The following databases were accessed in April 2021. All ampir models
were trained using data from one or more of these databases (see details
below).
Four recently updated antimicrobial peptide (AMP) databases were used:
- APD 3 by Wang et al. 2016
- DRAMP 2.0 by Kang et al. 2019.
- dbAMP by Jhong et al. 2018
- UniProt using the keyword “Antimicrobial [KW-0929]”.
Raw downloads for these databases are included in the data distribution. After unpacking they should be present at the following file locations:
| Database Name | File |
|---|---|
| APD 3 | raw_data/amp_databases/APD_sequence_release_09142020.fasta |
| DRAMP Natural | raw_data/amp_databases/dramp_nat_tidy.fasta |
| dbAMP | raw_data/amp_databases/dbAMPv1.4.xlsx |
| UniProt | raw_data/amp_databases/uniprot-keyword__Antimicrobial+[KW-0929]_.xlsx |
The Antimicrobial Peptide Database appears to be regularly updated. An AMP sequence FASTA file from September 2020 is downloadable from the APD download page which currently contains 3,167 sequences. To include a more updated AMP (October 2020) list, the web query interface was used to obtain the full 3,257 AMPs along with sufficient metadata to filter unwanted entries.
DRAMP’s download page provides access to download a general AMP dataset (which contains both natural and synthetic AMPs). However, there is also a natural dataset. Because the natural dataset contains the AMPs we are interested in and is likely more regularly updated (it currently contains 4414 sequences), the natural AMP data was obtained using a scrape script.
The latest release of dbAMP is from 06/2019 and was downloaded from their download page. It contains 4213 experimentally verified natural AMPs (synthetic AMPs were removed).
note: Unable to access dbAMP database on 7 April 2021 (website unavailable) and was therefore unable an updated version of this database
AMPs were downloaded from UniProt on 09-April-2021 using the search term: “keyword:Antimicrobial [KW-0929]”. This included 3350 reviewed and 42745 unreviewed proteins.
One of the most striking differences between AMP databases becomes clear
simply by looking at the length distributions. The APD and DRAMP
databases emphasise short peptides (mostly < 50 amino acids) which
reflects their focus on mature peptides rather than full length
precursor proteins.
The SwissProt database provides a Peptide field which allows us to
distinguish between entries for mature peptides and precursors. If the
Peptide length is the same as the total length, it is a mature peptide.
For entries where such information is available we see a very clear
split with mature peptides having a typical length of 25 amino acids
(AA) whereas precursors are slightly longer at around 60-75 AA. Note
that the length distribution of mature peptides matches the APD length
distribution very well but all other databases, including dbAMP include
longer sequences and therefore are likely to include some fraction of
precursors. Note that there are a total of 784 mature peptides, 884
precursors with peptide annotation information, and 1682 reviewed AMPs
without Peptide information. The Unknown category has a notably
broader distribution of lengths reflecting the possibility that it
includes a mix of both types.
Another indicator that a protein is an AMP precursor is the presence of a signal peptide. 767 of the 884 precursors identified above have well defined signal peptide sequences. The length distribution plot below shows that only precursors longer than about 60 AA are likely to have a signal peptide. Manual inspection of precursors without signal peptides revealed that many are annotated with a pro-peptide indicating that even in this group there is some post-translational processing to produce a mature product.
SwissProt also includes a small number of larger proteins (>500 AA) that are listed under the keyword Antimicrobial but are very different from classical AMPs. These include some large viral proteins (e.g. EXLYS_BPDPK) which show evidence of antibacterial activity but their mode of action and vast difference in size make them outliers from the point of view of building a machine learning model.
AmPEP Training Data
The AmPEP AMP predictor provides its training data available directly for download. The length distribution of sequences in this database is interesting. It shows that sequences classified as AMPs form a clear peak corresponding to mature peptides whereas non-AMP (background) sequences are clearly larger and more likely to represent full length proteins.
AMP Scanner v2 Data
AMP Scanner data used for training, testing and evaluation are available directly for download from https://www.dveltri.com/ascan/v2/about.html. In contrast with the AmPEP data the lengths of positive and negative datasets are much more closely matched for AMPScanv2. This reflects specific steps taken by the authors of AMP scanner v2 (detailed in Veltri et al. 2018) to extract random sub-sequences from non-AMPs that have a matching length distribution to the AMPs themselves.
Xiao et al. Benchmark data
The benchmark data provided by Xiao et al. 2013 has been used in several studies to provide a somewhat independent estimate of prediction accuracy. Although this data is restricted to sequences less than 100 AA it otherwise resembles the AmPEP data in overall length distribution. This helps explain the extraordinary accuracy of AmPEP when tested with this benchmark. A more worrying issue is that this benchmark, and the AmPEP training data, have length distributions which suggest that positive cases are mature peptides while negative cases are full length proteins. A predictor optimised to perform well on such data will therefore be effective at distinguishing mature peptides from precursor proteins but perhaps not so effective at distinguishing between AMP and non-AMP mature peptides (arguably a more important and interesting task). We would therefore recommend that future work use a negative dataset that is as close as possible to a set of non-AMP mature peptides. Since this is difficult to obtain, the negative data should at least have a similar length distribution to the positive dataset (i.e. reflecting other types of mature peptides).
Since our goal with ampir is to obtain the maximum possible utility
for genome-wide scans, we sought to build a positive AMP dataset
consisting entirely of precursor proteins. In typical genome-scanning
operations this is the only information available.
To achieve this we used the following criteria to design our database:
Firstly we start with the UniProt database and only include proteins if they are present in either the reviewed or unreviewed elements of this database. Although this removes a small number of proteins from custom AMP databases it allows us to make use of extensive metadata included for all proteins in UniProt. The following filters are then applied:
- Exclude all mature peptides
- Exclude unreviewed proteins unless they also appear in APD, DRAMP or dbAMP
- Remove proteins with lengths < 50 AA since these might be mature peptides included in APD, DRAMP or dbAMP
- Remove very large proteins (>500 AA) since these are likely to have very different physicochemical properties and are not amenable to prediction by this method
- Remove identical sequences
- Remove sequences with nonstandard amino acids
This leaves an initial database with 2294 entries, of which 213 are unreviewed.
cd-hit -i raw_data/amp_databases/ampir_positive.fasta -o raw_data/amp_databases/ampir_positive90.fasta -c 0.90 -g 1As a final step we write the database to a FASTA file and then use
cd-hit to cluster sequences to 90% identity, keeping only a single
representative sequence for each cluster. This reduces the database size
(to 1483 sequences) but roughly maintains the same length distribution.
Certain organisms are particularly well annotated for AMPs. We find that our final database contains a large number of Arabidopsis, human, mouse, chicken and rat sequences.
| Organism | nentries | n90_entries |
|---|---|---|
| Arabidopsis thaliana (Mouse-ear cress) | 289 | 281 |
| Mus musculus (Mouse) | 96 | 77 |
| Homo sapiens (Human) | 85 | 59 |
| Rattus norvegicus (Rat) | 65 | 52 |
| Bos taurus (Bovine) | 45 | 34 |
| Gallus gallus (Chicken) | 23 | 19 |
- Full database (prior to
cd-hitclustering) along with SwissProt metadata is available atraw_data/amp_databases/ampir_db.tsv - A FASTA formatted file with 90% clustered sequences
raw_data/amp_databases/ampir_positive90.fasta
Another approach to AMP prediction is to focus entirely on mature peptides as these are the most likely to have shared/convergent physicochemical properties since they are the active molecules.
For this approach we build a database as follows:
- Include all AMPs from the APD, DRAMP and dbAMP databases with lengths >10 AA and < 60 AA
- Include mature peptides from SwissProt (also with length >10 AA and <60 AA)
- Remove sequences that are identical or that contain non-standard amino acids
The resulting database has a high proportion of peptides
Finally cluster these sequences to 90% identity with cd-hit
cd-hit -i raw_data/amp_databases/ampir_mature_positive.fasta -o raw_data/amp_databases/ampir_mature_positive90.fasta -c 0.90 -g 1- Full database (prior to
cd-hitclustering) along with SwissProt metadata is available atraw_data/amp_databases/ampir_mature_positive.fasta - A FASTA formatted file with 90% clustered sequences
raw_data/amp_databases/ampir_mature_positive90.fasta