GOFlowLLM

GOFlowLLM uses large language models to curate Gene Ontology annotations from open access papers that describe miRNA-mRNA interactions. These interactions cover a small part of the GO graph, for which the GO consortium has well defined curation guidelines. GOFlowLLM follows the flowchart in those guidelines step by step, and reproduces the annotations of expert curators in around 90% of cases. The approach is described in GOFlowLLM: curating miRNA literature with large language models and flowcharts.

Browse all sequences with GOFlowLLM annotations →

Annotations on RNAcentral pages

Sequences curated by GOFlowLLM have a GOFlowLLM tab. For each paper, it shows the decisions the model made on its way through the flowchart, the evidence it took from the paper, and the model’s reasoning. For example, see the results for human hsa-miR-21-5p.

How it works

The flowchart from the GO curation guidelines is encoded as a graph. At each step, the model is asked a yes or no question and given the relevant part of the paper, for example the methods section when looking for an assay. Following the answers leads either to an annotation or to a decision that no annotation can be made. A reasoning model is used, so it can work through the biology before answering.

The model’s output is also constrained, so that it can only:

  • answer yes or no to each flowchart question;

  • quote evidence that appears word for word in the paper;

  • choose genes and proteins from those that Europe PMC has tagged in the paper.

When four expert GO curators checked the results for a set of papers, GOFlowLLM reproduced their curation in around 90% of cases. Where it differed from older annotations, this was often because the guidelines had changed since, so GOFlowLLM can also help bring existing annotations up to date.

Limitations

  • GOFlowLLM follows the curators’ flowchart rigidly, so it is weaker on the parts of the flowchart that curators use least.

  • Gene and protein choices depend on the Europe PMC annotations, which are sometimes wrong.

  • A paper curated by GOFlowLLM is not fully curated. It only extracts what the miRNA guidelines cover, and a human curator would probably find more. These papers are a good starting set for further curation.