Bickmann, L. | bioRxiv Science

bioRxiv2024

PuMA: PubMed Gene-Celltype-Relation Atlas

Rapid extraction and visualization of cell-specific gene expression is important for automatic celltype annotation, e.g. in single cell analysis. There is an emerging field in which tools such as curated databases or Machine Learning methods are used to support celltype annotation. However, complementing approaches to efficiently incorporate latest knowledge of free-text articles from literature databases, such as PubMed are understudied. This work introduces the PubMed Gene/Celltype-Relation Atlas (PuMA) which provides a local, easy-to-use web-interface to facilitate automatic celltype annotation. It utilizes pretrained large language models in order to extract gene and celltype concepts from Pub-Med and links biomedical ontologies to suggest gene to celltype relations. It includes a search tool for genes and cells, additionally providing an interactive graph visualization for exploring cross-relations. Each result is fully traceable by linking the relevant PubMed articles. The software framework is freely available and enables regular article imports for incremental knowledge updates. GitLab: imigitlab.uni-muenster.de/published/PuMA

bioinformatics↗

bioRxiv2023

TEclass2: Classification of transposable elements using Transformers

MotivationTransposable elements (TEs) are interspersed repetitive sequences that are major constituents of most eukaryotic genomes and are crucial for genome evolution. Despite the existence of multiple tools for their classification and annotation, none of them can achieve completely reliable results making it a challenge for genomic studies. In this work, we introduce TEclass2, a new software that uses a deep learning approach based upon a linear Transformer architecture with a k-mer to-kenizer and further adaptations to handle DNA sequences. This software has an easy configuration that allows training models on new datasets and the classification of TE models providing multiple metrics for a reliable evaluation of the results. ResultsThis work shows a successful adaptation of deep learning with Transformers for the classification of TE models from consensus sequences, and these results lay a foundation for novel methodologies in bioinformatics. We provide a tool for the training of models and the classification of consensus sequences from TE models on custom data and a web page interface with a pre-trained dataset based on curated and non-curated TE libraries allowing a fast and simple classification of TEs. Availabilityhttps://bioinformatics.uni-muenster.de/tools/teclass2/index.pl Contactwojmak@uni-muenster.de Supplementary informationSupplementary data are available at Bioinformatics online.

bioinformatics↗