WARNING: These profile HMMs are based on antibiotic resistance genes from later releases of the databases than the original annotated_profile_HMMs, but the sequences also include predicted (non-experimentally validated ARGs) and this remains work in progress.
CARD version 3.0.0 protein homolog models and Resfinder version 3.1 sequences were downloaded and merged with CD-HIT to create a non-redundant database of experimentally curated ARGs. Then, these known ARGs were clustered with all the bacterial proteins in orthoDB version 10 (over 19 million proteins) with linclust from the MMSeqs package with 80% sequence identity threshold cut-offs. Clusters with at least 2 known ARGs were then taken forwards and trimmed MSAs, and subsiquently profile HMMs, were constructed as described below but MAFFT was used instead of ClustalOmega for building the MSAs and the profile HMMs were built without adding GA thresholds.
Sets of protein familes represented as profile HMMs annotated for antibiotic resistance (AR) and virulence function.
Each sub-directory contains a catalogue of HMMs and some corresponding .txtfiles that describe the profile HMMs. There is also a .pdf file that shows how the gathering (GA) thresholds were set for each HMM.
The profile HMMs can be scaned with HMMER3 to identify likely AR or virulence protein sequences. It is recommend that the scans are running using the model-specific gathering thresholds, using the --cut_ga flag, for example:
cat hmms/*.hmm > all.hmm
hmmsearch --cut_ga --cpu 1 --tblout output.table all.hmm target.faa
Note: The HMMs only include genes where presence or absence of the gene causes the resistance/virulence phenotype. Therefore, genes are not detected where only specific mutations (SNPs or indels) underly the phenotype as these will require a different approach to detect. The HMMs are recommended for research and exploratory studies and not validated for clinical use.
The initial sequencing databases were card 1.1.3 homolog, ResFinder 2.1, and ResFinderFG 1.0 antibiotic resistance sequences, and VFDB July 2017 and VirulenceFinder 1.5 virulence sequences. Profile HMMs were constructed by first clustering these sequence databases using CD-hit at 40% sequence identity, then aligning sequences with ClustalOmega, trimming multiple sequence alignments with TrimAl, and building HMMs with HMMBuild. Subsequently, all input proteins were scanned against the HMMs with HMMsearch and for each HMM a model-specific bitscore threshold (gathering threshold; GA) was set based on the distribution of bitscores for the sequences that were used to build the original model. Specifically, the threshold was set so that any sequences were be called as homologous to those in the profile HMM if the bitscore hit against the HMM was greater than or equal to the mean - 2 sigma of bitscores from the sequences used to build the HMMs (i.e. the known true positive sequences).