Agile in-litero experiments : how can semi-automated information extraction from neuroscientific literature help neuroscience model building?

In neuroscience, as in many other scientific domains, the primary form of knowledge dissemination is through published articles in peer-reviewed journals. One challenge for modern neuroinformatics is to design methods to make the knowledge from the tremendous backlog of publications accessible for search, analysis and its integration into computational models. In this thesis, we introduce novel natural language processing (NLP) models and systems to mine the neuroscientific literature. In addition to in vivo, in vitro or in silico experiments, we coin the NLP methods developed in this thesis as in litero experiments, aiming at analyzing and making accessible the extended body of neuroscientific literature. In particular, we focus on two important neuroscientific entities: brain regions and neural cells. An integrated NLP model is designed to automatically extract brain region connectivity statements from very large corpora. This system is applied to a large corpus of 25M PubMed abstracts and 600K full-text articles. Central to this system is the creation of a searchable database of brain region connectivity statements, allowing neuroscientists to gain an overview of all brain regions connected to a given region of interest. More importantly, the database enables researcher to provide feedback on connectivity results and links back to the original article sentence to provide the relevant context. The database is evaluated by neuroanatomists on real connectomics tasks (targets of Nucleus Accumbens) and results in significant effort reduction in comparison to previous manual methods (from 1 week to 2h). Subsequently, we introduce neuroNER to identify, normalize and compare instances of identify neuronsneurons in the scientific literature. Our method relies on identifying and analyzing each of the domain features used to annotate a specific neuron mention, like the morphological term 'basket' or brain region 'hippocampus'. We apply our method to the same corpus of 25M PubMed abstracts and 600K full-text articles and find over 500K unique neuron type mentions. To demonstrate the utility of our approach, we also apply our method towards cross-comparing the NeuroLex and Human Brain Project (HBP) cell type ontologies. By decoupling a neuron mention's identity into its specific compositional features, our method can successfully identify specific neuron types even if they are not explicitly listed within a predefined neuron type lexicon, thus greatly facilitating cross-laboratory studies. In order to build such large databases, several tools and infrastructureslarge-scale NLP were developed: a robust pipeline to preprocess full-text PDF articles, as well as bluima, an NLP processing pipeline specialized on neuroscience to perform text-mining at PubMed scale. During the development of those two NLP systems, we acknowledged the need for novel NLP approaches to rapidly develop custom text mining solutions. This led to the formalization of the agile text mini