Abdullateef I. Almudaifer (University of Alabama at Birmingham), Kuleen Sasse (University of Alabama at Birmingham), Kaiwen He (University of Alabama at Birmingham), Michael J. Patton (University of Alabama at Birmingham), Akhil Nadimpalli (University of Alabama at Birmingham), Clementino Vong Do Rosario (University of Alabama at Birmingham), William Byrd (University of Alabama at Birmingham), John D. Osborne (University of Alabama at Birmingham)

The extraction of treatment-relevant biomedical relations from scientific literature is a critical component for downstream treatment tasks (DTTs) like drug repurposing and precision medicine. However, there is a misalignment between upstream annotators of Natural Language Processing (NLP) corpora who annotate task-invariant “core” biomedical relationships and the needs of DTTs. DTTs rely on pre-computed knowledge graphs (KGs) of relations, derived from relation extraction corpora, such as SemMedDB (Kilicoglu et al. 2012). However, these corpora include relations with weak statistical association, low penetrance, occurrence in irrelevant genetic backgrounds or other limited treatment applicability. We propose the use of multi- task training to flag relations as treatment- relevant, and test our method on the BioRed (Luo et al. 2022) corpus as part of the NIH LitCoin Challenge. We use a majority voting ensemble of BioBERT(Lee et al. 2020) models to jointly predict one of 4 document- level relation types (association, positive_correlation, negative_correlation and no_relation) between biomedical entities in the BioRed corpus and the novel relation modifier that serves as a stand-in for treatment relevance. We placed among the top teams with a F1 score on the testing set of 0.49, the highest individual model had an accuracy of 88.81% for novelty and 87.74% for relation finding.