You have encountered a common problem that occurs when trying to move from a consensus-based search strategy (CD-HIT) to a pairwise based search strategy (BLASTP). In general, consensus based strategies are designed to capture deep evolutionary relationships with a single model. But sometimes, there will be sequences that are closely related (> 50% identity, E() < 1e-40) to each other, but one of the proteins can be detected by the consensus model (but is perhaps distant from its "center"), while the other cannot. (Think of two leaves on a tree on nearby branches, one of which is close enough to the root to be found with CD-HIT, but the other is just beyond detection.) The same problem occurs with PFAM.

One solution would be to use pairwise searches, rather than CD-HIT. Use BLASTP to find the proteins that are shared by the pathogenic organisms but not by non-pathogens (or use ggsearch, which I think will be better suited to this problem).

And forget about 30% identity. There will be many homologous proteins with E()-values < 1e-10 that are clearly homologous but less than 30% identical. E-values are much more reliable indicators of homology than percent identity.

CD-HIT uses heuristics to find clusters of proteins with high similarity. (The name stands for "Cluster Database at High Identity with Tolerance".) So a threshold of 30% is well outside the intended parameter range. At such a low identity threshold the heuristic will miss many pairs that have >30% identity.

Thus, you'll have to rely on the BLAST results. Note that the e-values are dependent on the size of the database, so perhaps instead of an e-value cutoff you want to use a bitscore cutoff. Bitscores have a meaning independent of the number of genes that are in the database.

Hello Nicole,

I want to know the name of databases or list of non-pathogenic bacteria for human? because all the available databases showing only pathogenic bacteria for human.