Biology subjects

Sikic, M.

Publications and source records attributed to Sikic, M..

At least 19 recordsLinked to original sources

bioRxiv2026

Binary node clustering via contrastive learning for haplotype phasing in de novo genome assembly

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 grapHiCaccurately 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.

bioinformatics↗

bioRxiv2026

A complete human pancreatic cancer genome

Cancer genome sequencing is essential for understanding tumor evolution and advancing precision medicine.1 However, reference gaps and germline variants obscure detection of small and large somatic variants and methylation in repetitive regions.1-3 It is common for tumor cells to gain or lose chromosome arms due to somatic structural changes that occur inside highly repetitive satellite DNA sequences in the centromeres.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. The tumor assembly completely recapitulates all 35 tumor chromosomes observed with karyotyping, with multiple translocation-induced hybrid chromosomes. The hybrid chromosomes contain putative functional dicentric and fused centromeres, nested foldback inversions causing 14 breakpoints with a haplotype switch in a single event, and centromeric satellite tandem duplications up to 136 kbp. Direct comparison of tumor and normal assembly haplotypes uncovers >7,000 variants altering >1 Mbp of sequence in repetitive regions that have been hidden by reference gaps and germline variants. 44 % of somatic small variants change representation because they alter germline variants on GRCh38, impacting mutational signatures and kataegis/omikli clusters. Most somatic LINE insertions originate from two hypomethylated non-reference germline LINE insertions, highlighting their impact on insertion mutation burden. These assemblies demonstrate that centromeric, acrocentric, and telomeric regions conventionally excluded from analysis harbor extensive somatic and epigenetic changes. Resolving complete tumor genomes enables a deeper understanding of cancer structural plasticity and the endpoints of breakage-fusion-bridge cycles. These assembled, curated paired normal-tumor benchmarks will serve as a critical foundation for developing future algorithms to characterize the most intractable regions of cancer genomes.

genomics↗

bioRxiv2026

Metaxa: A Transformer-Based Deep Learning Model for Taxonomic Classification of Long Nanopore Reads

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.

bioinformatics↗

bioRxiv2025

DipGNNome: Diploid de novo genome assembly with geometric deep learning and beam-search

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.

genomics↗

bioRxiv2025

RNA-xLSTM: Evaluating xLSTM as an Alternative Foundation to Transformers in RNA Modeling

Transformer-based architectures currently achieve state-of-the-art performance across a wide range of domains, including biological sequence modeling. Motivated by the recent introduction of the xLSTM architecture, we investigate its effectiveness for RNA sequence modeling by comparing a 33.7M-parameter RNA-xLSTM model against two leading RNA language models: RNA-FM and RiNALMo-33M. We pretrain RNA-xLSTM on the RNAcentral database and evaluate its performance on two downstream tasks: RNA secondary structure prediction and splice site prediction. Our results show that while RNA-xLSTM underperforms compared to the similarly sized RiNALMo, it does outperform the larger RNA-FM model on certain tasks. However, its overall performance remains inconsistent, and its advantages over transformer-based models are unclear, suggesting that further work is needed to assess its true potential in RNA modeling.

bioinformatics↗

bioRxiv2025

A Complete Telomere-to-Telomere Diploid Reference Genome for Indian Population

Human reference genomes have been instrumental in advancing genomic and biomedical research, but South and Southeast Asian populations are underrepresented, despite accounting for a large proportion of world population. As a part of effort on generating reference genomes for these populations, we present the first gapless, telomere-to-telomere (T2T) diploid genome assembly created by using a trio sample set of Indian ancestry (I002C), with NG50 of 154.89 Mb and 146.27 Mb for the maternal and paternal haplotypes, including the fully assembled rDNA array for the maternal chromosome 21 and Y chromosome. With the Merqury QVs of 82.05, 83.08 and 82.64 for the maternal, paternal and haploid assemblies respectively, I002C represents the highest-quality human genome assembled in both diploid and haploid forms to date. Compared to CHM13, the I002C genome displays substantial sequence diversity, resulting in 14,943 structural variants, including 3,236 novel variants absent from public databases. Analysis of trio-phased haplotypes further revealed elevated inter-haplotype divergence within centromeric and subtelomeric regions, along with identification of differentially methylated regions (DMRs) as candidates for novel imprinting loci. As a result of substantial SVs between them, I002C is a more suitable reference than CHM13 for the genomic analysis of South Asian samples with less reference bias and better performance in mapping and variant calling, particularly for long read sequencing data. As the first high-quality T2T diploid reference genome for Indian, the largest worlds population, I002C contributes to the growing set of population-specific reference genomes and helps to overcome a significant gap in human genome diversity.

genomics↗

bioRxiv2025

Campolina: A Deep Neural Framework for Accurate Segmentation of Nanopore Signals

Nanopore sequencing enables real-time, long-read analysis by processing raw signals as they are produced. A key step, segmentation of signals into events, is typically handled algorithmically, struggling in noisy regions. We present Campolina, a first deep-learning frame-work for accurate segmentation of raw nanopore signals. Campolina uses a convolutional model to identify event boundaries and significantly outperforms the traditional Scrappie algorithm on R9.4.1 and R10.4.1 datasets. We introduce a comprehensive evaluation pipeline and show that Campolina aligns better with reference-guided ground-truth segmentation. We show that integrating Campolina segmentation into real-time frameworks, Sigmoni and RawHash2, improves their performance while maintaining time efficiency.

bioinformatics↗

bioRxiv2025

MADRe: Strain-Level Metagenomic Classification Through Assembly-Driven Database Reduction

AbstractStrain-level metagenomic classification is essential for understanding microbial diversity and functional potential, but remains challenging, par- ticularly in the absence of prior knowledge about the composition of the sample. In this paper we present MADRe, a modular and scalable pipeline for long-read strain-level metagenomic classification, enhanced with Metagenome Assembly-Driven Database Reduction. MADRe com- bines long-read metagenome assembly, contig-to-reference mapping reas- signment based on an expectation-maximization algorithm for database reduction, and probabilistic read mapping reassignment to achieve sensi- tive and precise classification. We extensively evaluated MADRe on sim- ulated datasets, mock communities, and a real anaerobic digester sludge metagenome, demonstrating that it consistently outperforms existing tools by achieving higher precision with reduced false positives. MADRes de- sign allows users to apply either the database reduction or read classi- fication step individually. Using only the read classification step shows results on par with other tested tools. MADRe is open source and pub- licly available at https://github.com/lbcb-sci/MADRe.

bioinformatics↗

bioRxiv2025

GnnDebugger: GNN based error correction in De Bruijn Graphs

MotivationModern 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. 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. ResultsOur 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. Availability and implementationBoth GnnDebugger (https://github.com/m5imunovic/gnndebugger) and LJA (https://github.com/AntonBankevich/LJA/tree/gnndebugger) are available on GitHub. Datasets used for training and testing of ML model are available at Zenodo: https://doi.org/10.5281/zenodo.15073168. HG002 reference and reads are available at https://github.com/marbl/HG002. Primates references and reads are available at https://github.com/marbl/Primates.

bioinformatics↗

bioRxiv2024

A Comparative Review of Deep Learning Methods for RNA Tertiary Structure Prediction

Several deep learning-based tools for RNA 3D structure prediction have recently emerged, including DRfold, DeepFoldRNA, RhoFold, RoseTTAFoldNA, trRosettaRNA, and AlphaFold 3. In this study, we systematically evaluate these six models on three datasets: RNA Puzzles, CASP15 RNA targets, and a newly generated dataset of sequentially distinct RNAs, which serves as a benchmark for generalization capabilities. To ensure a robust evaluation, we also introduce a fourth, more stringent dataset that contains both sequentially and structurally distinct RNAs. We observed that each model predicts the best structure for certain RNAs, and evaluated whether commonly used scoring functions, Rosetta score and ARES, can reliably identify the most accurate structure from the predictions. Finally, since many RNA chains in the Protein Data Bank are part of complexes, we compare the performance of RoseTTAFoldNA and AlphaFold 3 in predicting RNA structures within complexes versus isolated RNA chains extracted from these complexes. This comprehensive evaluation highlights the strengths and limitations of current deep learning-based tools and provides valuable insights for advancing RNA 3D structure prediction.

bioinformatics↗

bioRxiv2024

Telomere-to-telomere phased genome assembly using error-corrected Simplex nanopore reads

Telomere-to-telomere phased assemblies have become the norm in genomics. To achieve these for diploid and even polyploid genomes, the contemporary approach involves a combination of two long-read sequencing technologies: high-accuracy long reads, e.g. Pacific Biosciences (PacBio) HiFi or Oxford Nanopore (ONT) Duplex reads, and ultra-long ONT Simplex reads. Using two different technologies 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 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 Verkko assembler, we achieve high contiguity on several human genomes by reconstructing many chromosomes telomere-to-telomere, including chromosomes X and Y. HERRO supports both R9.4.1 and R10.4.1 ONT Simplex reads and generalizes well to other species. These results provide an opportunity to reduce the cost of genome sequencing and use corrected ONT reads to analyze more complex genomes with different levels of ploidy or even aneuploidy.

bioinformatics↗

bioRxiv2024

The Hitchhiker's Guide to Sequencing Data Types and Volumes for Population-Scale Pangenome Construction

Long-read (LR) technologies from Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) have transformed genomics research by providing diverse data types like HiFi, Duplex, and ultra-long ONT (ULONT). Despite recent strides in achieving haplotype-phased gapless genome assemblies using long-read technologies, concerns persist regarding the representation of genetic diversity, prompting the development of pangenome references. However, pangenome studies face challenges related to data types, volumes, and cost considerations for each assembled genome, while striving to maintain sensitivity. The absence of comprehensive guidance on optimal data selection exacerbates these challenges. To fill this gap, our study evaluates available data types, their significance, and the required volumes for robust de novo assembly in population-level pangenome projects. The results show that achieving chromosome-level haplotype-resolved assembly requires 20x high-quality long reads (HQLR) such as PacBio HiFi or ONT duplex, combined with 15-20x of ULONT per haplotype and 30x of long-range data such as Omni-C. High-quality long reads from both platforms yield assemblies with comparable contiguity, with HiFi excelling in NG50 and phasing accuracies, while usage of duplex generates more T2T contigs. As Long-Read Technologies advance, our study reevaluates recommended data types and volumes, providing practical guidelines for selecting sequencing platforms and coverage. These insights aim to be vital to the pangenome research community, contributing to their efforts and pushing genomic studies with broader impacts.

bioinformatics↗

bioRxiv2024

Geometric deep learning framework for de novo genome assembly

The critical stage of every de novo genome assembler is identifying paths in assembly graphs that correspond to the reconstructed genomic sequences. The existing algorithmic methods struggle with this, primarily due to repetitive regions causing complex graph tangles, leading to fragmented assemblies. Here, we introduce GNNome, a framework for path identification based on geometric deep learning that enables training models on assembly graphs without relying on existing assembly strategies. By leveraging symmetries inherent to the problem, GNNome reconstructs assemblies with similar or superior contiguity compared to the state-of-the-art tools across several species, sequenced with PacBio HiFi or Oxford Nanopore. With every new genome assembled telomere-to-telomere, the amount of reliable training data at our disposal increases. Combining the straightforward generation of abundant simulated data for diverse genomic structures with the AI approach makes the proposed framework a plausible cornerstone for future work on reconstructing complex genomes with different ploidy and aneuploidy degrees. To facilitate such developments, we make the framework and the best-performing model publicly available, provided as a tool that can directly be used to assemble new haploid genomes.

bioinformatics↗

bioRxiv2022

Rockfish: A Transformer-based Model for Accurate 5-Methylcytosine Prediction from Nanopore Sequencing

DNA methylation plays a crucial role in various biological processes, including cell differentiation, ageing, and cancer development. The most important methylation in mammals is 5-methylcytosine (5mC) which is present in the context of CpG dinucleotides. Sequencing methods such as whole-genome bisulfite sequencing (WGBS) successfully detect 5mC DNA modifications. However, they suffer from the serious drawbacks of short read lengths and might introduce an amplification bias. Here we present Rockfish, a deep learning algorithm that significantly improves read-level 5mC detection by using Nanopore sequencing. Compared to other methods based on Nanopore sequencing, there is an increase in the single-base accuracy and the F1 measure of up to 5% and 12%, respectively. Furthermore, Rockfish shows a high correlation with WGBS and requires lower read depth while being computationally efficient. We deem that Rockfish is broadly applicable to study 5mC methylation in diverse organisms and disease systems to yield biological insights.

bioinformatics↗

bioRxiv2021

A systematic benchmark of Nanopore long read RNA sequencing for transcript level analysis in human cell lines

The human genome contains more than 200,000 gene isoforms. However, different isoforms can be highly similar, and with an average length of 1.5kb remain difficult to study with short read sequencing. To systematically evaluate the ability to study the transcriptome at a resolution of individual isoforms we profiled 5 human cell lines with short read cDNA sequencing and Nanopore long read direct RNA, amplification-free direct cDNA, PCR-cDNA sequencing. The long read protocols showed a high level of consistency, with amplification-free RNA and cDNA sequencing being most similar. While short and long reads generated comparable gene expression estimates, they differed substantially for individual isoforms. We find that increased read length improves read-to-transcript assignment, identifies interactions between alternative promoters and splicing, enables the discovery of novel transcripts from repetitive regions, facilitates the quantification of full-length fusion isoforms and enables the simultaneous profiling of m6A RNA modifications when RNA is sequenced directly. Our study demonstrates the advantage of long read RNA sequencing and provides a comprehensive resource that will enable the development and benchmarking of computational methods for profiling complex transcriptional events at isoform-level resolution.

genomics↗

bioRxiv2020

Benchmarking metagenomic classification tools for long-read sequencing data

BackgroundLong reads have gained popularity in the analysis of metagenomics data. Therefore, we comprehensively assessed metagenomics classification tools on the species taxonomic level. We analysed kmer-based tools, mapping-based tools and two general-purpose long reads mappers. We evaluated more than 20 pipelines which use either nucleotide or protein databases and selected 13 for an extensive benchmark. We prepared seven synthetic datasets to test various scenarios, including the presence of a host, unknown species and related species. Moreover, we used available sequencing data from three well-defined mock communities, including a dataset with abundance varying from 0.0001% to 20% and six real gut microbiomes. ResultsGeneral-purpose mappers Minimap2 and Ram achieved similar or better accuracy on most testing metrics than best-performing classification tools. They were up to ten times slower than the fastest kmer-based tools requiring up to four times less RAM. All tested tools were prone to report organisms not present in datasets, except CLARK-S, and they underperformed in the case of the high presence of the hosts genetic material. Tools which use a protein database performed worse than those based on a nucleotide database. Longer read lengths made classification easier, but due to the difference in read length distributions among species, the usage of only the longest reads reduced the accuracy. The comparison of real gut microbiome datasets shows a similar abundance profiles for the same type of tools but discordance in the number of reported organisms and abundances between types. Most assessments showed the influence of database completeness on the reports. ConclusionThe findings indicate that kmer-based tools are well-suited for rapid analysis of long reads data. However, when heightened accuracy is essential, off-the-shelf mappers demonstrate slightly superior performance, albeit at a considerably slower pace. Nevertheless, a combination of diverse categories of tools and databases will likely be necessary to analyse complex samples. Discrepancies observed among tools when applied to real gut datasets, as well as a reduced performance in cases where unknown species or a significant proportion of the host genome is present in the sample, highlight the need for continuous improvement of existing tools. Additionally, regular updates and curation of databases are important to ensure their effectiveness.

bioinformatics↗

bioRxiv2020

Raven: a de novo genome assembler for long reads

We present new methods for the improvement of de novo genome assembly from erroneous long-reads incorporated into a straightforward tool called Raven (https://github.com/lbcb-sci/raven). Raven maintains similar performance for various genomes and has accuracy on par with other assemblers which support third-generation sequencing data. It is one of the fastest options while having the lowest memory consumption on the majority of benchmarked datasets.

bioinformatics↗

bioRxiv2019

Graphmap2 - splice-aware RNA-seq mapper for long reads

In this paper we present Graphmap2, a splice-aware mapper built on our previously developed DNA mapper Graphmap. Graphmap2 is tailored for long reads produced by Pacific Biosciences and Oxford Nanopore devices. It uses several newly developed algorithms which enable higher precision and recall of correctly detected transcripts and exon boundaries. We compared its performance with the state-of-the-art tools Minimap2 and Gmap. On both simulated and real datasets Graphmap2 achieves higher mappability and more correctly recognized exons and their ends. In addition we present an analysis of potential of splice aware mappers and long reads for the identification of previously unknown isoforms and even genes. The Graphmap2 tool is publicly available at https://github.com/lbcb-sci/graphmap2.

bioinformatics↗