Yes some of them are. We had so many data points that it was impossible to just repeat and standardize it all with one cell line. For a few that I did k562 side by side with T2, the difference between pep vs no pep was the same in both.

We had no reactivity against K562 WT. I know it's phenotypically HLA- but I don't know if it's B2M- or HLA-. Would that be what you were looking for?

I think conditioned kn that you know these are binders you probably can find the most self-reactive (usually with ridiculously high binding signal). But a lot of the others wouldn’t score poorly and still be non specific.

Yes. I definitely think there’s transient associations with MHC only strong enough to signal. But I also have moved to think that T2 also has a good variety of peptides presented despite tap deficiency.

Surprisingly high! I think out of 120 only some 2-3 didn’t express and 4-5 expressed but noticeably lower. For all fixed scaffold designs, we actually had absolute 100% expression rate.

I think it would perform poorly. Maybe you can come up with metrics averaging over many many predictions and they would be a decent predictor of self-reactivity or not. However, it’s hard to exactly predict where the off-target would come from.

I think in terms of sequence and structure they look very normal. But they’re usually more active than native TCRs. I guess the fact that they haven’t undergone any thymic selection…

On the epitope side it’s just the allele and the non anchor residues of the peptide. Maybe adding in 10 mers. But the models do terribly on 10 mers

You could be right. I think someone should try to make sure. On the TCR side you have triple the design space at minimum even with fixed length. You could vary the alpha and beta and also sample different lengths for each CDR.

that utilized de novo pairings of alpha and beta chains.

The paired CDR3s are only an initial seed. Honestly you could use unpaired ones and crank the compute a bit more and you will still get convergence to binders that work. They undergo quite a lot of changes during the design process. We even had data that didn't make it into the paper for brevity

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.

Sorry for missing some details in the methods section. For A02:01, the antigen presenting cells were mostly T2 cells. For some data they might have been through HEK293T and for some K562-A02:01. For MAGE-A3, we used HEK293T-A01:01 cells because K562 is MAGE-A3+. For all others, they were K562 + MHC