Biology subjects

Kulmanov, M.

Publications and source records attributed to Kulmanov, M..

At least 19 recordsLinked to original sources

bioRxiv2026

METEOR: joint genome-scale reconstruction and enzyme prediction

MotivationCurrent machine learning methods for enzyme function prediction primarily treat proteins as independent entities, ignoring the metabolic context in which they operate. This reductionist approach often generates biologically implausible annotations that fail to satisfy stoichiometric or thermodynamic constraints. While genome-scale metabolic models (GEMs) enforce systemic coherence, traditional reconstruction workflows decouple functional annotation from network assembly, resulting in information loss during the discretization of enzymatic evidence. ResultsWe developed METEOR, a framework that integrates continuous confidence scores from deep learning models directly into a Mixed-Integer Linear Programming (MILP) formulation. By conditioning predictions on global metabolic requirements, METEOR resolves functional ambiguities where sequence signals are weak. Evaluation on the Price-149 dataset shows that our method yields physiologically viable networks capable of sustaining growth and improve protein-level predictions over baseline methods. Validation on 6,894 bacterial genomes reveals that system-aware refinement increases the recall of experimentally observed phenotypes substantially. Furthermore, we show that constraint-driven selection can still improve annotation performance in highly incomplete genomes. Our results suggest that functional annotation should be treated as a unified inference problem where global system constraints supervise local predictions. AvailabilityMETEOR is freely available at https://github.com/bio-ontology-research-group/Meteor Contactrobert.hoehndorf@kaust.edu.sa Supplementary informationSupplementary data are available at Bioinformatics online.

bioinformatics↗

bioRxiv2025

Beyond protein functions: evaluating completeness, coherence, and consistency of genome-scale function annotations

Protein function annotation has traditionally followed a reductionist approach, assigning functions to individual proteins acting in isolation. This paradigm treats each annotation as an independent fact, disconnected from the broader biological system. However, proteins operate within integrated cellular networks where their functions depend on genomic context and the presence of interacting partners. Here, we develop a genome-scale evaluation framework that assesses whether annotated protein functions could plausibly coexist within a living organism. We formalize three criteria grounded in systems biology principles: completeness (presence of essential functions), coherence (satisfaction of functional dependencies), and consistency (absence of mutually exclusive functions). Applying this framework to bacterial genomes, we evaluated manually curated annotations from six model organisms and computational predictions from six methods. While model organism annotations largely satisfied our constraints -- with violations primarily reflecting host- pathogen interactions -- all computational prediction methods systematically failed to produce biologically plausible genome-scale annotations. Methods achieved high accuracy for individual proteins yet produced incomplete metabolic pathways, incoherent protein complexes, and taxonomically impossible function combinations. These results reveal a fundamental disconnect between the reductionist annotation model and the systems-level requirements of biological organisms. Current computational methods amplify this disconnect as they are optimized for protein-level accuracy while ignoring genome-scale constraints. Our framework provides quantitative metrics for evaluating biological plausibility and establishes a foundation for developing system-aware annotation approaches. The shift from reductionist to systems-level perspectives will be essential for annotating the rapidly growing collection of sequenced genomes and metagenomes. Significance StatementProtein function prediction methods are evaluated by their accuracy on individual proteins, but proteins operate within integrated biological systems with strict functional requirements. We developed a framework that evaluates whether predicted protein functions could plausibly coexist in a living organism by checking for completeness of essential functions, coherence of functional dependencies, and consistency with biological constraints. While manually curated annotations largely satisfy these requirements, all computational prediction methods systematically fail to produce biologically viable genome-scale annotations. This reveals a fundamental disconnect between current evaluation paradigms and the systems-level requirements of biology, highlighting the need for prediction methods that consider genome-scale constraints rather than optimizing for individual protein accuracy.

bioinformatics↗

bioRxiv2025

Molecular Basis and Cellular Effects of Janus-Class-Driven Cytoplasmic PYK2 Coacervates

Kinase activity is increasingly associated with biomolecular phase separation. Focal adhesion kinase (FAK) forms membrane-associated condensates with paxillin to promote adhesion. Here we show that its paralogue, proline-rich tyrosine kinase 2 (PYK2), undergoes phase separation via a distinct mechanism. PYK2 forms cytoplasmic condensates primarily driven by its kinase- FAT linker (KFL) region. Overexpression of PYK2 induces condensates enriched in its autophosphorylated form, which sequester paxillin from focal adhesions and impair cell adhesion. We uncover an autoregulatory mechanism involving the KFL, linking self-association, autophosphorylation, and condensation. Uncommon among known phase separation drivers, KFL condensation is phosphorylation-independent and its sequence belongs to the "Janus" class. Using a transformer-based protein language model, we identified non-homologous sequences with similar features, many from adhesion and cytoskeletal regulators. We validated the phase-separating potential of several of these sequences in cells. These findings reveal a novel mechanism linking phase separation with kinase activation, and demonstrate distinct condensation behavior in close homologues. Our results also highlight how protein concentration modulates condensate function, with implications for disease, and expand the landscape of phase separation drivers.

cell biology↗

bioRxiv2025

Genomic surveillance of Methicillin Resistant Staphylococcus aureus in Saudi Arabia.

Methicillin-resistant Staphylococcus aureus (MRSA) surveillance in regions with mass gatherings presents unique challenges for public health systems. Saudi Arabia, hosting millions of pilgrims annually, provides a distinctive setting for studying how human mobility shapes bacterial populations, yet comprehensive genomic surveillance data from this region remains limited. Here, we present an integrated analysis of S. aureus isolates collected across seven Saudi Arabian regions, combining whole-genome sequencing with extensive antimicrobial susceptibility testing and standardized metadata following FAIR data principles. Our analysis revealed striking differences between pilgrimage and non-pilgrimage cities. Pilgrimage cities showed significantly higher genetic diversity and antimicrobial resistance rates, harboring numerous international strains including recognized clones from diverse geographic origins. Reported lineage dynamics is changing, expanding toward community clones. While genomic prediction of antimicrobial resistance showed high accuracy for some antibiotics, particularly beta-lactams, with varying performance for others, highlighting the necessity for phenotypic testing in clinical settings. Our findings demonstrate how mass gatherings drive bacterial population structures and emphasize the importance of integrated surveillance approaches in regions with significant global connectivity and travel. ImportanceGenomic data enables the tracking of pathogens by revealing clonal expansions within populations and identifying successful lineages. However, comprehensive national-level data from Saudi Arabia remains limited on a large scale. The adoption of FAIR principles and reproducible workflows ensures robust, consistent analysis, fostering effective data sharing. The OneHealth approachs success depends on the integration and collaboration across diverse domains in todays digital landscape.

genomics↗

bioRxiv2024

Phased genome assemblies and pangenome graphs of human populations of Japan and Saudi Arabia

The selection of a reference sequence in genome analysis is critical, as it serves as the foundation for all downstream analyses. Recently, the pangenome graph has been proposed as a data model that incorporates haplotypes from multiple individuals. Here we present JaSaPaGe, a pangenome graph reference for Saudi Arabian and Japanese populations, both of which have been significantly underrepresented in previous genomic studies. We constructed JaSaPaGe from high-quality phased diploid assemblies which were made utilizing PacBio high-fidelity long reads, Nanopore long reads, and Hi-C short reads of 9 Saudi and 10 Japanese individuals. Quality evaluation of the pangenome graph by variant calling showed that our pangenome outperformed earlier linear reference genomes (GRCh38 and T2T-CHM13) and showed comparable performance to the pangenome graph provided by the Human Pangenome Reference Consortium (HPRC), with more variants found in Japanese and Saudi samples using their population-specific pangenomes. This pangenome reference will serve as a valuable resource for both the research and clinical communities in Japan and Saudi Arabia.

genomics↗

bioRxiv2024

Context-based protein function prediction in bacterial genomes

MotivationThe rapid growth of sequencing data from high-throughput technologies has emphasized the need to uncover the functions of unannotated genes. Recent advancements in deep learning algorithms have enabled researchers to utilize various features to predict protein functions. Traditionally, these algorithms treat proteins as independent functional units or consider interactions only at the protein level. However, prokaryotes often preserve specific genomic neighborhoods over evolutionary time, providing valuable context for predicting protein functions. This context can arise from genes near the gene of interest or synteny regions, where the conserved order of genes on chromosomes results from common ancestry. ResultsWe developed a transformer-based model to pre-train representations of proteins based on their genomic context, and use this model for predicting protein functions. Our results show that context-based protein representations capture context-specific functional semantics and can effectively predict protein functions. We use our model to investigate the influence of phylogenetic distance and homology on the performance of context-dependent function prediction, and find that synteny affects the prediction performance substantially, except for some functions where the function is determined by the genomic context. Our experiments allow us to gain insights into the factors affecting the performance and applicability of context-based function prediction methods across diverse prokaryotic genomes and meta-genomes. Availability and implementationThe generated model, including all training code and generated data, is freely available at https://github.com/bio-ontology-research-group/Genomic_context. Contactrobert.hoehndorf@kaust.edu.sa

bioinformatics↗

bioRxiv2024

Predicting protein functions using positive-unlabeled ranking with ontology-based priors

Automated protein function prediction is a crucial and widely studied problem in bioinformatics. Computationally, protein function is a multilabel classification problem where only positive samples are defined and there is a large number of unlabeled annotations. Most existing methods rely on the assumption that the unlabeled set of protein function annotations are negatives, inducing the false negative issue, where potential positive samples are trained as negatives. We introduce a novel approach named PU-GO, wherein we address function prediction as a positive-unlabeled ranking problem. We apply empirical risk minimization, i.e., we minimize the classification risk of a classifier where class priors are obtained from the Gene Ontology hierarchical structure. We show that our approach is more robust than other state-of-the-art methods on similarity-based and time-based benchmark datasets. Data and code are available at https://github.com/bio-ontology-research-group/PU-GO.

bioinformatics↗

bioRxiv2024

DeepGOMeta: Predicting functions for microbes}

Analyzing microbial samples remains computationally challenging due to their diversity and complexity. The lack of robust de novo protein function prediction methods exacerbates the difficulty in deriving functional insights from these samples. Traditional prediction methods, dependent on homology and sequence similarity, often fail to predict functions for novel proteins and proteins without known homologs. Moreover, most of these methods have been trained on largely eukaryotic data, and have not been evaluated or applied to microbial datasets. This research introduces DeepGOMeta, a deep learning model designed for protein function prediction, as Gene Ontology (GO) terms, trained on a dataset relevant to microbes. The model is validated using novel evaluation strategies and applied to diverse microbial datasets. Data and code are available at https://github.com/bio-ontology-research-group/deepgometa

bioinformatics↗

bioRxiv2023

DeepGO-SE: Protein function prediction as Approximate Semantic Entailment

The Gene Ontology (GO) is one of the most successful ontologies in the biological domain. GO is a formal theory with over 100,000 axioms that describe the molecular functions, biological processes, and cellular locations of proteins in three sub-ontologies. Many methods have been developed to automatically predict protein functions. However, only few of them use the background knowledge provided in the axioms of GO for knowledge-enhanced machine learning, or adjust and evaluate the model for the differences between the sub-ontologies. We have developed DeepGO-SE, a novel method which predicts GO functions from protein sequences using a pretrained large language model combined with a neuro-symbolic model that exploits GO axioms and performs protein function prediction as a form of approximate semantic entailment. We specifically evaluate DeepGO-SE on proteins that have no significant similarity with training proteins and demonstrate that DeepGO-SE can improve function prediction for those proteins.

bioinformatics↗

bioRxiv2022

A personal, reference quality, fully annotated genome from a Saudi individual

We have used multiple sequencing approaches to sequence the genome of a volunteer from Saudi Arabia. We use the resulting data to generate a de novo assembly of the genome, and use different computational approaches to refine the assembly. As a consequence, we provide a contiguous assembly of the complete genome of an individual from Saudi Arabia for all chromosomes except chromosome Y, and label this assembly KSA001. We transferred genome annotations from reference genomes and predicted genome features using methods from Artificial Intelligence to fully annotate KSA001, and we make all primary sequencing data, the assembly, and the genome annotations freely available in public databases using the FAIR data principles.

bioinformatics↗

bioRxiv2022

DeepGOZero: Improving protein function prediction from sequence and zero-shot learning based on ontology axioms

MotivationProtein functions are often described using the Gene Ontology (GO) which is an ontology consisting of over 50,000 classes and a large set of formal axioms. Predicting the functions of proteins is one of the key challenges in computational biology and a variety of machine learning methods have been developed for this purpose. However, these methods usually require significant amount of training data and cannot make predictions for GO classes which have only few or no experimental annotations. ResultsWe developed DeepGOZero, a machine learning model which improves predictions for functions with no or only a small number of annotations. To achieve this goal, we rely on a model-theoretic approach for learning ontology embeddings and combine it with neural networks for protein function prediction. DeepGOZero can exploit formal axioms in the GO to make zero-shot predictions, i.e., predict protein functions even if not a single protein in the training phase was associated with that function. Furthermore, the zero-shot prediction method employed by DeepGOZero is generic and can be applied whenever associations with ontology classes need to be predicted. Availabilityhttp://github.com/bio-ontology-research-group/deepgozero Contactrobert.hoehndorf@kaust.edu.sa

bioinformatics↗

bioRxiv2021

DeepMOCCA: A pan-cancer prognostic model identifies personalized prognostic markers through graph attention and multi-omics data integration

Combining multiple types of genomic, transcriptional, proteomic, and epigenetic datasets has the potential to reveal biological mechanisms across multiple scales, and may lead to more accurate models for clinical decision support. Developing efficient models that can derive clinical outcomes from high-dimensional data remains problematical; challenges include the integration of multiple types of omics data, inclusion of biological background knowledge, and developing machine learning models that are able to deal with this high dimensionality while having only few samples from which to derive a model. We developed DeepMOCCA, a framework for multi-omics cancer analysis. We combine different types of omics data using biological relations between genes, transcripts, and proteins, combine the multi-omics data with background knowledge in the form of protein-protein interaction networks, and use graph convolution neural networks to exploit this combination of multi-omics data and background knowledge. DeepMOCCA predicts survival time for individual patient samples for 33 cancer types and outperforms most existing survival prediction methods. Moreover, DeepMOCCA includes a graph attention mechanism which prioritizes driver genes and prognostic markers in a patient-specific manner; the attention mechanism can be used to identify drivers and prognostic markers within cohorts and individual patients. Author summaryLinking the features of tumors to a prognosis for the patient is a critical part of managing cancer. Many methods have been applied to this problem but we still lack accurate prognostic markers for many cancers. We now have more information than ever before on the state of the cancer genome, the epigenetic changes in tumors, and gene expression at both RNA and protein levels. Here, we address the question of how this data can be used to predict cancer survival and discover which tumor genes make the greatest contribution to the prognosis in individual tumor samples. We have developed a computational model, DeepMOCCA, that uses artificial neural networks underpinned by a large graph constructed from background knowledge concerning the functional interactions between genes and their products. We show that with our method, DeepMOCCA can predict cancer survival time based entirely on features of the tumor at a cellular and molecular level. The method confirms many existing genes that affect survival but for some cancers suggests new genes, either not implicated in survival before or not known to be important in that particular cancer. The ability to predict the important features in individual tumors provided by our method raises the possibility of personalized therapy based on the gene or network dominating the prognosis for that patient.

bioinformatics↗

bioRxiv2020

Machine learning with biomedical ontologies

Ontologies have long been employed in the life sciences to formally represent and reason over domain knowledge, and they are employed in almost every major biological database. Recently, ontologies are increasingly being used to provide background knowledge in similarity-based analysis and machine learning models. The methods employed to combine ontologies and machine learning are still novel and actively being developed. We provide an overview over the methods that use ontologies to compute similarity and incorporate them in machine learning methods; in particular, we outline how semantic similarity measures and ontology embeddings can exploit the background knowledge in biomedical ontologies, and how ontologies can provide constraints that improve machine learning models. The methods and experiments we describe are available as a set of executable notebooks, and we also provide a set of slides and additional resources at https://github.com/bio-ontology-research-group/machine-learning-with-ontologies. Key pointsO_LIOntologies provide background knowledge that can be exploited in machine learning models. C_LIO_LIOntology embeddings are structure-preserving maps from ontologies into vector spaces and provide an important method for utilizing ontologies in machine learning. Embeddings can preserve different structures in ontologies, including their graph structures, syntactic regularities, or their model-theoretic semantics. C_LIO_LIAxioms in ontologies, in particular those involving negation, can be used as constraints in optimization and machine learning to reduce the search space. C_LI

bioinformatics↗

bioRxiv2020

Functional pangenome analysis provides insights into the origin, function and pathways to therapy of SARS-CoV-2 coronavirus

The spread of the novel coronavirus (SARS-CoV-2) has triggered a global emergency, that demands urgent solutions for detection and therapy to prevent escalating health, social and economic impacts. The spike protein (S) of this virus enables binding to the human receptor ACE2, and hence presents a prime target for vaccines preventing viral entry into host cells1. The S proteins from SARS-CoV-1 and SARS-CoV-2 are similar2, but structural differences in the receptor binding domain (RBD) preclude the use of SARS-CoV-1-specific neutralizing antibodies to inhibit SARS-CoV-23. Here we used comparative pangenomic analysis of all sequenced Betacoronaviruses to reveal that, among all core gene clusters present in these viruses, the envelope protein E shows a variant shared by SARS and SARS-Cov2 with two completely-conserved key functional features, an ion-channel and a PDZ-binding Motif (PBM). These features trigger a cytokine storm that activates the inflammasome, leading to increased edema in lungs causing the acute respiratory distress syndrome (ARDS)4-6, the leading cause of death in SARS-CoV-1 and SARS-CoV-2 infection7,8. However, three drugs approved for human use may inhibit SARS-CoV-1 and SARS-CoV-2 Protein E, either acting upon the ion channel (Amantadine and Hexamethylene amiloride9,10) or the PBM (SB2035805), thereby potentially increasing the survival of the host, as already demonstrated for SARS-CoV-1in animal models. Hence, blocking the SARS protein E inhibits development of ARDS in vivo. Given that our results demonstrate that the protein E subcluster for the SARS clade is quasi-identical for the key functional regions of SARS-CoV-1 and SARS-CoV-2, we conclude that use of approved drugs shown to act as SARS E protein inhibitors can help prevent further casualties from COVID-2019 while vaccines and other preventive measures are being developed.

genomics↗

bioRxiv2019

DeepPheno: Predicting single gene knockout phenotypes

MotivationPredicting the phenotypes resulting from molecular perturbations is one of the key challenges in genetics. Both forward and reverse genetic screen are employed to identify the molecular mechanisms underlying phenotypes and disease, and these resulted in a large number of genotype-phenotype association being available for humans and model organisms. Combined with recent advances in machine learning, it may now be possible to predict human phenotypes resulting from particular molecular aberrations. ResultsWe developed DeepPheno, a neural network based hierarchical multi-class multi-label classification method for predicting the phenotypes resulting from complete loss-of-function in single genes. DeepPheno uses the functional annotations with gene products to predict the phenotypes resulting from a loss-of-function; additionally, we employ a two-step procedure in which we predict these functions first and then predict phenotypes. Prediction of phenotypes is ontology-based and we propose a novel ontology-based classifier suitable for very large hierarchical classification tasks. These methods allow us to predict phenotypes associated with any known protein-coding gene. We evaluate our approach using evaluation metrics established by the CAFA challenge and compare with top performing CAFA2 methods as well as several state of the art phenotype prediction approaches, demonstrating the improvement of DeepPheno over state of the art methods. Furthermore, we show that predictions generated by DeepPheno are applicable to predicting gene-disease associations based on comparing phenotypes, and that a large number of new predictions made by DeepPheno interact with a gene that is already associated with the predicted phenotype. Availabilityhttps://github.com/bio-ontology-research-group/deeppheno Contactrobert.hoehndorf@kaust.edu.sa

bioinformatics↗

bioRxiv2019

DeepGOPlus: Improved protein function prediction from sequence

Protein function prediction is one of the major tasks of bioinformatics that can help in wide range of biological problems such as understanding disease mechanisms or finding drug targets. Many methods are available for predicting protein functions from sequence based features, protein-protein interaction networks, protein structure or literature. However, other than sequence, most of the features are difficult to obtain or not available for many proteins thereby limiting their scope. Furthermore, the performance of sequence-based function prediction methods is often lower than methods that incorporate multiple features and predicting protein functions may require a lot of time.\n\nWe developed a novel method for predicting protein functions from sequence alone which combines deep convolutional neural network (CNN) model with sequence similarity based predictions. Our CNN model scans the sequence for motifs which are predictive for protein functions and combines this with functions of similar proteins. We evaluate the performance of DeepGOPlus on the CAFA3 dataset and significantly improve the performance of predictions of biological processes and cellular components with Fmax of 0.47 and 0.70, respectively, using only the amino acid sequence of proteins as input. DeepGOPlus can annotate around 40 protein sequences per second, thereby making fast and accurate function predictions available for a wide range of proteins.

bioinformatics↗

bioRxiv2018

PathoPhenoDB: linking human pathogens to their disease phenotypes in support of infectious disease research

Understanding the relationship between the pathophysiology of infectious disease, the biology of the causative agent and the development of therapeutic and diagnostic approaches is dependent on the synthesis of a wide range of types of information. Provision of a comprehensive and integrated disease phenotype knowledgebase has the potential to provide novel and orthogonal sources of information for the understanding of infectious agent pathogenesis, and support for research on disease mechanisms. We have developed PathoPhenoDB, a database containing pathogen-to-phenotype associations. PathoPhenoDB relies on manual curation of pathogen-disease relations, on ontology-based text mining as well as manual curation to associate phenotypes with infectious disease. Using Semantic Web technologies, PathoPhenoDB also links to knowledge about drug resistance mechanisms and drugs used in the treatment of infectious diseases. PathoPhenoDB is accessible at http://patho.phenomebrowser.net/, and the data is freely available through a public SPARQL endpoint.

bioinformatics↗

bioRxiv2018

Vec2SPARQL: integrating SPARQL queries and knowledge graph embeddings

Recent developments in machine learning have lead to a rise of large number of methods for extracting features from structured data. The features are represented as a vectors and may encode for some semantic aspects of data. They can be used in a machine learning models for different tasks or to compute similarities between the entities of the data. SPARQL is a query language for structured data originally developed for querying Resource Description Framework (RDF) data. It has been in use for over a decade as a standardized NoSQL query language. Many different tools have been developed to enable data sharing with SPARQL. For example, SPARQL endpoints make your data interoperable and available to the world. SPARQL queries can be executed across multiple endpoints. We have developed a Vec2SPARQL, which is a general framework for integrating structured data and their vector space representations. Vec2SPARQL allows jointly querying vector functions such as computing similarities (cosine, correlations) or classifications with machine learning models within a single SPARQL query. We demonstrate applications of our approach for biomedical and clinical use cases. Our source code is freely available at https://github.com/bio-ontology-research-group/vec2sparql and we make a Vec2SPARQL endpoint available at http://sparql.bio2vec.net/.

bioinformatics↗