MetaboListem and TABoLiSTM: Two Deep Learning Algorithms for Metabolite Named Entity Recognition

Area of Science:

• Bioinformatics
• Computational Biology
• Metabolomics

Background:

• The rapid growth of metabolomics literature presents challenges for manual literature reviews.
• Efficient text-mining tools are essential for navigating and analyzing this expanding body of research.
• Automated methods are needed to extract key information, such as metabolite names, from scientific publications.

Purpose of the Study:

• To develop and evaluate novel metabolite named entity recognition (NER) methods for the field of metabolomics.
• To create a standardized corpus of full-text metabolomics publications for training and evaluating NER models.
• To improve the efficiency and accuracy of literature reviews in metabolomics through advanced text-mining technologies.

Main Methods:

• Development of two metabolite NER models: MetaboListem (using GloVe embeddings) and TABoLiSTM (using BERT/BioBERT embeddings), both based on Bidirectional Long Short-Term Memory (BiLSTM) networks.
• Creation of a novel, standardized training corpus from over 1000 Open Access metabolomics publications, featuring 105,335 annotated metabolites.
• Evaluation of models on a manually annotated test corpus of 19,138 metabolite annotations.

Main Results:

• MetaboListem achieved an F1-score of 0.890 (precision 0.892, recall 0.888).
• TABoLiSTM (BioBERT version) demonstrated superior performance with an F1-score of 0.909 (precision 0.926, recall 0.893), achieving state-of-the-art results.
• The developed models accurately and efficiently identify metabolite names in text.

Conclusions:

• Deep learning algorithms, particularly BiLSTM networks with advanced transfer learning techniques, are highly effective for metabolite NER.
• The created corpus and NER algorithms provide valuable resources for various metabolomics text-mining applications, including information retrieval and literature-based discovery.
• These tools can significantly enhance the process of reviewing and analyzing metabolomics literature.