Paperonium

Taste chemistry · Nucleic Acids Res · 2021

The paper rates a sweetness-guessing program 88% accurate, but that’s its “best” round. Ten-round average: 76.5%.

Overall76.5%Its spreadsheet’sten-round average88%Its table (“best” rounds)too highSweet ones74%97%Not sweet79%96%70%8090100
Overall76.5%Its spreadsheet’sten-round average88%Its table(“best” rounds)too highSweet ones74%97%Not sweet79%96%70%8090100

Share of chemicals the program guessed right in its ten-round test

The paper says

A computer program learns which chemicals taste sweet from known examples. It was tested in ten rounds: each time it learned from nine-tenths of the examples and guessed the other tenth. The paper’s table: right on 97% of sweet chemicals, 96% of others, 88% overall.

Its own data show

That can’t be: the overall score mixes the two group scores, so it must land between 96% and 97%. A spreadsheet published with the paper explains why: each figure comes from the round it labels “best” for that score, picked separately, not all from one round. Its ten-round averages: 74% sweet, 79% others, 76.5% overall.

A ten-round test is summed up by its average, the way a grade averages all your exams instead of keeping your best. By that measure the program got 76.5% right, not 88%; its scores swung widely between rounds. Its separate test on 403 new chemicals holds up: 89%.

“The summary” means the paper’s abstract: the short summary printed at its top, which is the part most readers see.

Read the paper →Full technical write-up →

VirtualTaste: a web server for the prediction of organoleptic properties of chemical compounds · Nucleic Acids Res (2021) · doi:10.1093/nar/gkab292