Curation at Scale with EPITOME: Extraction Pipeline for Immunological Texts and Open-Source Multimodal Enquiry.

Area of Science:

• Bioinformatics
• Computational Biology
• Immunology

Background:

• The Immune Epitope Database (IEDB) manually curates epitope data from scientific literature.
• Traditional curation methods struggle to keep pace with the increasing volume of published research.
• Scientific papers contain multimodal data (text, tables, figures) that are challenging to extract.

Purpose of the Study:

• To develop an open-source tool to assist human curators in automating biological data extraction.
• To leverage vision language models (VLMs) for enhanced biocuration.
• To create a multimodal document processing pipeline for scientific literature.

Main Methods:

• Developed EPITOME, a pipeline combining Optical Character Recognition (OCR), text matching, and VLM capabilities.
• Implemented a three-stage processing system: regex-based identification, visual element extraction, and contextual indexing.
• Linked peptide sequences, MHC molecules, and assays to their locations within documents for VLM Question-Answering (QnA).

Main Results:

• EPITOME demonstrated promising zero-shot performance using open-source VLMs.
• The system effectively extracts and contextualizes biological data from scientific documents.
• Evaluation identified key points for curator intervention to improve accuracy.

Conclusions:

• EPITOME shows potential for accelerating biocuration through a curator-in-the-loop approach.
• VLMs can significantly enhance the efficiency of biological data extraction from multimodal sources.
• Automated tools are crucial for managing the growing scale of scientific literature.