Table 2 Comparison of the FART models with previously published work using state-of-the-art binary classifiers as given in refs. ^{9,46,55,56,57,58,59}

	Reference	Model name	Classifier	n	Accuracy	F1	AUROC
Sweet/non-sweet	Tuwani et al.⁵⁵	BitterSweet	AdaBoost	161	0.834	0.856	0.883
	Fritz et al.⁵⁶	VirtualSweet	RF	403	0.893	0.888	0.951
	Bo et al.⁵⁷	SweetMLP-Fingerprint	MLP	444	0.900	–	0.940
	Lee et al.⁵⁸	BoostSweet	Consensus	459	0.899	0.907	0.961
	Yang et al.⁵⁹	ChemSweet	RF	241	0.920	–	0.971
	This work	FART augmented	Transformer	2254	0.926	0.944	0.978
	This work	FART confidence	Transformer	2129	0.938	0.954	0.984
Bitter/non-bitter	Tuwani et al.⁵⁵	BitterSweet	RF	154	0.819	0.838	0.880
	Fritz et al.⁵⁶	VirtualBitter	RF	323	0.898	0.882	0.956
	Charoenkwan et al.⁴⁶	BERT4Bitter	Transformer	128	0.922	–	0.964
	Bo et al.⁵⁷	BitterMLP-Descriptor	MLP	446	0.820	–	0.940
	This work	FART augmented	Transformer	2254	0.958	0.778	0.951
	This work	FART confidence	Transformer	2129	0.970	0.830	0.965
Sour/non-sour	Fritz et al.⁵⁶	VirtualSour	RF	133	0.977	0.842	0.994
	This work	FART augmented	Transformer	2254	0.980	0.906	0.994
	This work	FART confidence	Transformer	2129	0.986	0.935	0.997
Umami/non-umami	This work	FART augmented	Transformer	2254	0.998	0.5	0.989
	This work	FART confidence	Transformer	2129	0.998	0.5	0.989

The best scores in each taste category are highlighted in bold. Despite being trained for multi-class prediction, FART outperforms state-of-the-art methods specifically trained on predicting only sweet/non-sweet, bitter/non-bitter, sour/non-sour and umami/non-umami, respectively. The size of the test set (n) is also reported.
RF Random forest, MLP Multi-layer perceptron.

Quick links

Search