Improvements in viral gene annotation using large language models and soft alignments

BMC Bioinformatics

Table 2 Confusion matrices based on (a) soft alignment-based approach with a soft alignment score threshold of 20, (b) blastp with an e-value cutoff of 1e−3, and (c) pooled embeddings similarity using KNN search for capsid proteins (CP), regulatory and accessory proteins (RAP), envelope proteins (EP), replication and transcription protein (RTP), and assembly and release proteins (ARP)

	CP	RAP	EP	RTP	ARP	Class sensitivity	Class specificity
(a) Soft alignment (minimum soft alignment score 20)
CP	603	0	0	65	15	0.88	0.98
RAP	1	414	0	42	4	0.9	0.99
EP	4	1	83	2	0	0.92	1
RTP	59	59	2	1758	68	0.9	0.98
ARP	13	21	2	19	3249	0.98	0.97
(b) blastp (e-value cutoff 1e–3)
CP	51	0	0	0	7	0.88	1
RAP	0	187	0	53	2	0.77	0.93
EP	1	0	5	0	0	0.83	1
RTP	6	15	0	843	21	0.95	0.95
ARP	2	7	0	7	974	0.98	0.91
(c) Best match using pooled embeddings
CP	675	9	7	80	133	0.75	0.97
RAP	7	435	3	87	76	0.72	0.97
EP	46	6	92	17	8	0.55	0.97
RTP	112	96	11	1680	108	0.84	0.94
ARP	34	66	48	41	3289	0.95	0.94

ISSN: 1471-2105