Biology subjects

Ligeti, B.

Publications and source records attributed to Ligeti, B..

5 recordsLinked to original sources

bioRxiv2024

ProkBERT PhaStyle: Accurate Phage Lifestyle Prediction with Pretrained Genomic Language Models

BackgroundPhage lifestyle prediction, i.e. classifying phage sequences as virulent or temperate, is crucial in biomedical and ecological applications. Phage sequences from metagenome or metavirome assemblies are often fragmented, and the diversity of environmental phages is not well known. Current computational approaches often rely on database comparisons and machine learning algorithms that require significant effort and expertise to update. We propose using genomic language models for phage lifestyle classification, allowing efficient direct analysis from nucleotide sequences without the need for sophisticated preprocessing pipelines or manually curated databases. MethodsWe trained three genomic language models (DNABERT-2, Nucleotide Transformer, and ProkBERT) on datasets of short, fragmented sequences. These models were then compared with dedicated phage lifestyle prediction methods (PhaTYP, DeePhage, BACPHLIP) in terms of accuracy, prediction speed, and generalization capability. ResultsProkBERT PhaStyle consistently outperforms existing models in various scenarios. It generalizes well for out-of-sample data, accurately classifies phages from extreme environments, and also demonstrates high inference speed. Despite having up to 20 times fewer parameters, it proved to be better performing than much larger genomic language models. ConclusionsGenomic language models offer a simple and computationally efficient alternative for solving complex classification tasks, such as phage lifestyle prediction. ProkBERT PhaStyles simplicity, speed, and performance suggest its utility in various ecological and clinical applications.

bioinformatics↗

bioRxiv2023

ProkBERT Family: Genomic Language Models for Microbiome Applications

Machine learning offers transformative capabilities in microbiology and microbiome analysis, deciphering intricate microbial interactions, predicting functionalities, and unveiling novel patterns in vast datasets. This enriches our comprehension of microbial ecosystems and their influence on health and disease. However, the integration of machine learning in these fields contends with issues like the scarcity of labeled datasets, the immense volume and complexity of microbial data, and the subtle interactions within microbial communities. Addressing these challenges, we introduce the ProkBERT model family. Built on transfer learning and self-supervised methodologies, ProkBERT models capitalize on the abundant available data, demonstrating adaptability across diverse scenarios. The models learned representations align with established biological understanding, shedding light on phylogenetic relationships. With the novel Local Context-Aware (LCA) tokenization, the ProkBERT family overcomes the context size limitations of traditional transformer models without sacrificing performance or the information rich local context. In bioinformatics tasks like promoter prediction and phage identification, ProkBERT models excel. For promoter predictions, the best performing model achieved an MCC of 0.74 for E. coli and 0.62 in mixed-species contexts. In phage identification, they all consistently outperformed tools like VirSorter2 and DeepVirFinder, registering an MCC of 0.85. Compact yet powerful, the ProkBERT models are efficient, generalizable, and swift. They cater to both supervised and unsupervised tasks, providing an accessible tool for the community. The models are available on GitHub and HuggingFace.

bioinformatics↗

bioRxiv2022

CovidOutcome2: a tool for SARS-CoV2 mutation identification and for disease severity prediction

Our goal was to develop a platform, CovidOutcome2, capable of predicting disease severity from viral mutation profiles using automated machine learning (autoML) and deep neural networks applied to the available large corpus of sequenced SARS-CoV2 genomes. CovidOutcome2 accepts either user-submitted genomes or user defined mutation combinations as the input. The output is a predicted severity score plus a list of identified, annotated mutations and their functional effects in VCF format. The best model performance is a ROC-AUC 0.899 for the model including patient age and ROC-AUC 0.83 for the model without patient age. AvailabilityCovidOutcome is freely available online under the URL https://www.covidoutcome.bio-ml.com as well as in a standalone version https://github.com/bio-apps/covid-outcome.

bioinformatics↗

bioRxiv2021

COVIDOUTCOME - Estimating COVID Severity Based on Mutation Signatures in the SARS-CoV-2 Genome

IntroductionNumerous studies demonstrate frequent mutations in the genome of SARS-CoV-2. Our goal was to statistically link mutations to severe disease outcome. MethodsWe used an automated machine learning approach where 1,594 viral genomes with available clinical follow-up data were used as the training set (797 "severe" and 797 "mild"). The best algorithm, based on random forest classification combined with the LASSO feature selection algorithm was employed to the training set to link mutation signatures and outcome. The performance of the final model was estimated by repeated, stratified, 10-fold cross validation (CV), then adjusted for multiple testing with Bootstrap Bias Corrected CV. ResultsWe identified 26 protein and UTR mutations significantly linked to severe outcome. The best classification algorithm uses a mutation signature of 22 mutations as well as the patients age as the input and shows high classification efficiency with an AUC of 0.94 (CI: [0.912, 0.962]) and a prediction accuracy of 87% (CI: [0.830, 0.903]). Finally, we established an online platform (https://covidoutcome.com/) which is capable to use a viral sequence and the patients age as the input and provides a percentage estimation of disease severity. DiscussionWe demonstrate a statistical association between mutation signatures of SARS-CoV-2 and severe outcome of COVID-19. The established analysis platform enables a real-time analysis of new viral genomes. KEY MESSAGESO_LIA statistical link between SARS-Cov-2 mutation status and severe COVID outcome was established using automated machine learning techniques based on random forest and logistic regression combined with feature selection algorithms. C_LIO_LIA mutation signature based on 3,779 protein coding and 36 UTR mutations capable to identify severe outcome cases was established. C_LIO_LIThe trained model showed high classification performance (AUC=0.94 (CI: [0.912, 0.962]), accuracy=0.87 (CI: [0.830, 0.903])). C_LIO_LIA registration-free web-server for automated classification of new samples was set up and is accessible at http://www.covidoutcome.com. C_LIO_LIThe established pipeline provides a quick assessment of future patients warranting a prospective clinical validation. C_LI

bioinformatics↗

bioRxiv2018

Multiple polyvalency provided by intrinsically disordered segments is a key feature of postsynaptic scaffold proteins

The human postsynaptic density is an elaborate network comprising thousands of proteins, playing a vital role in the molecular events of learning and the formation of memory. Despite our growing knowledge of specific proteins and their interactions, atomic-level details of their full three-dimensional structure and their rearrangements are mostly elusive. Advancements in structural bioinformatics enabled us to depict the characteristic features of proteins involved in different processes aiding neurotransmission. We show that postsynaptic protein-protein interactions are mediated through the delicate balance of intrinsically disordered regions and folded domains, and this duality is also imprinted in the amino acid sequence. We introduce Diversity of Potential Interactions (DPI), a structure and regulation based descriptor to assess the diversity of interactions. Our approach reveals that the postsynaptic proteome has its own characteristic features and these properties reliably discriminate them from other proteins of the human proteome. Our results suggest that postsynaptic proteins are especially susceptible to forming diverse interactions with each other, which might be key in the reorganization of the PSD in molecular processes related to learning and memory.

bioinformatics↗