Home » Abstracts » AI and Bioinformatics » Bridging expert curation and LLMs for automated metadata extraction in lesSDRF 2.0

Bridging expert curation and LLMs for automated metadata extraction in lesSDRF 2.0

AI and Bioinformatics Other

Abstract #230 Posted on28 March 202529 April 2025 29 April 2025 22h54

Author(s)

Armin Soleymaniniya | Computational Mass Spectrometry Department, School of Life Sciences, Technical University of Munich | Maximus-von-Imhof-Forum 3, 85354, Freising, Germany
Harikrishnan Ramadasan (Presenting Author) | CompOmics - VIB Ghent University Center for Medical Biotechnology | FSVM II - Technologiepark Zwijnaarde 75, 9052, Gent, Belgium
Julian Uszkoreit | Ruhr University Bochum, Medical Faculty, Medical Bioinformatics | Universitätsstr. 150, D-44801, Bochum, Germany
Katerina Nastou | Statens Serum Institut | Artillerivej 5, 2300, Copenhagen, Denmark
Lev Levitsky | Protein Research Group, Department of Biochemistry and Molecular Biology, University of Southern Denmark | Campusvej , 5230, Odense, Denmark
Magnus Palmblad | Center for Proteomics and Metabolomics | Postbus 9600, 2300 RC, Leiden, Netherlands
Natalia Tischenko | CompOmics - VIB - Ghent University Center for Medical Biotechnology | FSVM II - Technologiepark Zwijnaarde 75, 9052, Ghent, Belgium
Samuel de la Cámara Fuentes | Proteomics Unit, Faculty of Pharmacy, Complutense University of Madrid | Plaza Ramón y Cajal, 28040, Madrid, Spain
Tim Van Den Bossche | CompOmics - VIB Ghent University Center for Medical Biotechnology | FSVM II - Technologiepark Zwijnaarde 75,, 9052, Ghent, Belgium
Veit Schwämmle | University of Southern Denmark | Campusvej 55, 5230, Odense, Denmark
Tine Claeys | CompOmics - VIB Ghent University Center for Medical Biotechnology | FSVM II - Technologiepark Zwijnaarde 75, 9052, Ghent, Belgium

Abstract

Originating from the 2025 EuBIC-MS Developers Meeting, our initiative introduces an automated metadata annotation pipeline integrating natural language processing (NLP) and large language models (LLMs) with robust manual curation. Implemented in lesSDRF 2.0, this pipeline extracts biological and technical metadata from open-access publications, supplementary files, and mzML files in accordance with SDRF guidelines. To ensure accuracy, we employed a hybrid approach: 42 open-access papers were manually annotated following SDRF guidelines to fine-tune a Generative Pretrained Transformer (GPT) model. To mitigate hallucinations, the model is reinforced with a Named Entity Recognition (NER) pipeline incorporating 18 SDRF-compatible ontologies. Once refined, the GPT model will be used to annotate a vast corpus of open-access papers, which, after expert validation, will serve as training data for a BERT-based LLM, selected for its state-of-the-art performance in NER tasks. In parallel, we are developing workflows to extract metadata from supplementary files and MS data formats, ensuring coverage of both biological context and technical parameters. This dual strategy consolidates experimental and technical details into a streamlined and FAIR-compliant framework. By combining careful manual curation with advanced NLP models, this initiative bridges human expertise and computational efficiency, marking a significant step forward in automated proteomics metadata annotation.

Abstract N°

Review committee

Sélectionnez le(s) utilisateur(s) chargé(s) de la revue de cet abstract.

Communication type

Speaker abstract
Oral presentation
Flash communication
Poster

Communication status

Accepted
Rejected

Final abstract N°

Note globale

Abstract content *

Max 2000 characters. Line breaks are supported.

Originating from the 2025 EuBIC-MS Developers Meeting, our initiative introduces an automated metadata annotation pipeline integrating natural language processing (NLP) and large language models (LLMs) with robust manual curation. Implemented in lesSDRF 2.0, this pipeline extracts biological and technical metadata from open-access publications, supplementary files, and mzML files in accordance with SDRF guidelines. To ensure accuracy, we employed a hybrid approach: 42 open-access papers were manually annotated following SDRF guidelines to fine-tune a Generative Pretrained Transformer (GPT) model. To mitigate hallucinations, the model is reinforced with a Named Entity Recognition (NER) pipeline incorporating 18 SDRF-compatible ontologies. Once refined, the GPT model will be used to annotate a vast corpus of open-access papers, which, after expert validation, will serve as training data for a BERT-based LLM, selected for its state-of-the-art performance in NER tasks. In parallel, we are developing workflows to extract metadata from supplementary files and MS data formats, ensuring  coverage of both biological context and technical parameters. This dual strategy consolidates experimental and technical details into a streamlined and FAIR-compliant framework. By combining careful manual curation with advanced NLP models, this initiative bridges human expertise and computational efficiency, marking a significant step forward in automated proteomics metadata annotation.

Author(s) informations *

1	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
2	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
3	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
4	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
5	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
6	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
7	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
8	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
9	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
10	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
11	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *
1	First name * Last name * email * Presenting author * At least and only one presenting author must be selected per abstract Presenting author Do you want this email to appear in the book of abstracts? If you refuse, only the administrators and the review commitee will have access to this email as part of the review process. YesNo Laboratory * Address * ZIP code * City * Country *

Add another author

Are you an ECR? *

<7 years after last degree

Are you a Junior PI? *

If you have started your own group less than 3 years ago and are not chosen by the scientific committee for a talk, you will have the opportunity to give a keynote during an ECR parallel session

Are you an invited speaker? *

Consent *

By submitting this form I agree
- that the administrators use my personal information for the abstract revision process.
- that my name, affiliation and abstract content may be publicly displayed.

Captcha