To start with, I would completely forget about using CD-HIT for this purpose. The point of CD-HIT is that it is a very fast algorithm for finding highly similar sequences. 30% similarity is not high similarity. Moreover bacterial genomes are not so big that speed is your primary concern here.

The exact way to do it depends a bit on whether you are looking for global or local similarity scores. Assuming that you are interested in local alignments, I would use BLAST and filter the results by whatever combination of identity, similarity, alignment length, bit score, or E-value cutoff you desire.

I would try jackhmmer from the HMMER 3.0 package. Something like:

~/hmmer3.0/jackhmmer proteome1.fasta proteome2.fasta

the cd-hit package contains a perl script called PSI-CD-HIT for low ID cutoffs. this script uses BLAST as part of the clustering process to calculate similarities. it is not as fast as regular CD-HIT, mind. http://weizhong-lab.ucsd.edu/cd-hit/wiki/doku.php?id=cd-hit_user_guide#psi-cd-hit_clustering

however, I'm not sure this is the right tool for you. an All-VS-All BLAST of the two species can easily be done on most desktops and would probably be better.

I would blast the first against the second proteome at some low e-value cutoff (< e-3). Then you could either take the similarity values from blast (local alignment), or run a global alignment on significant hits to get global alignment similarities.

Thanks everyone for the comments. Actually I want to identify the proteins (probably the virulence factors) which exist only in those pathogens and not in the non-pathogens. I wonder I can compare the proteomes of the pathogens and purged out the orthologs at 25% or 30%, and at the end I can get those proteins which exist in all the pathogens that I analyzed. Anyone can give me a better suggestion on how to do this? Thanks in advance.