diploid one it is harder still, because the maternal and paternal copies of each chromosome are similar but not identical, and an assembler must keep them apart rather than collapse them into a consensus. The assemblers that do this well today — hifiasm and Verkko — are built almost entirely from hand-designed graph heuristics.

DipGNNome is a deep-learning assembler for diploid de novo genome assembly. It recasts assembly as an edge classification and traversal problem on a haplotype-aware assembly graph: a graph neural network scores every edge for correctness, and a beam search then walks the scored graph to emit two independent sets of contigs, one per haplotype. It was developed by Martin Schmitz, Lovro Vrček, Kenji Kawaguchi, and Mile Šikić at the National University of Singapore School of Computing, the Genome Institute of Singapore at A*STAR, and the University of Zagreb, posted as a preprint in September 2025 and published in the RECOMB-CG 2026 proceedings.

The work extends GNNome, the same group's earlier haploid GNN assembler, into the diploid regime. The harder half of that extension is supervision rather than modelling: producing ground-truth edge labels when a read's coordinates live in one haplotype's reference frame and its overlap partner's live in the other. DipGNNome ships the pipeline that solves this, along with the resulting labelled dataset.

#Key Features

• Haplotype-aware graph construction: HiFi reads are assembled into a unitig graph with hifiasm, transitively reduced, and annotated with maternal and paternal unique k-mer counts from parental Illumina reads via yak trio binning.
• Ground-truth labelling across two coordinate systems: Liftover-based translation between homologous loci lets an overlap be validated in either haplotype's frame, yielding the first systematic supervision signal for diploid assembly graphs.
• Diploid-adapted beam search: Contig construction adds best-beam tracking, complement enforcement preventing a node and its reverse complement from co-occurring, and beam merging that collapses redundant paths — consistently beating a greedy variant.
• Released checkpoint applied without retraining: A single model trained on synthetic human data (dipgnnome_trained.pt) is applied directly to seven real genomes, including six non-human apes.
• Open code and training data: The repository ships the graph-generation pipeline, training code, decoding algorithm, and checkpoint; the labelled dataset is distributed separately in DGL/PyG-ready form.

#Technical Details

The scoring network is a SymGatedGCN — GNNome's symmetric extension of GatedGCN, which runs double message passing to capture edge directionality, with PairNorm after each layer to prevent oversmoothing. The released configuration uses 8 GNN layers, 512 hidden dimensions, 4 node features (read support, in/out degree, haplotype k-mer counts) and a single edge feature (overlap length), trained with dropout 0.2 at a learning rate of 1e-4 for up to 200 epochs. A jointly trained MLP head emits two scores per edge — a general correctness score and a translocation probability — optimized with a symmetric binary cross-entropy loss.

Training data is entirely synthetic: PBSIM3 simulates 40x HiFi coverage (20x per haplotype) from the maternal and paternal references of every chromosome of the I002C human genome, producing roughly 7.7M nodes and 10.5M edges, of which about 85% of edges are correct. METIS partitions each graph into subgraphs of at most 40,000 nodes for mini-batch training. Evaluation is on real data only: HG002 plus the six telomere-to-telomere ape references. On human, beam search reaches an NG50 of 63.0/65.2 Mb and NGA50 of 48.4/49.3 Mb (paternal/maternal) against hifiasm's 49.7/58.6 Mb NGA50; on P. pygmaeus it reaches 78.9/92.2 Mb NGA50 versus hifiasm's 55.7/47.4 Mb. hifiasm retains lower Hamming error across most genomes (0.8/0.8% versus 1.9/3.0% on human), which the authors attribute to the read-level polishing steps their pipeline lacks.

#Applications

The immediate audience is groups producing reference-quality phased assemblies from PacBio HiFi reads with parental short-read data, particularly in comparative and evolutionary genomics, where the great-ape results show the checkpoint transfers across species without retraining. Where parental data are unavailable, haplotype k-mers can be derived from an existing reference. The released dataset is arguably the broader contribution: it gives graph-learning researchers a large benchmark for path-finding on string-like graphs without requiring a genomics pipeline first.

#Impact

DipGNNome establishes that diploid assembly is tractable as a learned graph problem, and supplies the labelled training data that had been the blocking gap. Its limits are stated plainly by the authors: it depends on hifiasm for the initial overlap graph, carries no polishing stage and therefore higher Hamming error, and its checkpoint was trained on a single human reference. Reuse terms are unsettled — the README declares a BSD 3-Clause license but ships no license file, and the preprint is CC-BY-NC-ND. It claims no generality beyond diploid assembly, but is a concrete bridge between algorithmic genomics and graph representation learning.