Question about MSi score/pranks

[mjsl...@gmail.com]

Hello,

I've managed to run the Terra implementation of HLAthena (see below for a potential bug in the website implementation), and I have a question about the prediction scores & %ranks. I'm somewhat familiar with netMHCpan, and they suggest %rank cutoffs of 0.5 & 2 as designations of strong binders and weak binders, respectively. From your paper, it looks like you use 0.1% as a criterion for assessing PPV. Is this also the cutoff you would define as representing a "strong binder"? Are there other cutoffs that you would recommend for selecting peptides for further assessment? Thanks for any advice on this issue.

Potential bug: Under the "Predict" tab of the website, I attempted to upload a peptide file that appeared to be in the proper format (and contained only valid amino acids, etc.). The upload step took a very long time, then appeared to complete; shortly after, I got the generic error message that was referred to in an earlier query (suggesting to check the log files, though no log files are apparent). I'm pretty sure I found the source of the problem: standard text-file formats on Unix-like platforms & on Windows have different end-of-line sequences to separate lines of a text file (LF for *nix & CR/LF for Windows). I uploaded a file from a Windows machine and got the behavior described. When I uploaded the same file - transformed into a Unix x format - it worked properly. Given that this is a web implementation, it would seem better to allow the prediction tool to accept any kind of text file (or to include a warning about the necessary text-file format).

Anyway, thanks for making this resource available, and I look forward to hearing from you.

Michael

[Sisi Sarkizova]

Hi Michael, I'm glad you were able to run HLAthena on Terra. Let me know if we can help clarify this setup further.

Re prediction scores and ranks - The ranks and scores in HLAthena are very similar to those in NetMHC (scores are most analogous to the EL mode of NetMHCpan 4.0). Choosing 0.1%rank as cutoff was motivated by the estimated number of presented peptides by a cell in an endogenous setting, a more detailed argument from our first paper is provided below (Abelin et al., 2017). In practice, the cutoffs to select would depend on the intended application. If interested in predictions with lower false positive rates then a more stringent cutoff should be applied; if interested in exploring possible presented peptides with secondary validation experiments that can reduce the larger set of false positives then a less stringent cutoff could be beneficial. That being said, the 0.5% and 2% cutoffs for strong and weak binders respectively can still be used as a general guideline.

"MS peptides among a 999-fold excess of decoy peptides; the top-scoring 0.1% of peptides were selected as positives. Because there are approximately 10 million 9-mers in the human proteome, and each allele presents approximately 10,000 of these, the 1:1000 ratio closely mimics the reality of the epitope selection problem. [...] We ruled out 0.01% because we have directly observed more than 1000 9mers for some alleles,  and 1% would imply that 100,000 peptides are presented per allele, which is inconsistent with previous biochemical estimates (Walz et al., 2015)."

Re end-of-line characters - I tried to reproduce the problem by using the sample input file and switching the format (available to download from the bottom of the left-hand side menu). The upload and predictions ran without an issue on my end in both cases. Could you try on your side with the sample file and let me know if you see the same problem? If not, then it would be very helpful if  you can share the input file with which you see the error.

Thanks,

Sisi

--
You received this message because you are subscribed to the Google Groups "HLAthena" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hlathena+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/hlathena/35250ea8-d2a1-44c9-bb0d-6a861065088d%40googlegroups.com.