The results files contain some or all of the following tab-separated
fields:

Field 1:	GenBank ID
Field 2:	Whole protein score
Field 3:	Protein Length
Field 4:	Number of Predicted transmembrane strands
Field 5:	Classification (Gram-, Gram+, or Atypical Gram+)
Field 6:	Organism
Field 7:	Annotation or Homology to Annotation ('---' if none)
Field 8:	4-state per-residue Prediction
Field 9:	Amino Acid Sequence

The four states in the prediction are:

U	'upstrand'
D	'downstrand'
-	'outer loop'
i	'inner loop'

The procedure used to generate these results was the following:

1.  Gather all proteins from 78 Gram-Negative bacterial genomes from
NCBI ftp site ftp://ftp.ncbi.nih.gov/genomes/Bacteria/*/*.faa and in
several cases ftp://ftp.ncbi.nih.gov/genbank/genomes/Bacteria/*/*.faa
and supplementing this with additional SWISS-PROT proteins from those
genomes not appearing in the *.faa files from the NCBI.  (248669
proteins)

2.  For each, generate a psiblast profile with the command:

blastpgp -d <big> -i <sequence> -h 1 -j 2 -b 10000 -o /dev/null -Q <profile>

where <big> is a combination of SWISS-PROT, TREMBL, and the NEW
sequence databases.  Note that we are using an evalue cutoff of 1
(option -h) and the PSSM generated after one search. (-j 2)

3.  Run PROFtmb on these profiles, generating 4-state per-residue
predictions only for the ones scoring above 25 bits. (2369 proteins)

4.  Filter out proteins with more than 20% low-complexity content as
determined by seg, with the command:

seg -x -c 1000000

(both -x and -c are formatting parameters)

Field 2: Whole protein score is calculated as:

bits-(Length*0.066-6).

This linear adjustment was a quick fix for the length-dependence of
the bits score, and was found to be nearly optimal in separating
positives and negatives.  (See the Methods section of this website at
http://cubic.bioc.columbia.edu/services/proftmb/methods.html)


Field 7: To classify whether each PROFtmb hit was previously known to
be either an Integral Outer Membrane or an Outer Membrane protein, or
homologous to one, we created four independent reference sets of
proteins, based on SWISS-PROT files or SWISS-PROT formatted TrEMBL
files, and parsed them using a program called Meta-(A)nnotator), with
specialized configuration files (see Methods section of the PROFtmb
paper).  Then, for each protein PROFtmb hit in files proteomes_*, we
did a blast search using the protein as a query against all annotated
proteins in these four sets.  Proteins coming within an e-value of 1
(or 10) to a member of the set were deemed to be close enough for
annotation transfer by homology.  The set to which each protein is
related is listed in field 7.  Files proteomes_full_* list all PROFtmb
hits, while files proteomes_new_* list only those with '---' in field
7.  In other words, proteins which are not similar to any previously
known beta-barrel.


These predictions are offered for completeness, even though
annotations may already exist directly or through homology transfer.
We may further classify these proteins by existing and homology
transferable annotations of localization to the outer membrane.