Enabling Scientific Reproducibility through FAIR Data Management: An ontology-driven deep learning approach in the NeuroBridge Project.

Scientific reproducibility that effectively leverages existing study data is critical to the advancement of research in many disciplines including neuroscience, which uses imaging and electrophysiology modalities as primary endpoints or key dependency in studies. We are developing an integrated search platform called NeuroBridge to enable researchers to search for relevant study datasets that can be used to test a hypothesis or replicate a published finding without having to perform a difficult search from scratch, including contacting individual study authors and locating the site to download the data. In this paper, we describe the development of a metadata ontology based on the World Wide Web Consortium (W3C) PROV specifications to create a corpus of semantically annotated published papers. This annotated corpus was used in a deep learning model to support automated identification of candidate datasets related to neurocognitive assessment of subjects with drug abuse or schizophrenia using neuroimaging. We built on our previous work in the Provenance for Clinical and Health Research (ProvCaRe) project to model metadata information in the NeuroBridge ontology and used this ontology to annotate 51 articles using a Web-based tool called Inception. The Bidirectional Encoder Representations from Transformers (BERT) neural network model, which was trained using the annotated corpus, is used to classify and rank papers relevant to five research hypotheses and the results were evaluated independently by three users for accuracy and recall. Our combined use of the NeuroBridge ontology together with the deep learning model outperforms the existing PubMed Central (PMC) search engine and manifests considerable trainability and transparency compared with typical free-text search. An initial version of the NeuroBridge portal is available at: https://neurobridges.org/.