Authors: Senay Kafkas, Robert Hoehndorf
Affiliation: King Abdullah University of Science and Technology Thuwal, Saudi Arabia
Project Description:
Studying gene functions is essential as it reveals fundamental biological processes governing life. Genes contain instructions for building proteins, which are crucial for cellular functions. This understanding aids scientists in disease mechanism exploration, identifying drug targets, and developing personalized therapies. This knowledge advances medical research, and personalized medicine, and enhances human health.
The Gene Ontology (GO) captures gene products' functions in classes and establishes a relationship between them. Annotating proteins with Gene Ontology (GO) functions by hand from biomedical literature is a laborious task that necessitates automation. Earlier approaches primarily depended on unsupervised (dictionary-based [1,2], text similarity [3]) and supervised machine learning-based [4] methods to extract gene functions from the literature.
However, none of the existing studies explored the potential of LLMs in gene function extraction from scientific articles. To unveil the potential of LLMs in gene function extraction, we will use several recent cutting-edge LLMs (e.g. GPT3.5-turbo, GPT4) and design zero-shot, few-shot, and fine-tuning (will be explored by using weakly supervision approach only if zero-shot/one-shot does not work) experiments. We will compare the performance of the LLMs against baseline/other methods (e.g. dictionary-based approach) on the BC4GO dataset [5]. This dataset comprises 200 full-text articles curated for genes, corresponding evidence text, and the functions of these genes linked to Gene Ontology (GO) terms. To create a manageable test set for BLAH, we randomly chose 25 full-text articles annotated with 663 triples, each consisting of gene, evidence text, and GO term.
During BLAH, we will experiment with different prompts. The input to the LLM will be a piece of text (depending on the max token size allowed by the LLM) containing the protein name of interest and the output will be a list of functions of this protein. Functions can be returned in free text form (entity recognition) along with their GO IDs (Normalization). Depending on the performance, we may implement a normalization method (which could be based on text similarity). The scalability of the LLM-based approaches to a larger dataset/whole literature depends on the budget of the one who likes to apply.
References:
[1] Sumyyah Toonsi. Automatic Protein Function Annotation Through Text Mining”. MS Thesis (2019).
[2] Jia-Hong Wang et al. GenCLiP 3: mining human genes’ functions and regulatory networks from PubMed based on co-occurrences and natural language processing, Bioinformatics, Vol. 36, Issue 6, pp. 1973–1975, (2019), https://doi.org/10.1093/bioinformatics/btz807
[3] Couto FM, et. al. GOAnnotator: linking protein GO annotations to evidence text. J Biomed Discov Collab. 1:19. (2006), doi: 10.1186/1747-5333-1-19.
[4] Andrea Fossati et al. Towards a systematic characterization of protein complex function: a natural language processing and machine-learning framework. Pre-print: bioRxiv, (2021), doi: https://doi.org/10.1101/2021.02.24.432789
[5] Van Auken K, et. al. BC4GO: a full-text corpus for the BioCreative IV GO task. Database (Oxford). 2014 Jul 28;2014:bau074. doi: 10.1093/database/bau074.