Table 1.

F1 score (harmonic mean of precision and recall).

Method	Dataset	F1-metal	F1-XNA	F1-small	F1-all
bindEmbed21DL	DevSet1014	24 ± 2%	18 ± 3%	26 ± 2%	39 ± 2%
bindEmbed21DL	TestSet300	22 ± 4%	24 ± 6%	33 ± 3%	43 ± 2%
bindEmbed21DL	TestSetNew46	26 ± 14%	19 ± 11%	29 ± 9%	37 ± 6%
Random	TestSet300	1 ± 1%	6 ± 2%	6 ± 1%	9 ± 1%
bindEmbed21DL	TestSet225	n/a	n/a	n/a	47 ± 2%
bindPredictML17	TestSet225	n/a	n/a	n/a	34 ± 2%
bindEmbed21DL	TestSet300XNA66	n/a	31 ± 5%	n/a	n/a
ProNA2020	TestSet300XNA66	n/a	33 ± 7%	n/a	n/a
bindEmbed21DL	TestSet300Zinc51	58 ± 8%	n/a	n/a	n/a
PredZinc	TestSet300Zinc51	58 ± 10%	n/a	n/a	n/a
ZincBindPredict	TestSet300Zinc51	17 ± 9%	n/a	n/a	n/a

Measure: F1 (Eq. 3); ± : 95% confidence intervals (1.96 standard errors); Methods: bindEMbed21DL: method introduced here, bindPredictML17⁵: MSA-based method predicting binding, ProNA2020¹⁹: method specialized on predicting binding to DNA, RNA, and other proteins; PredZinc²⁰ and ZincBindPredict²⁹: methods specialized on predicting zinc-binding; Random: random prediction by randomly shuffling the original output probabilities of bindEmbed21DL; Data: DevSet1014: development set (validation) set with 1014 proteins, TestSet300: Test set created during method development with 300 proteins, TestSet225: subset of test set shared with bindPredictML17, TestSetNew46: 46 sequence-unique proteins added since development of this work began—all sequence-unique with respect to each other and all other proteins used, TestSet300_XNA66: subset with DNA or RNA (dubbed XNA) binding proteins from our test set. TestSet300_Zinc51: subset with zinc-binding proteins from our test set.