Hi fellow researchers, I have got a noob query regarding the question of the post. I am trying to build a data set of enzymes (from UniProt data base) that would contain protein sequences and their corresponding parent enzyme class (say hydrolase or kinase or similar). Now, I need to get rid of those proteins that have similar sequences (that 30% similarity benchmark that bioinformaticians use) and keep sequentially dissimilar proteins in my data set. Which one would be more appropriate - sequence similarity vs sequence identity?
I am not a bioinformatics researcher, however, my project requires this task to get accomplished. So, I don’t have much of a coherent idea regarding these concepts (I am willing to learn). I would also love to learn which tools are used to assess sequence similarity and identity of protein sequences in an in-house prepared data set (not an online data base).
Feel free to refer me to resources that may contain relevant and comprehensive information. I duly appreciate any suggestions. Thanks a million in advance.
Dear Sumon,
The sequence identity is considered in its literal sense by comparing (in pair-wise alignment) when two amino acid residues are matched. On the other hand, sequence similarity accounts for sequence identity and conservative substitutions with positive scores in substitution matrices. In this manner, sequence similarity is a broader term, and I see it a measure of homology among sequences. In order to generate datasets for machine learning models generation or for in silico structural studies, I think you should consider Sequence Similarity as a measure to keep or rule out sequences.
Kind Regards
Dear Dr. Mazhar Hussain,
Thank you for your prompt answer. I was thinking the same and completely agree with you that similarity is a better measure compared to identity.
However, I keep getting new search results on Google scholar, which are quite relevant to the project that I am trying to build, and those researchers primarily use sequence identity over similarity. They utilise a program, named CD-HIT to do so. For instance, this article Article GrAPFI: predicting enzymatic function of proteins from domai...
(one of many examples). Can identity be better compared to similarity to cluster the sequences?And if you have already known any tool that aims to cluster sequences based on similarity, I shall truly appreciate it if you could redirect me to that.
Once again, thank you so much for your upfront and coherent answer.
Perhaps UCLUST algorithm (which is a part of USEARCH, https://drive5.com/usearch/) for clustering of protein sequences will help you accomplish the clustering? However, as far as I got this algorithm works on identity percent of 50 and lower percentage is discouraged by the author (https://drive5.com/usearch/manual/uclust_algo.html).
P.S. I personally programmed and used mean shift clustering algorithm (see e.g. https://datascientest.com/en/meanshift-everything-you-need-to-know-about-the-data-clustering-algorithm), so if you are interested in controlling the process of clustering, this also could be useful. When I tried this method, I relied either on pairwise BLAST Score or % identity in the BLAST alignment, however this could actually be a lame approach - Dr. Edgar, the author of USEARCH and MUSCLE, wrote the following:
"Identities are computed using a global alignment. Clustering based on local alignments could easily be implemented in the USEARCH software, but I believe local clustering is fundamentally flawed. If you have a application that really needs it, I'll add support for local clustering"
Dear Dr. Daria V. Dibrova,
Firstly, thank you for being upfront.
Also, thank you for referring me to UCLUST, and certainly, I couldn't help but notice that 50% SI is the bottom threshold for that algo. However, we are trying to get as minimum as possible and recently, could do that with CD-HIT at 40% SI.
I looked up the mean shift clustering, and it appears genuinely interesting. Good have known such methods too. Thank you for sharing that link.