Table 3.

AUC and total runtime for each method (listed from the least to most accurate) on pfam-local benchmark with 15,273 pairs of proteins

	UBLAST	USEARCH	FASTA	phmmer	BLAST	CS-BLAST	HHsearch	DCTdomain^a
AUC	0.840	0.906	0.906	0.924	0.951	0.952	0.971	0.972
Time	237 sec	156 sec	749 sec	993 sec	468 sec	50 min	5.7 h	6.6 sec/47 min

^aOur program reports both DCTglobal scores and DCTdomain scores and the reported time is for computing both; using DCTglobal scores resulted in very low accuracy with AUC of 0.665 only; see Figure 5. PROST is not included in this table as its AUC is also very low. Given the DCT fingerprints, DCTdomain is extremely fast using only a few seconds for comparing all the pairs; it is still relatively fast (47 min) even including the DCT fingerprint generation (ESM-2 embedding and RecCut) for all of the proteins in those pairs (13,407 proteins). All programs were run on the same Linux computer using one CPU and GPU (embedding was done using GPU).

Protein domain embeddings for fast and accurate similarity search