Automated cognome construction and semi-automated hypothesis generation

Abstract

Modern neuroscientific research stands on the shoulders of countless giants. PubMed alone contains more than 21 million peer-reviewed articles with 40–50,000 more published every month. Understanding the human brain, cognition, and disease will require integrating facts from dozens of scientific fields spread amongst millions of studies locked away in static documents, making any such integration daunting, at best. The future of scientific progress will be aided by bridging the gap between the millions of published research articles and modern databases such as the Allen brain atlas (ABA). To that end, we have analyzed the text of over 3.5 million scientific abstracts to find associations between neuroscientific concepts. From the literature alone, we show that we can blindly and algorithmically extract a “cognome”: relationships between brain structure, function, and disease. We demonstrate the potential of data-mining and cross-platform data-integration with the ABA by introducing two methods for semi-automated hypothesis generation. By analyzing statistical “holes” and discrepancies in the literature we can find understudied or overlooked research paths. That is, we have added a layer of semi-automation to a part of the scientific process itself. This is an important step toward fundamentally incorporating data-mining algorithms into the scientific method in a manner that is generalizable to any scientific or medical field.

Introduction

The scientific method begins with a hypothesis about our reality that can be tested via experimental observation. Hypothesis formation is iterative, building off prior scientific knowledge. Before one can form a hypothesis, one must have a thorough understanding of previous research to ensure that the path of inquiry is founded upon a stable base of established facts. But how can a researcher perform a thorough, unbiased literature review when over one million scientific articles are published annually (Björk et al., 2008)? The rate of scientific discovery has outpaced our ability to integrate knowledge in an unbiased, principled fashion. One solution may be via automated information aggregation (Akil et al., 2001). In this manuscript we show that, by calculating associations between concepts in the peer-reviewed literature, we can algorithmically synthesize scientific information and use that knowledge to help formulate plausible low-level hypotheses.

Neuroscience is a particularly complex discipline that relies upon expertise from many disparate fields (Akil et al., 2001). The aim of neuroscience is to understand relationships between brain, behavior, and disease; yet, no one person or group can possibly unify all neuroscientific understanding into a coherent framework. In this paper, we show that the literature contains a hidden network of connected facts that, by definition, recapitulate known neuroscientific relationships. Neuroanatomical, behavioral, and disease associations can be quantified and visualized to speed research and education or to discover understudied research paths (Yarkoni et al., 2010, Wren et al., 2004, Bilder et al., 2009). Rather than allowing our limited ability to review the entire scientific literature bias our hypotheses, we can algorithmically integrate millions of scientific research papers in a principled fashion.

To accomplish this, we used a co-occurrence algorithm to calculate the pair-wise association index (AI) between neuroscientific terms (and their synonyms) contained within more than 3.5 million papers indexed in PubMed (see Section 2). The primary assumption is that the frequency with which terms appeared together across the titles or abstracts of manuscripts is proportional to their probability of association. That is, we assumed an underlying structure within the peer-reviewed neuroscientific literature that we could leverage to our advantage. We conceive of our system as a proof-of-concept tool for knowledge discovery limited only by the size and quality of the inputs. We believe that, in its current state, when combined with the website search and visualization system we created to accompany it (http://www.brainscanr.com), it acts as a more sophisticated complement to normal PubMed searches. Furthermore, it provides, for the first time, a method for quantifying the relationship between disparate neuroscientific concepts, paving the way for researchers to incorporate statistical decision making into their future research.