Deep forecasting of translational impact in medical research

• Published in: Patterns (New York, N.Y.) N.Y.), 2022-05, Vol.3 (5), p.100483-100483, Article 100483
• Main Authors: Nelson, Amy P.K., Gray, Robert J., Ruffle, James K., Watkins, Henry C., Herron, Daniel, Sorros, Nick, Mikhailov, Danil, Cardoso, M. Jorge, Ourselin, Sebastien, McNally, Nick, Williams, Bryan, Rees, Geraint E., Nachev, Parashkev
• Format: Article
• Language: English
• Subjects: deep learning; DSML3: Development/Pre-production: Data science output has been rolled out/validated across multiple domains/problems; natural language processing; representation learning; research impact; translational research
• ISSN: 2666-3899; 2666-3899
• DOI: 10.1016/j.patter.2022.100483

The value of biomedical research—a $1.7 trillion annual investment—is ultimately determined by its downstream, real-world impact, whose predictability from simple citation metrics remains unquantified. Here we sought to determine the comparative predictability of future real-world translation—as indexed by inclusion in patents, guidelines, or policy documents—from complex models of title/abstract-level content versus citations and metadata alone. We quantify predictive performance out of sample, ahead of time, across major domains, using the entire corpus of biomedical research captured by Microsoft Academic Graph from 1990–2019, encompassing 43.3 million papers. We show that citations are only moderately predictive of translational impact. In contrast, high-dimensional models of titles, abstracts, and metadata exhibit high fidelity (area under the receiver operating curve [AUROC] > 0.9), generalize across time and domain, and transfer to recognizing papers of Nobel laureates. We argue that content-based impact models are superior to conventional, citation-based measures and sustain a stronger evidence-based claim to the objective measurement of translational potential. •Deep learning models of biomedical paper content can accurately predict translation•Deep content models substantially outperform traditional citation metrics•Models trained on patent inclusion transfer to predicting Nobel Prize-preceding papers•Science policy is potentially better informed by deep content models than by citations The relationship of scientific activity to real-world impact is hard to describe and even harder to quantify. Analyzing 43.3 million biomedical papers from 1990–2019, we show that deep learning models of publication, title, and abstract content can predict inclusion of a scientific paper in a patent, guideline, or policy document. We show that the best of these models, incorporating the richest information, substantially outperforms traditional metrics of paper success—citations per year—and transfers to the task of predicting Nobel Prize-preceding papers. If judgments of the translational potential of science are to be based on objective metrics, then complex models of paper content should be preferred over citations. Our approach is naturally extensible to richer scientific content and diverse measures of impact. Its wider application could maximize the real-world benefits of scientific activity in the biomedical realm and beyond. Analyzing 43.3 million biomedic