Self-supervised learning (SSL) has proven to be a powerful approach for extracting biologically meaningful representations from single-cell data. To advance our understanding of SSL methods applied to single-cell data, we present scSSL-Bench, a comprehensive benchmark that evaluates nineteen SSL methods. Our evaluation spans nine datasets and focuses on three common downstream tasks: batch correction, cell type annotation, and missing modality prediction. Furthermore, we systematically assess various data augmentation strategies. Our analysis reveals task-specific trade-offs: the specialized single-cell frameworks, scVI, CLAIRE, and the finetuned scGPT excel at uni-modal batch correction, while generic SSL methods, such as VICReg and SimCLR, demonstrate superior performance in cell typing and multi-modal data integration. Random masking emerges as the most effective augmentation technique across all tasks, surpassing domain-specific augmentations. Notably, our results indicate the need for a specialized single-cell multi-modal data integration framework. scSSL-Bench provides a standardized evaluation platform and concrete recommendations for applying SSL to single-cell analysis, advancing the convergence of deep learning and single-cell genomics.

Understanding the geometry of neural network loss landscapes is a central question in deep learning, with implications for generalization and optimization. A striking phenomenon is $\textit{linear mode connectivity}$ (LMC), where independently trained models can be connected by low- or zero-barrier paths, despite appearing to lie in separate loss basins. However, this is often obscured by symmetries in parameter space—such as neuron permutations—which make functionally equivalent models appear dissimilar. Prior work has predominantly focused on neuron reordering through permutations, but such approaches are limited in scope and fail to capture the richer symmetries exhibited by modern architectures such as Transformers. In this work, we introduce a unified framework that captures four symmetry classes—permutations, semi-permutations, orthogonal transformations, and general invertible maps—broadening the set of valid reparameterizations and subsuming many previous approaches as special cases. Crucially, this generalization enables, for the first time, the discovery of low- and zero-barrier linear interpolation paths between independently trained Vision Transformers and GPT-2 models. Furthermore, our framework extends beyond pairwise alignment, to multi-model and width-heterogeneous settings, enabling alignment across architectures of different sizes. These results reveal deeper structure in the loss landscape and underscore the importance of symmetry-aware analysis for understanding model space geometry.

Multi-omics data, which include genomic, transcriptomic, epigenetic, and proteomic data, are gaining increasing importance for determining the clinical outcomes of cancer patients. Several recent studies have evaluated various multimodal integration strategies for cancer survival prediction, highlighting the need for standardizing model performance results. Addressing this issue, we introduce SurvBoard, a benchmark framework that standardizes key experimental design choices. SurvBoard enables comparisons between single-cancer and pan-cancer data models and assesses the benefits of using patient data with missing modalities. We also address common pitfalls in preprocessing and validating multi-omics cancer survival models. We apply SurvBoard to several exemplary use cases, further confirming that statistical models tend to outperform deep learning methods, especially for metrics measuring survival function calibration. Moreover, most models exhibit better performance when trained in a pan-cancer context and can benefit from leveraging samples for which data of some omics modalities are missing. We provide a web service for model evaluation and to make our benchmark results easily accessible and viewable: https://www.survboard.science/. All code is available on GitHub: https://github.com/BoevaLab/survboard/. All benchmark outputs are available on Zenodo: 10.5281/zenodo.11066226. A video tutorial on how to use the Survboard leaderboard is available on YouTube at https://youtu.be/HJrdpJP8Vvk.

Humans struggle to perceive and interpret high-dimensional data. Therefore, high-dimensional data are often projected into two dimensions for visualization. Many applications benefit from complex nonlinear dimensionality reduction techniques, but the effects of individual high-dimensional features are hard to explain in the two-dimensional space. Most visualization solutions use multiple two-dimensional plots, each showing the effect of one high-dimensional feature in two dimensions; this approach creates a need for a visual inspection of k plots for a k-dimensional input space. Our solution, Feature Clock, provides a novel approach that reduces the need to inspect these k plots to grasp the influence of original features on the data structure depicted in two dimensions. Feature Clock enhances the explainability and compactness of visualizations of embedded data and is available in an open-source Python library1.

MOTIVATION: Sparse survival models are statistical models that select a subset of predictor variables while modeling the time until an event occurs, which can subsequently help interpretability and transportability. The subset of important features is often obtained with regularized models, such as the Cox Proportional Hazards model with Lasso regularization, which limit the number of non-zero coefficients. However, such models can be sensitive to the choice of regularization hyperparameter. RESULTS: In this work, we develop a software package and demonstrate how knowledge distillation, a powerful technique in machine learning that aims to transfer knowledge from a complex teacher model to a simpler student model, can be leveraged to learn sparse survival models while mitigating this challenge. For this purpose, we present sparsesurv, a Python package that contains a set of teacher-student model pairs, including the semi-parametric accelerated failure time and the extended hazards models as teachers, which currently do not have Python implementations. It also contains in-house survival function estimators, removing the need for external packages. Sparsesurv is validated against R-based Elastic Net regularized linear Cox proportional hazards models as implemented in the commonly used glmnet package. Our results reveal that knowledge distillation-based approaches achieve competitive discriminative performance relative to glmnet across the regularization path while making the choice of the regularization hyperparameter significantly easier. All of these features, combined with a sklearn-like API, make sparsesurv an easy-to-use Python package that enables survival analysis for high-dimensional datasets through fitting sparse survival models via knowledge distillation. AVAILABILITY AND IMPLEMENTATION: sparsesurv is freely available under a BSD 3 license on GitHub (https://github.com/BoevaLab/sparsesurv) and The Python Package Index (PyPi) (https://pypi.org/project/sparsesurv/).

Numerous cancer types have shown to present hypermethylation of CpG islands, also known as a CpG island methylator phenotype (CIMP), often associated with survival variation. Despite extensive research on CIMP, the etiology of this variability remains elusive, possibly due to lack of consistency in defining CIMP. In this work, we utilize a pan-cancer approach to further explore CIMP, focusing on 26 cancer types profiled in the Cancer Genome Atlas (TCGA). We defined CIMP systematically and agnostically, discarding any effects associated with age, gender or tumor purity. We then clustered samples based on their most variable DNA methylation values and analyzed resulting patient groups. Our results confirmed the existence of CIMP in 19 cancers, including gliomas and colorectal cancer. We further showed that CIMP was associated with survival differences in eight cancer types and, in five, represented a prognostic biomarker independent of clinical factors. By analyzing genetic and transcriptomic data, we further uncovered potential drivers of CIMP and classified them in four categories: mutations in genes directly involved in DNA demethylation; mutations in histone methyltransferases; mutations in genes not involved in methylation turnover, such as KRAS and BRAF; and microsatellite instability. Among the 19 CIMP-positive cancers, very few shared potential driver events, and those drivers were only IDH1 and SETD2 mutations. Finally, we found that CIMP was strongly correlated with tumor microenvironment characteristics, such as lymphocyte infiltration. Overall, our results indicate that CIMP does not exhibit a pan-cancer manifestation; rather, general dysregulation of CpG DNA methylation is caused by heterogeneous mechanisms.

BACKGROUND: Multiple studies rely on ChIP-seq experiments to assess the effect of gene modulation and drug treatments on protein binding and chromatin structure. However, most methods commonly used for the normalization of ChIP-seq binding intensity signals across conditions, e.g., the normalization to the same number of reads, either assume a constant signal-to-noise ratio across conditions or base the estimates of correction factors on genomic regions with intrinsically different signals between conditions. Inaccurate normalization of ChIP-seq signal may, in turn, lead to erroneous biological conclusions. RESULTS: We developed a new R package, CHIPIN, that allows normalizing ChIP-seq signals across different conditions/samples when spike-in information is not available, but gene expression data are at hand. Our normalization technique is based on the assumption that, on average, no differences in ChIP-seq signals should be observed in the regulatory regions of genes whose expression levels are constant across samples/conditions. In addition to normalizing ChIP-seq signals, CHIPIN provides as output a number of graphs and calculates statistics allowing the user to assess the efficiency of the normalization and qualify the specificity of the antibody used. In addition to ChIP-seq, CHIPIN can be used without restriction on open chromatin ATAC-seq or DNase hypersensitivity data. We validated the CHIPIN method on several ChIP-seq data sets and documented its superior performance in comparison to several commonly used normalization techniques. CONCLUSIONS: The CHIPIN method provides a new way for ChIP-seq signal normalization across conditions when spike-in experiments are not available. The method is implemented in a user-friendly R package available on GitHub: https://github.com/BoevaLab/CHIPIN.

Motivation: In cancer, clonal evolution is assessed based on information coming from single nucleotide variants and copy number alterations. Nonetheless, existing methods often fail to accurately combine information from both sources to truthfully reconstruct clonal populations in a given tumor sample or in a set of tumor samples coming from the same patient. Moreover, previously published methods detect clones from a single set of variants. As a result, compromises have to be done between stringent variant filtering [reducing dispersion in variant allele frequency estimates (VAFs)] and using all biologically relevant variants. Results: We present a framework for defining cancer clones using most reliable variants of high depth of coverage and assigning functional mutations to the detected clones. The key element of our framework is QuantumClone, a method for variant clustering into clones based on VAFs, genotypes of corresponding regions and information about tumor purity. We validated QuantumClone and our framework on simulated data. We then applied our framework to whole genome sequencing data for 19 neuroblastoma trios each including constitutional, diagnosis and relapse samples. We confirmed an enrichment of damaging variants within such pathways as MAPK (mitogen-activated protein kinases), neuritogenesis, epithelial-mesenchymal transition, cell survival and DNA repair. Most pathways had more damaging variants in the expanding clones compared to shrinking ones, which can be explained by the increased total number of variants between these two populations. Functional mutational rate varied for ancestral clones and clones shrinking or expanding upon treatment, suggesting changes in clone selection mechanisms at different time points of tumor evolution. Availability and implementation: Source code and binaries of the QuantumClone R package are freely available for download at https://CRAN.R-project.org/package=QuantumClone. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

MOTIVATION: Read simulators combined with alignment evaluation tools provide the most straightforward way to evaluate and compare mappers. Simulation of reads is accompanied by information about their positions in the source genome. This information is then used to evaluate alignments produced by the mapper. Finally, reports containing statistics of successful read alignments are created.In default of standards for encoding read origins, every evaluation tool has to be made explicitly compatible with the simulator used to generate reads. RESULTS: To solve this obstacle, we have created a generic format Read Naming Format (Rnf) for assigning read names with encoded information about original positions. Futhermore, we have developed an associated software package RnfTools containing two principal components. MIShmash applies one of popular read simulating tools (among DwgSim, Art, Mason, CuReSim, etc.) and transforms the generated reads into Rnf format. LAVEnder evaluates then a given read mapper using simulated reads in Rnf format. A special attention is payed to mapping qualities that serve for parametrization of Roc curves, and to evaluation of the effect of read sample contamination. AVAILABILITY AND IMPLEMENTATION: RnfTools: http://karel-brinda.github.io/rnftools Spec. of Rnf: http://karel-brinda.github.io/rnf-spec CONTACT: [email protected].

MOTIVATION: Whole genome sequencing of paired-end reads can be applied to characterize the landscape of large somatic rearrangements of cancer genomes. Several methods for detecting structural variants with whole genome sequencing data have been developed. So far, none of these methods has combined information about abnormally mapped read pairs connecting rearranged regions and associated global copy number changes automatically inferred from the same sequencing data file. Our aim was to create a computational method that could use both types of information, i.e. normal and abnormal reads, and demonstrate that by doing so we can highly improve both sensitivity and specificity rates of structural variant prediction. RESULTS: We developed a computational method, SV-Bay, to detect structural variants from whole genome sequencing mate-pair or paired-end data using a probabilistic Bayesian approach. This approach takes into account depth of coverage by normal reads and abnormalities in read pair mappings. To estimate the model likelihood, SV-Bay considers GC-content and read mappability of the genome, thus making important corrections to the expected read count. For the detection of somatic variants, SV-Bay makes use of a matched normal sample when it is available. We validated SV-Bay on simulated datasets and an experimental mate-pair dataset for the CLB-GA neuroblastoma cell line. The comparison of SV-Bay with several other methods for structural variant detection demonstrated that SV-Bay has better prediction accuracy both in terms of sensitivity and false-positive detection rate. AVAILABILITY AND IMPLEMENTATION: https://github.com/InstitutCurie/SV-Bay CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Because of its low cost, amplicon sequencing, also known as ultra-deep targeted sequencing, is now becoming widely used in oncology for detection of actionable mutations, i.e. mutations influencing cell sensitivity to targeted therapies. Amplicon sequencing is based on the polymerase chain reaction amplification of the regions of interest, a process that considerably distorts the information on copy numbers initially present in the tumor DNA. Therefore, additional experiments such as single nucleotide polymorphism (SNP) or comparative genomic hybridization (CGH) arrays often complement amplicon sequencing in clinics to identify copy number status of genes whose amplification or deletion has direct consequences on the efficacy of a particular cancer treatment. So far, there has been no proven method to extract the information on gene copy number aberrations based solely on amplicon sequencing. RESULTS: Here we present ONCOCNV, a method that includes a multifactor normalization and annotation technique enabling the detection of large copy number changes from amplicon sequencing data. We validated our approach on high and low amplicon density datasets and demonstrated that ONCOCNV can achieve a precision comparable with that of array CGH techniques in detecting copy number aberrations. Thus, ONCOCNV applied on amplicon sequencing data would make the use of additional array CGH or SNP array experiments unnecessary.

MOTIVATION: DNA copy number profiles characterize regions of chromosome gains, losses and breakpoints in tumor genomes. Although many models have been proposed to detect these alterations, it is not clear which model is appropriate before visual inspection the signal, noise and models for a particular profile. RESULTS: We propose SegAnnDB, a Web-based computer vision system for genomic segmentation: first, visually inspect the profiles and manually annotate altered regions, then SegAnnDB determines the precise alteration locations using a mathematical model of the data and annotations. SegAnnDB facilitates collaboration between biologists and bioinformaticians, and uses the University of California, Santa Cruz genome browser to visualize copy number alterations alongside known genes. AVAILABILITY AND IMPLEMENTATION: The breakpoints project on INRIA GForge hosts the source code, an Amazon Machine Image can be launched and a demonstration Web site is http://bioviz.rocq.inria.fr.

MOTIVATION: Cancer cells are often characterized by epigenetic changes, which include aberrant histone modifications. In particular, local or regional epigenetic silencing is a common mechanism in cancer for silencing expression of tumor suppressor genes. Though several tools have been created to enable detection of histone marks in ChIP-seq data from normal samples, it is unclear whether these tools can be efficiently applied to ChIP-seq data generated from cancer samples. Indeed, cancer genomes are often characterized by frequent copy number alterations: gains and losses of large regions of chromosomal material. Copy number alterations may create a substantial statistical bias in the evaluation of histone mark signal enrichment and result in underdetection of the signal in the regions of loss and overdetection of the signal in the regions of gain. RESULTS: We present HMCan (Histone modifications in cancer), a tool specially designed to analyze histone modification ChIP-seq data produced from cancer genomes. HMCan corrects for the GC-content and copy number bias and then applies Hidden Markov Models to detect the signal from the corrected data. On simulated data, HMCan outperformed several commonly used tools developed to analyze histone modification data produced from genomes without copy number alterations. HMCan also showed superior results on a ChIP-seq dataset generated for the repressive histone mark H3K27me3 in a bladder cancer cell line. HMCan predictions matched well with experimental data (qPCR validated regions) and included, for example, the previously detected H3K27me3 mark in the promoter of the DLEC1 gene, missed by other tools we tested.

BACKGROUND: Many models have been proposed to detect copy number alterations in chromosomal copy number profiles, but it is usually not obvious to decide which is most effective for a given data set. Furthermore, most methods have a smoothing parameter that determines the number of breakpoints and must be chosen using various heuristics. RESULTS: We present three contributions for copy number profile smoothing model selection. First, we propose to select the model and degree of smoothness that maximizes agreement with visual breakpoint region annotations. Second, we develop cross-validation procedures to estimate the error of the trained models. Third, we apply these methods to compare 17 smoothing models on a new database of 575 annotated neuroblastoma copy number profiles, which we make available as a public benchmark for testing new algorithms. CONCLUSIONS: Whereas previous studies have been qualitative or limited to simulated data, our annotation-guided approach is quantitative and suggests which algorithms are fastest and most accurate in practice on real data. In the neuroblastoma data, the equivalent pelt.n and cghseg.k methods were the best breakpoint detectors, and exhibited reasonable computation times.

MOTIVATION: ChIP-seq consists of chromatin immunoprecipitation and deep sequencing of the extracted DNA fragments. It is the technique of choice for accurate characterization of the binding sites of transcription factors and other DNA-associated proteins. We present a web service, Nebula, which allows inexperienced users to perform a complete bioinformatics analysis of ChIP-seq data. RESULTS: Nebula was designed for both bioinformaticians and biologists. It is based on the Galaxy open source framework. Galaxy already includes a large number of functionalities for mapping reads and peak calling. We added the following to Galaxy: (i) peak calling with FindPeaks and a module for immunoprecipitation quality control, (ii) de novo motif discovery with ChIPMunk, (iii) calculation of the density and the cumulative distribution of peak locations relative to gene transcription start sites, (iv) annotation of peaks with genomic features and (v) annotation of genes with peak information. Nebula generates the graphs and the enrichment statistics at each step of the process. During Steps 3-5, Nebula optionally repeats the analysis on a control dataset and compares these results with those from the main dataset. Nebula can also incorporate gene expression (or gene modulation) data during these steps. In summary, Nebula is an innovative web service that provides an advanced ChIP-seq analysis pipeline providing ready-to-publish results. AVAILABILITY: Nebula is available at http://nebula.curie.fr/ SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: More and more cancer studies use next-generation sequencing (NGS) data to detect various types of genomic variation. However, even when researchers have such data at hand, single-nucleotide polymorphism arrays have been considered necessary to assess copy number alterations and especially loss of heterozygosity (LOH). Here, we present the tool Control-FREEC that enables automatic calculation of copy number and allelic content profiles from NGS data, and consequently predicts regions of genomic alteration such as gains, losses and LOH. Taking as input aligned reads, Control-FREEC constructs copy number and B-allele frequency profiles. The profiles are then normalized, segmented and analyzed in order to assign genotype status (copy number and allelic content) to each genomic region. When a matched normal sample is provided, Control-FREEC discriminates somatic from germline events. Control-FREEC is able to analyze overdiploid tumor samples and samples contaminated by normal cells. Low mappability regions can be excluded from the analysis using provided mappability tracks. AVAILABILITY: C++ source code is available at: http://bioinfo.curie.fr/projects/freec/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: We present a tool for control-free copy number alteration (CNA) detection using deep-sequencing data, particularly useful for cancer studies. The tool deals with two frequent problems in the analysis of cancer deep-sequencing data: absence of control sample and possible polyploidy of cancer cells. FREEC (control-FREE Copy number caller) automatically normalizes and segments copy number profiles (CNPs) and calls CNAs. If ploidy is known, FREEC assigns absolute copy number to each predicted CNA. To normalize raw CNPs, the user can provide a control dataset if available; otherwise GC content is used. We demonstrate that for Illumina single-end, mate-pair or paired-end sequencing, GC-contentr normalization provides smooth profiles that can be further segmented and analyzed in order to predict CNAs. AVAILABILITY: Source code and sample data are available at http://bioinfo-out.curie.fr/projects/freec/.

SUMMARY: ChIP-Seq data are a new challenge for motif discovery. Such a data typically consists of thousands of DNA segments with base-specific coverage values. We present a new version of our DNA motif discovery software ChIPMunk adapted for ChIP-Seq data. ChIPMunk is an iterative algorithm that combines greedy optimization with bootstrapping and uses coverage profiles as motif positional preferences. ChIPMunk does not require truncation of long DNA segments and it is practical for processing up to tens of thousands of data sequences. Comparison with traditional (MEME) or ChIP-Seq-oriented (HMS) motif discovery tools shows that ChIPMunk identifies the correct motifs with the same or better quality but works dramatically faster. AVAILABILITY AND IMPLEMENTATION: ChIPMunk is freely available within the ru_genetika Java package: http://line.imb.ac.ru/ChIPMunk. Web-based version is also available. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: We present SVDetect, a program designed to identify genomic structural variations from paired-end and mate-pair next-generation sequencing data produced by the Illumina GA and ABI SOLiD platforms. Applying both sliding-window and clustering strategies, we use anomalously mapped read pairs provided by current short read aligners to localize genomic rearrangements and classify them according to their type, e.g. large insertions-deletions, inversions, duplications and balanced or unbalanced inter-chromosomal translocations. SVDetect outputs predicted structural variants in various file formats for appropriate graphical visualization. AVAILABILITY: Source code and sample data are available at http://svdetect.sourceforge.net/

MOTIVATION: Genomic sequences are highly redundant and contain many types of repetitive DNA. Fuzzy tandem repeats (FTRs) are of particular interest. They are found in regulatory regions of eukaryotic genes and are reported to interact with transcription factors. However, accurate assessment of FTR occurrences in different genome segments requires specific algorithm for efficient FTR identification and classification. RESULTS: We have obtained formulas for P-values of FTR occurrence and developed an FTR identification algorithm implemented in TandemSWAN software. Using TandemSWAN we compared the structure and the occurrence of FTRs with short period length (up to 24 bp) in coding and non-coding regions including UTRs, heterochromatic, intergenic and enhancer sequences of Drosophila melanogaster and Drosophila pseudoobscura. Tandems with period three and its multiples were found in coding segments, whereas FTRs with periods multiple of six are overrepresented in all non-coding segment. Periods equal to 5-7 and 11-14 were characteristic of the enhancer regions and other non-coding regions close to genes. AVAILABILITY: TandemSWAN web page, stand-alone version and documentation can be found at http://bioinform.genetika.ru/projects/swan/www/ SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.