Functional Domain Annotation by Structural Similarity
We aim to annotate the functional domains of T. brucei based on the structural similarity of its proteins to annotated domains. In this regard, we gathered a database of structure of Pfam instances. It was done by cutting the structure of Proteins used as Pfam instance on their Pfam borders. Structure database of Pfam instances can be downloaded from shorturl.at/noXZ0.
We searched proteome of T. brucei and three other organisms against the Pfam instances by foldseek using following commands:
foldseek search {OrgDB} {PfamDB} aln_{Org}_pf tmpFolder -a --max-seqs 10000000 --cov-mode 1 -c 0.8 -e 3
foldseek convertalis {OrgDB} {PfamDB} aln_{Org}_pf aln_{Org}_pf_e3.tsv --format-output \
"query,target,fident,alnlen,mismatch,gapopen,qstart,qend,tstart,tend,qlen,tlen,evalue,bits,alntmscore,lddt"
In the notebooks, we have shown that based on alignment characteristics of query-pfam instance we can train a model to predict if the aligned part of the query has the same pfam as the instance it has aligned to. We have used the trained model to predict new domains in T. brucei.
To reproduce the data, first run DownloadFilesAndMakeDirs.sh to download all the data needed for the next steps. AliLabeler.py and AliLabelerSeq.py label the alignments. TrainingAndBenchmarking.ipynb trains a model and benchmarks the trained model.
Preprint available at: