AI has gone too far

K562 is a funny one - it still has it's B2M and HLA, but usually just doesn't express them (unless given an IFN kick or something).

So does that mean say the no peptide green/orange dots in Fig4A are responding to K562 + transgenic A2? If so yeah that would go a long way towards HLA reactivity!

Agreed, there'll definitely be TAP-independent peptides (and there are some immunopeptidomes out for T2). Probably not much though, certainly by number (going off W6/32 staining).

May not be a priority, but if you were of a mind to check then B2M KO control APCs could be instructive.

Although having said that, if your APC was mostly T2 and you're still getting super strong activation without peptide, presumably a lot of that activation isn't even from (p)HLA, at least for those near the diagonal.

You could test it out on the more selective ones you did have with tested peptide variants, e.g. from fig 4E. At the very least though, presumably if a TCR were to score strongly against *most* pMHC structures, then it's a decent bet that it's probably going to be a promiscuous binder, perhaps?

OK just some bsky brainstorming: once you have a hit, how feasible would in silico thymic selection be in your system? Like, for a given donor HLA type, run the human proteome, predict all binding self 9-mers, and then generate QC metrics for all of those for that given TCR (say 0.5-1e6 pMHC)?

Yea it looks like activity isn't the issue, so much as specificity right

Computational X-scan didn't work super well for improving the specificity of the design process. There's potential for deorphanization but the design problem is actually simpler than deorphanization since you're free to sample a huge space.

...but I guess the issue is that then have no choice but to brute force your way through those 7^20 options (versus all the different places you could tweak the TCR and happen upon something quicker)?

Also, if I can get a cheeky follow up Q: how 'normal' do the final TCRs look?

Thanks for the follow up Amir, I really appreciate it!

Huh that's interesting, and shows that my intuition was way off. My naive presumption would've guessed that it would be easier to only have the 7/9ths or whatever of non-anchor positions you could vary in the epitope...

I shudder to think at the amount of super important cells that have been sorted straight into the waste tank

4) This is using naturally paired CDR3s, but shorn of framework context... so do you need pairs? Why not just mess it up, do all vs all? Intrigued to see if there's enough data to see rules about which CDR3 works where.

5) I can't wait for the de novo TCR prediction competitions to kick off!

3) Presumably a similar process could be used to iterate over a bunch of related pMHC variants and check for potential clashes that way? Computational X-scans of the epitope, would be huge. Similarly, could you not just use natural TCRs and iterate over a bunch of pMHC to deorphanise?

It's hard to say much about what they could be responding to though (as I can't seem to see what the APCs were for a lot of these expts), but I wonder if some of those are just made very HLA-bindey TCRs. However the fact they could find any specifics is remarkable.

2) Some of those off-target activation results are absolutely crackers. I mean, we always knew both that TCRs are super cross-reactive and that selection is super important, but it looks like for a lot of these TCRs just about everycell would be be catching strays...

Wow, this is very damn cool! It's pretty dense and I'm in deadline season, so I've only given it a once-over, but a few thoughts leap to mind:

1) I'm amazed they can do all that with such modest computational demands, I would've naively presumed it'd take a lot more.