Mile Šikić

Accurate haplotype phasing is essential for high-quality genome assembly, yet de novo phasing of complex genomes without parental data remains challenging. We formulate haplotype phasing as a node clustering problem with overlapping clusters on augmented unitig graphs, where nodes represent contiguous, non-branching DNA sequence fragments and two edge types can encode sequence overlap or Hi-C proximity information. We introduce a contrastive learning framework with a custom objective function and train a graph-transformer–based model, termed grapHiC, to phase paternal, maternal, and homozygous unitig nodes. grapHiC ; is the first machine-learning–based method to perform reference-free haplotype phasing and the first approach to directly phase raw unitig graphs without prior simplification. We show that grapHiC accurately clusters nodes on human-genome-scale graphs and that its predictions can effectively guide phased de novo genome assembly, producing human assemblies with contiguity and phasing quality comparable to the state of the art when integrated with the DipGNNome assembler.

Modern sequencing technologies have enabled the reconstruction of complete mammalian genomes from telomere to telomere. However, scaling this achievement to thousands of species and population-level studies remains a challenge. Key bottlenecks include the low quality of the draft assemblies and the high coverage requirements. In particular, reconstructing complete and accurate sequences of both haplotypes in diploid genomes is especially difficult since the sequencing depth is not always sufficient to properly reconstruct diverged regions. We aim to explore the use of machine learning, specifically graph neural networks, for scalable error correction in De Bruijn Graphs, addressing the limitations of existing heuristic methods in genome assembly. Inspired by the success of neural networks in extracting patterns from the data on a massive scale, we introduce a method for correcting errors in De Bruijn Graphs using Graph Neural Networks. Our model provides a reliable classification of edges into correct and erroneous, especially for diploid genomes with coverage depth 35 and lower. We demonstrate that these predictions can guide the downstream read error correction algorithm and genome assembly, ultimately allowing for more accurate genome assembly. Machine learning methods have the potential to replace heuristic methods commonly used in genome assembly. Learning-based approaches can enhance the performance of existing assemblers in challenging scenarios and facilitate adaptation to newly sequenced species.

Summary: Genomic changes are central to cancer, so sequencing is essential for understanding tumor evolution and advancing precision medicine, but resolving somatic variants and methylation is challenging in regions with repeats, reference gaps, and germline variants. ^1–4 To identify the full spectrum of somatic variants, including complex rearrangements, we construct and curate near-complete, haplotype-resolved assemblies of the most recent common ancestor of an early-passage, broadly-consented hypodiploid pancreatic cancer cell line and matched normal tissues. ⁵ The tumor assembly recapitulates all 35 chromosomes observed by karyotyping, including translocation-induced hybrid chromosomes containing dynamic dicentric and fused centromeres, nested foldback inversions causing 14 breakpoints with a haplotype switch in a single derivative chromosome, and centromeric satellite tandem duplications up to 136 kbp. Our tumor-normal assembly-based benchmark expands previous multi-technology ensemble benchmark approaches, adding >8,000 somatic small variants altering ∼40 kbp and 149 somatic SVs altering 939 kbp, including 2,764 small variants and 126 SVs not represented on GRCh38. Additionally, 5,111 somatic small variants change representation because they alter germline variants on GRCh38, impacting mutational signatures. Most source-resolved somatic LINE insertions originate from two non-reference germline LINE insertions with accessible promoter elements. These assemblies demonstrate that centromeric, acrocentric, telomeric, and other polymorphic regions, conventionally excluded from analysis, harbor extensive somatic and epigenetic changes and provide curated paired tumor-normal benchmarks to advance algorithms for repetitive regions of cancer genomes.

Telomere-to-telomere (T2T) phased assemblies are emerging as a benchmark for reference-quality genomes^1,17, though they remain technically and financially demanding, particularly at scale. Generating such assemblies for diploid and polyploid genomes typically involves combining high-accuracy long reads, such as PacBio HiFi¹⁶ or the now-deprecated ONT Duplex² reads, with ultra-long ONT Simplex reads. Using multiple platforms or methods increases the cost and the required amount of genomic DNA. Here, we show that comparable results are possible using error correction of ultra-long ONT Simplex reads and then assembling them using state-of-the-art de novo assembly methods. To achieve this, we have developed the deep learning-based HERRO (Haplotype-aware ERRor cOrrection) framework, which corrects ONT Simplex reads while carefully preserving differences in related genomic sequences. Taking into account informative positions that differentiate the haplotypes or genomic repeat copies, HERRO achieves an increase of read accuracy of up to 100-fold for diploid human genomes. By combining HERRO with the Verkko¹⁷ assembler, we reconstruct up to 32 chromosomes telomere-to-telomere, including chromosomes X and Y, and consistently achieve NGA50 values of 100 Mb or higher across several human genomes. HERRO supports both R9.4.1 and R10.4.1 ONT Simplex reads and generalizes well to other species. These results show that error-corrected ONT reads can lower sequencing costs and improve the quality of genomic analyses.

A significant fraction of the microbial diversity remains unclassified, hindering our understanding of microbial roles in health and ecosystems. State-of-the-art methods like Kraken 2 perform well for taxa that are present in the database. However, their accuracy drops significantly when classifying taxa that are not included. While deep learning has advanced many fields, its applications in metagenomics remain limited, and its full potential has yet to be realized. Here, we present Metaxa, a transformer-based deep learning model designed for the taxonomic classification of long-read Nanopore sequences. Metaxa leverages the sequential context of Nanopore reads, enabling robust classification beyond fixed k-mer profiles. Our results show that Metaxa matches Kraken 2 on in-sample data at both the species and genus levels, and significantly outperforms both Kraken 2 and MetageNN at the genus level on out-of-sample datasets where the species genome is absent from the reference database but a different species from the same genus is present. Furthermore, Metaxa demonstrates strong generalization across different Nanopore chemistries (R9.4.1 and R10.4.1). This work highlights the potential of deep learning models to improve metagenomic classification accuracy, especially in complex or underexplored environments where traditional tools fall short.

Nanopore sequencing enables real-time, long-read analysis by processing raw signals as they are produced. A key step, segmentation of signals into events,

Strain-level metagenomic classification is essential for understanding microbial diversity and functional potential, yet remains challenging, particularly when sample composition is unknown and reference databases are large and redundant. Here, we present MADRe, a modular and scalable pipeline for long-read strain-level metagenomic classification based on Metagenome Assembly-Driven Database Reduction. Beyond system-level integration, MADRe introduces statistical strategies that leverage assembly-derived genomic context to guide database reduction and probabilistic read reassignment. Specifically, it combines long-read metagenome assembly, contig-to-reference reassignment using an expectation–maximization framework for reference reduction, and probabilistic read mapping reassignment on a reduced database to achieve sensitive and precise strain-level classification. We extensively evaluated MADRe on simulated datasets, mock communities, and a real anaerobic digester sludge metagenome. Across diverse similarity and coverage conditions, MADRe consistently improves precision by reducing false-positive strain detections. MADRe’s design allows users to apply either the database reduction or read classification step individually. Using only the read classification step shows results on par with other tested tools. MADRe is open source and publicly available at https://github.com/lbcb-sci/MADRe.

Human genome resequencing typically involves mapping reads to a reference genome to call variants; however, this approach suffers from both technical and reference biases, leaving many duplicated and structurally polymorphic regions of the genome unmapped. Consequently, existing variant benchmarks, generated by the same methods, fail to assess these complex regions. To address this limitation, we present a telomere-to-telomere genome benchmark that achieves near-perfect accuracy (i.e. no detectable errors) across 99.4% of the complete, diploid HG002 genome. This benchmark adds 701.4 Mb of autosomal sequence and both sex chromosomes (216.8 Mb), totaling 15.3% of the genome that was absent from prior benchmarks. We also provide a diploid annotation of genes, transposable elements, segmental duplications, and satellite repeats, including 39,144 protein-coding genes across both haplotypes. To facilitate application of the benchmark, we developed tools for measuring the accuracy of sequencing reads, phased variant call sets, and genome assemblies against a diploid reference. Genome-wide analyses show that state-of-the-art de novo assembly methods resolve 2–7% more sequence and outperform variant calling accuracy by an order of magnitude, yielding just one error per 100 kb across 99.9% of the benchmark regions. Adoption of genome-based benchmarking is expected to accelerate the development of cost-effective methods for complete genome sequencing, expanding the reach of genomic medicine to the entire genome and enabling a new era of personalized genomics.

De novo genome assembly remains a central challenge in computational biology, particularly for diploid genomes where maternal and paternal haplotypes must be accurately resolved. Existing assemblers achieve impressive results through carefully designed heuristics, yet modern deep learning methods remain largely unexplored in the diploid setting. We present DipGNNome , the first deep learning–based framework for diploid de novo genome assembly. Our approach formulates assembly as an edge classification and traversal problem on haplotype-aware assembly graphs, training graph neural networks (GNNs) to guide contig construction. To enable this, we establish a novel pipeline for generating diploid graphs with ground-truth edge labels, providing the first systematic way to produce training data for machine learning models in this domain. This framework creates a foundation for applying and extending graph-based deep learning to diploid assembly. DipGNNome creates assemblies comparable to state-of-the-art and demonstrates the feasibility of deep learning for diploid assembly and introduces a paradigm that bridges algorithmic genomics with graph representation learning. Our code, dataset and trained model is openly available at https://github.com/lbcb-sci/DipGNNome .