Tuning Large language models to read biological literature

点击翻译按钮获取中文摘要

In this application, we focus on two related bioinformatics challenges that require interpretation and knowledge extraction from biological and biomedical literature at great scale.First, gene/genome databases store information on gene function, which is ultimately derived from scientific experiments with results reported in publications. It is exceptionally time-consuming and expensive for human curators to read all relevant scientific literature, interpret what has reported about the function or localisation of gene products, and assign specific controlled vocabulary terms (e.g. Gene Ontology terms) or short free text descriptions (gene names or product descriptions.Second, there are enormous volumes of raw data sets accompanying scientific publications, which are deposited in archival databases from expensive omics experiments, including mass spectrometry (MS) proteomics. Our group and others develop and apply pipelines for re-analysing MS data for new purposes, including annotating genomes, discovery of post-translational modifications and building quantitative atlases of species or tissues amongst others. There is a major bottleneck interpreting the original experimental design, sample descriptions and software parameters, which are currently described in blocks of free text submitted to the archival repository or within Materials and Methods sections of accompanying articles. For both challenges, we believe that with the recent extraordinary improvements in large language models (LLMs), they can be retrained and harnessed for these tasks, to remove the bottleneck in knowledge extraction from literature. Our group has significant expertise in bioinformatics and machine learning, but limited expertise in natural language processing (NLP) to date. In this international partnering application, we are collaborating with a leading group in artificial intelligence and NLP from the University of Pennsylvania (UPenn). The UPenn team will help to guide us in the optimal approach for re-training open source LLMs, using training data that our team has amassed over many years. We will produce open source code for the two challenge areas, with a longer term plan to put these into production within the context of major international databases and consortia, within which we have leading roles.