After a good bit of fighting with the code, I'm super excited to finally announce the release of my first ever app! AIMS - An Automated Immune Molecule Separator: https://t.co/P8n1QwOFVU
@VibeCoderMcSwag Love to hear it! Replicating the preprint from earlier this year? It was cool to see that paper, but I'm skeptical that their model will generalize to unseen data (AIMS didn't!).
I don't use discord, but let me know if anyone has questions! Always happy to help
After a good bit of fighting with the code, I'm super excited to finally announce the release of my first ever app! AIMS - An Automated Immune Molecule Separator: https://t.co/P8n1QwOFVU
A great thread with some much needed testing of ML-based binding predictions! When AF3 can't fall back on its "secret sauce" of leveraging MSA for predictions, it shows just how far we are. Similar to the fundamental issues in TCR/Ab binding predictions. Excited for this paper!
Here is the Alphafold3 prediction. It is worse than random. High false positive and high false negative (iPTM/PTM shown - other metrics look very similar). 4/n
A great talk by Andreas, strongly recommend it to everyone interested in TCR specificity! Big takeaway for me is that we are quite a ways away from ML being useful in the specificity prediction problem given a random TCR. 99.99% of Andreas' new sequenced TCRs were orphans!
What are the rules of the immune receptor code? In my talk @KITP_UCSB last Friday I summarised our recent findings, provide a perspective on why ML approaches have so far not achieved breakthrough success, and propose a path forward:
https://t.co/ClmNVKi8Ri
I was invited to write a textbook chapter for "Methods in Molecular Biology: Immunoproteomics"! The final version will be behind a paywall, so here's a free copy. In it, I outline a step-by-step protocol for using AIMS to analyze immunopeptidomics data.
https://t.co/z5rVakN3O2
@biophilross My pessimism on the ML especially comes from both the positive and negatives being ill-defined. No one ever tests the negative, and we aren't super sure of our positives. Does the interaction need to be high-affinity? Not necessarily. Catch bonds? Maybe, maybe not.
Recently been talking with a lot of people about the shortcomings of AI in biology, especially in T cell biology. Came across a decent summary of some of the issues today: https://t.co/QJp5WHgzeX
A solid read! I recommend it, even if "crisis" is a bit strong.
@biophilross By my last count we've got 149 human TCR-pMHC complexes. And if every new structure adds new information on binding, can we confidently say where this information gain will plateau? Hard to say, so I'm not sure I could throw out an estimate on what the "right" scale is
@biophilross @whoisyoo But we just haven't seen such degeneracy in the TCR-pMHC structures. The closest thing to a "rule" is the diagonal docking orientation, and even then we've found some wonky exceptions to this! Maybe somewhere in that 10^19 space there are even stranger binders that still signal.
@biophilross @whoisyoo So my gut reaction is to say still no. 10^4 interactions still falls well short of the estimated number of possible receptors (~10^19). Now to be a bit of an optimist, *maybe* there is serious degeneracy in the TCR-pMHC interaction to effectively reduce this number...
@biophilross @whoisyoo Well 2k interactions isn't all that large in the grand scheme of things. That actually falls short of covering all possible TRAV-TRBV pairings in humans... Then you have to add in CDR3s and peptides.
You could *maybe* train one to work for a single peptide-SIINFEKL or GIGLFVFTL
@biophilross @whoisyoo Read that piece but not the preprint yet. I think the future is very far off in this application. 1. We don't even fully understand how TCR-pMHC signaling works and 2. the training data (experiments thus far) have sampled far too little of the possible interaction space.
Also, I'm really happy this was published in a journal where you can see the reviews and my responses. I think the 35 pages (!!!!) back and forth mostly with one unnamed, pugilistic, cantankerous reviewer really hammers home how robust this analysis is. Enjoy!
Really proud that this paper is finally out today in @eLife! I think a lot of the analysis in this paper can/should change the way we think about a really fundamental aspect of immunology: the interaction between TCRs and MHC
https://t.co/sVDGB7YSJn
However, thymic selection is also really important for determining the binding orientation! Biophysically, there's no reason TCRs can't dock on the MHC in a reverse orientation. This has been observed a few times, but is likely rare *because* of selection.
My first publication as corresponding author is out today in its final form! I'm hoping this methods paper can essentially be the AIMS "bible". If you want to see how we can analyze TCRs, antibodies, peptides, and MHC all in one go, take a look:
https://t.co/dve3lVuvCI
We don't think this is the *only* way you can get polyreactivity. And the field's ability to rigorously classify an antibody as "polyreactive" is still a bit shaky. But we think that these two manuscripts help open up some new ways of thinking about polyreactivity!
My last bit of graduate school-related work is out today in @CellReports !
https://t.co/deYFBx1Bll
Some great experimental work by @mtborowska coupled with some MD simulations really nicely confirmed some of the AIMS predictions we previously published in eLife.
This increase in inter- and intra-CDR loop interactions may work in two ways to create an "inoffensive" binding surface to a range of binding partners. First, it can sequester side chains and promote backbone interactions. Second, it reduces the entropic penalty to binding!