Analogous to the Encyclopedia of DNA Elements (ENCODE) project, the Functional Annotation of ANimal Genomes (FAANG) consortium has produced chromatin annotations for domesticated animals, albeit in smaller amounts. Although acquiring experimental data is more accessible and affordable for many species, human and mouse organisms will remain the reference. Classical methods based on sequence conservation can be used to infer missing annotations, but are inappropriate for non-conserved sequences. While regulatory sequences share low to moderate conservation, they have retained their regulatory function during the evolution process. Here, we take advantage of three neural networks (DeepBind, DeepSEA, and Enformer) trained with human and mouse ENCODE data to infer chromatin annotations (transcription factors binding, chromatin accessibility, and histone marks) in cattle, pig, chicken, and European seabass. For this purpose, we comprehensively assessed the quality of predictions using experimental data from FAANG, through AUC-ROC and AUC-PR metrics. Our results showed low variability between tissues and similar performances for various annotations in mammals and chicken, with AUC-PR ranging from 0.663 ± 0.010 (chicken) to 0.765 ± 0.009 (pig) for the best predicted experiment, H3K4me3 (all tissues grouped), but lower (0.238 ± 0.005) in fish. Further analyses focused on pigs highlighted (i) accurate predictions even for non-conserved sequences, and (ii) variable predictions depending on genomic feature annotations. Our results advocate the widespread use of human-trained neural networks as a first step in cross-species genome annotation before training species-specific models.

Abstract Genome assembly is a computational pipeline designed to reconstruct chromosomes from small sequencing reads. Following their assembly, contiguous sequences (contigs) are arranged into chromosome-long sequences during scaffolding. Hi-C, a long-range linkage information between regions of the genome widely used in recent large sequencing projects, is often required to correctly order contigs. Several tools have been developed to automate this task following either statistical or deep-learning approaches. Statistical approaches summarise 2D Hi-C matrices into contact densities across sequences, thus ignoring informative visual patterns. The sole existing deep-learning tool uses a transformer-based computer vision model to correct the assembly. It has been trained on several species and uses Hi-C matrices directly. Yet it comes as a supplementary step in the scaffolding process, introducing extra computation time, and has been trained on a dataset that might contain labelling errors, which could provide sub-optimal results. We propose SPA-C , an hybrid pipeline combining the strengths of both approaches. Linkage prediction is handled with a frugal CNN-based model and a graph-solving algorithm is used to generate the scaffolds. Through our input’s design, the model is able to both correct errors within assemblies and link contigs, leveraging small, local Hi-C contact matrices. We handled low-complexity regions that might induce erroneous predictions using an external tool, improving the overall accuracy of generated assemblies. On a benchmark of six various genomes and four standard metrics, SPA-C outperformed four out of four state-of-the-art methods while achieving comparable start-to-end computation time. Python and Bash scripts are available on GitHub ( github.com/SPA-C/SPA-C.git ) and Zenodo ( 10.5281/zenodo.19000361 ).

Symptomatic lumbar spinal stenosis (LSS) is a disabling condition with a substantial economic impact. The determination of surgical candidacy for LSS relies on a subjective assessment of multiple clinical and imaging factors, leading to variability in recommendations. Artificial intelligence (AI), including traditional machine learning (ML) and emerging large language models (LLMs), holds promise for enhancing the accuracy and consistency of surgical decision-making. However, the comparative performance of LLMs against established ML methods for this specific task remains unclear. This study therefore benchmarks various LLM configurations against a conventional random forest model in predicting LSS surgery recommendation using structured clinical vignettes. Various LLMs, including zero-shot ChatGPT3.5, fine-tuned GPT-NEO (2.7B and 125 M parameters), and BioMedLM, were benchmarked against a random forest model. Using a ground truth established by five spine surgeons from 500 synthetic medical vignettes (36 variables each), models provided surgical recommendations. LLMs received these vignette data in text format; the random forest received structured tabular format, to simulate realistic LSS surgical candidates. Performance was evaluated using root mean square error (RMSE) for regression and area under the receiver operating curve (AUROC) for binary classification. For surgical recommendation, zero-shot ChatGPT3.5 achieved an RMSE of 0.37 (AUROC = 0.79). The random forest model, trained on the structured data, achieved superior performance (RMSE = 0.12, AUROC = 0.96). Fine-tuned LLMs showed improved accuracy: GPT-NEO 125 m achieved an RMSE of 0.224 (AUROC = 0.81), BioMedLM an RMSE of 0.258 (AUROC = 0.86), and GPT-NEO 2.7b achieved an RMSE of 0.275 (AUROC = 0.90). When benchmarked on structured clinical vignettes, LLMs demonstrate credible but inferior predictive accuracy for LSS surgery recommendation compared to a conventional random forest model. These findings highlight that for well-structured data, specialized ML models retain an advantage. The critical next step is to evaluate LLMs on unstructured, real-world clinical data where their natural language processing strengths may be decisive.

Analogous to the Encyclopedia of DNA Elements (ENCODE) project, the Functional Annotation of ANimal Genomes (FAANG) consortium has produced chromatin annotations for domesticated animals, albeit in smaller amounts. Classical methods based on sequence conservation can be used to infer missing annotations, but are inappropriate for non-conserved sequences. Here, we demonstrate the ability of neural networks trained with human data to infer the missing chromatin annotations in livestock species. For this purpose, we comprehensively assessed predictions of transcription factors, chromatin accessibility, and histone marks in several species. Our results showed good predictions for various annotations in mammalian genomes, and surprisingly, also for bird genomes, despite the large phylogenetic distance from the human genome. Moreover, predictions were accurate even for non-conserved sequences, unlike conservation-based methods. Our results advocate the widespread use of neural networks in cross-species genome annotation, a key step in understanding the genetic architecture of complex traits. ### Competing Interest Statement The authors have declared no competing interest. Agence Nationale de la Recherche, PDOC2 – ANR-22-PEAE-0015 PEAE 2022 INRAE, métaprogramme DIGIT-BIO, OBAMA

Despite their central role in transcription, it has been difficult to define universal sequences associated with eukaryotic promoters. Within the chromatin context, recruitment of transcriptional machinery requires promoter opening, but how DNA elements contribute to this process is unclear. Here we show that G-quadruplex (G4) secondary DNA structures are highly enriched at mammalian promoters. G4s are located at the deepest point of nucleosome exclusion at promoters and correlate with maximum promoter activity. We found that experimental G4s exclude nucleosomes in vivo and in vitro while favouring strong positioning. At model promoters, impairing G4s affected both transcriptional activity and chromatin opening. G4 destabilization also resulted in an inactive promoter state and affected the transition to effective RNA production. Finally, G4 stabilization resulted in global reduction of proximal promoter pausing. Altogether, our data introduce G4s as bona fide promoter elements allowing nucleosome exclusion and facilitating pause-release by RNA polymerase II.

Predicting molecular processes using deep learning is a promising approach to provide biological insights for non-coding single nucleotide polymorphisms identified in genome-wide association studies. However, most deep learning methods rely on supervised learning, which requires DNA sequences associated with functional data, and whose amount is severely limited by the finite size of the human genome. Conversely, the amount of mammalian DNA sequences is growing exponentially due to ongoing large-scale sequencing projects, but in most cases without functional data. To alleviate the limitations of supervised learning, we propose a novel semi-supervised learning (SSL) based on pseudo-labeling, which allows to exploit unlabeled DNA sequences from numerous genomes during model pre-training. We further improved it incorporating principles from the Noisy Student algorithm to predict the confidence in pseudo-labeled data used for pre-training, which showed improvements for transcription factor with very few binding (very small training data). The approach is very flexible and can be used to train any neural architecture including state-of-the-art models, and shows in most cases strong predictive performance improvements compared to standard supervised learning. Moreover, small models trained by SSL showed similar or better performance than large language model DNABERT2.

- OBJECTIVES: Symptomatic lumbar spinal stenosis (LSS) leads to functional impairment and pain. While radiologic characterization of the morphological stenosis grade can aid in the diagnosis, it may not always correlate with patient symptoms. Artificial intelligence (AI) may diagnose symptomatic LSS in patients solely based on self-reported history questionnaires. - METHODS: We evaluated multiple machine learning (ML) models to determine the likelihood of LSS using a self-reported questionnaire in patients experiencing low back pain and/or numbness in the legs. The questionnaire was built from peer-reviewed literature and a multidisciplinary panel of experts. Random forest, lasso logistic regression, support vector machine, gradient boosting trees, deep neural networks, and automated machine learning models were trained and performance metrics were compared. - RESULTS: Data from 4827 patients (4690 patients without LSS: mean age 62.44, range 27-84 years, 62.8% females, and 137 patients with LSS: mean age 50.59, range 30- 71 years, 59.9% females) were retrospectively collected. Among the evaluated models, the random forest model demonstrated the highest predictive accuracy with an area u nder the receiver operating characteristic curve (AUROC) between model prediction and LSS diagnosis of 0.96, a sensitivity of 0.94, a specificity of 0.88, a balanced accuracy of 0.91, and a Cohen's kappa of 0.85. - CONCLUSIONS: Our results indicate that ML can automate the diagnosis of LSS based on self-reported questionnaires with high accuracy. Implementation of standardized and intelligence-automated workflow may serve as a supportive diagnostic tool to streamline patient management and potentially lower health care costs.

The recent development of deep learning methods have undoubtedly led to great improvement in various machine learning tasks, especially in prediction tasks. This type of methods have also been adapted to answer various problems in bioinformatics, including automatic genome annotation, artificial genome generation or phenotype prediction. In particular, a specific type of deep learning method, called graph neural network (GNN) has repeatedly been reported as a good candidate to predict phenotypes from gene expression because its ability to embed information on gene regulation or co-expression through the use of a gene network. However, up to date, no complete and reproducible benchmark has ever been performed to analyze the trade-off between cost and benefit of this approach compared to more standard (and simpler) machine learning methods. In this article, we provide such a benchmark, based on clear and comparable policies to evaluate the different methods on several datasets. Our conclusion is that GNN rarely provides a real improvement in prediction performance, especially when compared to the computation effort required by the methods. Our findings on a limited but controlled simulated dataset shows that this could be explained by the limited quality or predictive power of the input biological gene network itself.

Genome-wide association studies have systematically identified thousands of single nucleotide polymorphisms (SNPs) associated with complex genetic diseases. However, the majority of those SNPs were found in non-coding genomic regions, preventing the understanding of the underlying causal mechanism. Predicting molecular processes based on the DNA sequence represents a promising approach to understand the role of those non-coding SNPs. Over the past years, deep learning was successfully applied to regulatory sequence prediction using supervised learning. Supervised learning required DNA sequences associated with functional data for training, whose amount is strongly limited by the finite size of the human genome. Conversely, the amount of mammalian DNA sequences is exponentially increasing due to ongoing large sequencing projects, but without functional data in most cases. To alleviate the limitations of supervised learning, we propose a paradigm shift with semi-supervised learning, which does not only exploit labeled sequences (e.g. human genome with ChIP-seq experiment), but also unlabeled sequences available in much larger amounts (e.g. from other species without ChIP-seq experiment, such as chimpanzee). Our approach is flexible and can be plugged into any neural architecture including shallow and deep networks, and shows strong predictive performance improvements compared to supervised learning in most cases (up to 70% ). https://forgemia.inra.fr/raphael.mourad/deepgnn .

STUDY DESIGN:Medical vignettes. OBJECTIVES:Lumbar spinal stenosis (LSS) is a degenerative condition with a high prevalence in the elderly population, that is associated with a significant economic burden and often requires spinal surgery. Prior authorization of surgical candidates is required before patients can be covered by a health plan and must be approved by medical directors (MDs), which is often subjective and clinician specific. In this study, we hypothesized that the prediction accuracy of machine learning (ML) methods regarding surgical candidates is comparable to that of a panel of MDs. METHODS:Based on patient demographic factors, previous therapeutic history, symptoms and physical examinations and imaging findings, we propose an ML which computes the probability of spinal surgical recommendations for LSS. The model implements a random forest model trained from medical vignette data reviewed by MDs. Sets of 400 and 100 medical vignettes reviewed by MDs were used for training and testing. RESULTS:The predictive accuracy of the machine learning model was with a root mean square error (RMSE) between model predictions and ground truth of .1123, while the average RMSE between individual MD's recommendations and ground truth was .2661. For binary classification, the AUROC and Cohen's kappa were .959 and .801, while the corresponding average metrics based on individual MD's recommendations were .844 and .564, respectively. CONCLUSIONS:Our results suggest that ML can be used to automate prior authorization approval of surgery for LSS with performance comparable to a panel of MDs.

The DNA damage response is essential to safeguard genome integrity. Although the contribution of chromatin in DNA repair has been investigated 1 , 2 , the contribution of chromosome folding to these processes remains unclear 3 . Here we report that, after the production of double-stranded breaks (DSBs) in mammalian cells, ATM drives the formation of a new chromatin compartment (D compartment) through the clustering of damaged topologically associating domains, decorated with γH2AX and 53BP1. This compartment forms by a mechanism that is consistent with polymer–polymer phase separation rather than liquid–liquid phase separation. The D compartment arises mostly in G1 phase, is independent of cohesin and is enhanced after pharmacological inhibition of DNA-dependent protein kinase (DNA-PK) or R-loop accumulation. Importantly, R-loop-enriched DNA-damage-responsive genes physically localize to the D compartment, and this contributes to their optimal activation, providing a function for DSB clustering in the DNA damage response. However, DSB-induced chromosome reorganization comes at the expense of an increased rate of translocations, also observed in cancer genomes. Overall, we characterize how DSB-induced compartmentalization orchestrates the DNA damage response and highlight the critical impact of chromosome architecture in genomic instability.

Predicting molecular processes using deep learning is a promising approach to provide biological insights for non-coding SNPs identified in genome-wide association studies. However, most deep learning methods rely on supervised learning, which requires DNA sequences associated with functional data, and whose amount is severely limited by the finite size of the human genome. Conversely, the amount of mammalian DNA sequences is growing exponentially due to ongoing large-scale sequencing projects, but in most cases without functional data. To alleviate the limitations of supervised learning, we propose a novel semi-supervised learning based on pseudo-labeling, which allows to exploit unlabeled DNA sequences from numerous genomes during model pre-training. The approach is very flexible and can be used to train any neural architecture including state-of-the-art models, and shows in certain situations strong predictive performance improvements compared to standard supervised learning in most cases. Moreover, small models trained by SSL showed similar or better performance than large language model DNABERT2.### Competing Interest StatementThe authors have declared no competing interest.

Objectives To develop a three-stage convolutional neural network (CNN) approach to segment anatomical structures, classify the presence of lumbar spinal stenosis (LSS) for all 3 stenosis types: central, lateral recess and foraminal and assess its severity on spine MRI and to demonstrate its efficacy as an accurate and consistent diagnostic tool. Methods The three-stage model was trained on 1635 annotated lumbar spine MRI studies consisting of T2-weighted sagittal and axial planes at each vertebral level. Accuracy of the model was evaluated on an external validation set of 150 MRI studies graded on a scale of absent, mild, moderate or severe by a panel of 7 radiologists. The reference standard for all types was determined by majority voting and in case of disagreement, adjudicated by an external radiologist. The radiologists’ diagnoses were then compared to the diagnoses of the model. Results The model showed comparable performance to the radiologist average both in terms of the determination of presence/absence of LSS as well as severity classification, for all 3 stenosis types. In the case of central canal stenosis, the sensitivity, specificity and AUROC of the CNN were (0.971, 0.864, 0.963) for binary (presence/absence) classification compared to the radiologist average of (0.786, 0.899, 0.842). For lateral recess stenosis, the sensitivity, specificity and AUROC of the CNN were (0.853, 0.787, 0.907) compared to the radiologist average of (0.713, 0.898, 805). For foraminal stenosis, the sensitivity, specificity and AUROC of the CNN were (0.942, 0.844, 0.950) compared to the radiologist average of (0.879, 0.877, 0.878). Multi-class severity classifications showed similarly comparable statistics. Conclusions The CNN showed comparable performance to radiologist subspecialists for the detection and classification of LSS. The integration of neural network models in the detection of LSS could bring higher accuracy, efficiency, consistency, and post-hoc interpretability in diagnostic practices.

SUMMARY There is a growing awareness that repeat sequences (RepSeq) - the main constituents of the human genome - are also prime players in its organization. Here we propose that the genome should be envisioned as a supersystem with three main subsystems, each composed of functionally redundant, cooperating elements. We define herein ProA and ProB RepSeqs as sequences that promote either the A/euchromatin or the B/heterochromatin compartment. ProA and ProB RepSeqs shape A/B partitioning, such that the relative proportions of ProA and ProB RepSeqs determine the propensity of a chromosome segment to adopt either an A or a B configuration. In human, core ProA RepSeqs are essentially made of Alu elements, whereas core ProB RepSeqs consist of young L1 and some Endogenous Retroviruses (ERVs) as well as a panel of AT-rich microsatellites and pericentromeric and telomeric satellites. Additionally, RepSeqs with more indefinite character and, importantly, their derivatives known as “transcriptional enhancers”, can shift between ProA and ProB functions and thus act to open or close specific chromatin domains depending on the cellular context. In this framework, genes and their promoters appear as a special class of RepSeqs that, in their active, transcribed state, reinforce the openness of their surroundings. Molecular mechanisms involve cooperativity between ProB elements, presumably underpinned by the condensate-like properties of heterochromatin, which ProA elements oppose in several ways. We provide strong arguments that altered CpG methylation patterns in cancer including a marked loss in the B compartment, result primarily from a global imbalance in the process of CpG methylation and its erasure. Our results suggest that the resulting altered methylation and impaired function of ProB RepSeqs globally weaken the B compartment, rendering it more plastic, which in turn may confer fate plasticity to the cancer cell.

Purpose Lumbar spinal stenosis (LSS) is a condition affecting several hundreds of thousands of adults in the United States each year and is associated with significant economic burden. The current decision-making practice to determine surgical candidacy for LSS is often subjective and clinician specific. In this study, we hypothesize that the performance of artificial intelligence (AI) methods could prove comparable in terms of prediction accuracy to that of a panel of spine experts. Methods We propose a novel hybrid AI model which computes the probability of spinal surgical recommendations for LSS, based on patient demographic factors, clinical symptom manifestations, and MRI findings. The hybrid model combines a random forest model trained from medical vignette data reviewed by surgeons, with an expert Bayesian network model built from peer-reviewed literature and the expert opinions of a multidisciplinary team in spinal surgery, rehabilitation medicine, interventional and diagnostic radiology. Sets of 400 and 100 medical vignettes reviewed by surgeons were used for training and testing. Results The model demonstrated high predictive accuracy, with a root mean square error (RMSE) between model predictions and ground truth of 0.0964, while the average RMSE between individual doctor's recommendations and ground truth was 0.1940. For dichotomous classification, the AUROC and Cohen's kappa were 0.9266 and 0.6298, while the corresponding average metrics based on individual doctor's recommendations were 0.8412 and 0.5659, respectively. Conclusions Our results suggest that AI can be used to automate the evaluation of surgical candidacy for LSS with performance comparable to a multidisciplinary panel of physicians.

Genome-wide association studies have systematically identified thousands of single nucleotide polymorphisms (SNPs) associated with complex genetic diseases. However, the majority of those SNPs were found in non-coding genomic regions, preventing the understanding of the underlying causal mechanism. Predicting molecular processes based on the DNA sequence represents a promising approach to understand the role of those non-coding SNPs. Over the past years, deep learning was successfully applied to regulatory sequence prediction. Such method required DNA sequences associated with functional data for training. However, the human genome has a finite size which strongly limits the amount of DNA sequence with functional data available for training. Conversely, the amount of mammalian DNA sequences is exponentially increasing due to ongoing large sequencing projects, but without functional data in most cases. Here, we propose a semi-supervised learning approach based on graph neural network which allows to borrow information from homologous mammal sequences during training. Our approach can be plugged into any existing deep learning model and showed improvements in many different situations, including classification and regression, and for different types of functional data.

Background/Aim In higher eukaryotes, the three-dimensional (3D) organization of the genome is intimately related to numerous key biological functions including gene expression, DNA repair and DNA replication regulations. Alteration of 3D organization, in particular topologically associating domains (TADs), is detrimental to the organism and can give rise to a broad range of diseases such as cancers. Methods Here, we propose a versatile regression framework which not only identifies TADs in a fast and accurate manner, but also detects differential TAD borders across conditions for which few methods exist, and predicts 3D genome reorganization after chromosomal rearrangement. Moreover, the framework is biologically meaningful, has an intuitive interpretation and is easy to visualize. Result and conclusion The novel regression ranks among top TAD callers. Moreover, it identifies new features of the genome we called TAD facilitators, and that are enriched with specific transcription factors. It also unveils the importance of cell-type specific transcription factors in establishing novel TAD borders during neuronal differentiation. Lastly, it compares favorably with the state-of-the-art method for predicting rearranged 3D genome.

Transcriptionally active loci are particularly prone to breakage and mounting evidence suggests that DNA Double-Strand Breaks arising in active genes are handled by a dedicated repair pathway, Transcription-Coupled DSB Repair (TC-DSBR), that entails R-loop accumulation and dissolution. Here, we uncover a function for the Bloom RecQ DNA helicase (BLM) in TC-DSBR in human cells. BLM is recruited in a transcription dependent-manner at DSBs where it fosters resection, RAD51 binding and accurate Homologous Recombination repair. However, in an R-loop dissolution-deficient background, we find that BLM promotes cell death. We report that upon excessive RNA:DNA hybrid accumulation, DNA synthesis is enhanced at DSBs, in a manner that depends on BLM and POLD3. Altogether our work unveils a role for BLM at DSBs in active chromatin, and highlights the toxic potential of RNA:DNA hybrids that accumulate at transcription-associated DSBs.

The repair of DNA double-strand breaks (DSBs) is essential for safeguarding genome integrity. When a DSB forms, the PI3K-related ATM kinase rapidly triggers the establishment of megabase-sized, chromatin domains decorated with phosphorylated histone H2AX (γH2AX), which act as seeds for the formation of DNA-damage response foci1. It is unclear how these foci are rapidly assembled to establish a 'repair-prone' environment within the nucleus. Topologically associating domains are a key feature of 3D genome organization that compartmentalize transcription and replication, but little is known about their contribution to DNA repair processes2,3. Here we show that topologically associating domains are functional units of the DNA damage response, and are instrumental for the correct establishment of γH2AX-53BP1 chromatin domains in a manner that involves one-sided cohesin-mediated loop extrusion on both sides of the DSB. We propose a model in which H2AX-containing nucleosomes are rapidly phosphorylated as they actively pass by DSB-anchored cohesin. Our work highlights the importance of chromosome conformation in the maintenance of genome integrity and demonstrates the establishment of a chromatin modification by loop extrusion.