Because of the long lag from publication to usage in a clinical study, the Office of Portfolio Analysis set out to develop an approach to identify early signatures of bench-to-bedside translation. Our measure, Approximate Potential to Translate, uses the scientific community’s reaction to an article to estimate the likelihood that the knowledge from that paper will be used and cited in later clinical articles. Approximate Potential to Translate is expressed as a probability from 0.05 (no detectable signatures of translation) to 0.95 (extremely strong signatures of translation), with intermediate values of 0.25, 0.50 and 0.75.
It is generated by training a machine learning model with information describing the Human, Animal, and Molecular/Cellular scores of the article in question as well as those of the later articles that cite the paper. The machine learning system learns, based on whether the training articles were eventually cited by a clinical article, which features serve as early signatures that knowledge from a publication is moving toward clinical applications. It then uses these signatures of bench-to-bedside translation make predictions for each article in iCite.
For example, 25% of the papers that have an Approximate Potential to Translate (APT) of 0.25 are cited by clinical articles on average. Likewise, 75% of papers that have an APT of 0.75 are cited by clinical articles on average. The fact that 25% of papers with an APT score of 0.75 are not cited by a clinical study is not an error, but rather the intended design of this metric.