Herbal medicine AI · PLoS Comput Biol · 2026
The winning F1 score, 0.9885, is impossible: its table allows at most 0.8368, behind a rival.
F1 score (0 to 1) on one test dataset: printed versus the most possible
The paper says
The paper’s AI predicts which herbs might treat which diseases. Its summary claims clear gains over the best rival on three scores, including F1 (0 is worst, 1 perfect). On one of five test datasets, its F1 is 0.9885; the best rival’s, 0.8400.
Its own data show
F1 is calculated from two other numbers in the same table row: the share of the AI’s predicted links that are real (72.17%) and the share of real links it finds (99.55%). Those allow at most 0.8368, below the rival. A second dataset: 0.9241 printed, 0.7625 at most, behind the best rival (0.8819).
F1 averages the two shares but leans toward the lower one: with 72% of picks right, even finding every real link gives only 0.8384. Averaging five test runs can only lower it. None of the 28 rival results is more than 0.0006 above its limit. So a number in the AI’s row is wrong. The paper’s main measure is unaffected.