Can predicting scientists improve discovery forecasts?

The excerpt's central claim is that a prediction model for science improves when it models the scientists who make discoveries, not only what their papers say. The authors build a mixed hypergraph of materials, properties and researchers, then sample random walks over it. Their headline result: "incorporating the distribution of human expertise by training unsupervised models on simulated inferences cognitively accessible to experts dramatically improves (up to 400%) AI prediction of future discoveries beyond those focused on research content alone, especially when relevant literature is sparse." In the drug repurposing test, the human-aware model gives "43% higher precision than identical models accounting for research content alone," measured 18 years after the 2001 prediction year.

The mechanism is a simulation of human inference. Each walk step stands for something a scientist could plausibly do: an author-author step for conversation between collaborators, an author-material step for deep familiarity with a material, and a material-material step for what a reader of two literatures could realize. The authors "hypothesize that identifying topics with high human expert density around them provides us with an informative signal regarding near-future discoveries." The excerpt verifies outcomes, not reasoning. Predictions are scored by overlap with discoveries published after the prediction year, and drug predictions are checked against expert-curated associations in the Comparative Toxicogenomics Database. "These models succeed by predicting human predictions and the scientists who will make them" is the authors' explanation. It is backed by a correlation between precision and how often a drug is mentioned in the literature (r=0.74, p<0.001), not by any checked account of a scientist's reasoning.

Against the nearest notes, this is a different kind of graph. The Mechanist note Can AI automate the discovery of how AI models work? grounds an agent in a knowledge graph of studies. This excerpt argues that content-side graphs cannot see the people positioned to make an inference. HypoEvolve How do collaboration rules shape hypothesis quality? also makes collaboration shape hypothesis quality, but its collaborators are explicit agents, where here they are inferred from co-authorship. The Large Discovery Model note Can language models reliably judge their own candidate quality? ranks candidates by a surrogate's value estimate; this paper ranks them by expert density. The sharpest turn runs the other way: tuned to avoid the crowd, the model yields "scientifically promising 'alien' hypotheses unlikely to be imagined or pursued without intervention until the distant future."

The excerpt does not establish that those alien hypotheses are good. It calls them promising and unlikely to arise otherwise, but reports no evaluation of them, and the COVID-19 progesterone case is an illustration, not a test. The materials-prediction section has only its heading in the excerpt, so the only quantified result in the body text is the 43% drug figure, from one prediction year; the 400% headline appears in the abstract and conclusion without its materials case. Co-authorship is the only human tie modeled, and the Limitations section says other ties, such as shared institutions and conference attendance, could be considered. The implication is narrow: in this drug and materials setting, with dense co-authorship and sparse literature, expert distribution adds predictive signal beyond content. Whether that generalizes, and whether the crowd-avoiding hypotheses are worth pursuing, the excerpt leaves open.