AutoCoV: tracking the early spread of COVID-19 in terms of the spatial and temporal patterns from embedding space by K-mer based deep learning

Table 2 Performance comparison results (mean ± std)

Method		LHS	MI	F1
Spatial
Dimension Reduction	PCA	0.280 ± 0.040	0.357 ± 0.050	0.739 ± 0.012
	t-SNE	0.215 ± 0.043	0.226 ± 0.034	0.690 ± 0.017
	UMAP	0.237 ± 0.030	0.448 ± 0.031	0.837 ± 0.011
Unsupervised	dna2vec	0.204 ± 0.035	0.108 ± 0.046	0.736 ± 0.019
Unsupervised	seq2vec	0.149 ± 0.041	0.238 ± 0.030	0.689 ± 0.018
Supervised	Seq2Seq+CF	0.131 ± 0.030	0.099 ± 0.013	0.689 ± 0.017
	BERT+CF	0.155 ± 0.054	0.327 ± 0.034	0.734 ± 0.013
	AutoCoV	0.529 ± 0.051	0.773 ± 0.045	0.881 ± 0.012

Method		LHS	MI	F1
Temporal
Dimension Reduction	PCA	0.147 ± 0.065	0.198 ± 0.040	0.665 ± 0.024
	t-SNE	0.140 ± 0.052	0.184 ± 0.033	0.384 ± 0.031
	UMAP	0.146 ± 0.053	0.329 ± 0.047	0.671 ± 0.027
Unsupervised	dna2vec	0.117 ± 0.053	0.087 ± 0.030	0.616 ± 0.021
Unsupervised	seq2vec	0.111 ± 0.048	0.156 ± 0.018	0.464 ± 0.015
Supervised	Seq2Seq+CF	0.116 ± 0.066	0.075 ± 0.013	0.474 ± 0.027
	BERT+CF	0.091 ± 0.058	0.181 ± 0.048	0.559 ± 0.037
	AutoCoV	0.355 ± 0.067	0.554 ± 0.065	0.761 ± 0.023

Three dimensional reduction methods (PCA, t-SNE, UMAP), two unsupervised methods (dna2vec, seq2vec), and three supervised methods (Seq2Seq+CF, BERT+CF, AutoCoV) are compared, and the bold values represent the best performance among them. In both patterns, AutoCoV outperforms the baselines in all metrics

ISSN: 1471-2105