Lab Newsletter — September 20, 2026: What the T-Cell Sees

Sun, 20 Sep 2026 03:00:21 +0000

Last week we watched AI design antibodies — the B-cell arm of adaptive immunity. Today we turn to the other arm, and a harder recognition problem: the T cell. Every cell in your body continuously chops up its proteins and displays the fragments on MHC molecules, like a shop holding up samples of everything inside. T cells patrol this display, and when a T-cell receptor (TCR) recognizes a peptide-MHC that looks foreign — a viral fragment, a tumor mutation — it raises the alarm. Two prediction problems sit at the heart of it: which peptides get presented, and which TCRs recognize them. Today’s digest is about the models learning to read that immune code.

🪟 The display: predicting what gets presented

Not every peptide makes it onto an MHC molecule — presentation is a filter, and it’s the crux. Reynisson et al. (Nucleic Acids Research, 2020) built the field’s workhorse, NetMHCpan-4.1. They frame the stakes precisely: MHC molecules “are expressed on the cell surface, where they present peptides to T cells, which gives them a key role in the development of T-cell immune responses,” and “the binding between MHC and antigenic peptides is the most selective step in the antigen presentation pathway.” Their pan-allele predictor improved accuracy “by concurrent motif deconvolution and integration of MS MHC eluted ligand data” — learning directly from peptides that were actually caught being presented. Predict this step well, and you can scan an entire proteome for what the immune system might see.

🧵 Beyond binding: model the whole processing pipeline

Binding to MHC is only the last step; before it, a peptide has to be cut out and transported. O’Donnell, Rubinsteyn & Laserson (Cell Systems, 2020) captured more of that pipeline in open source with MHCflurry 2.0. Their insight: mass-spec-identified ligands don’t just reveal binding motifs — “the identified ligands also reflect the antigen processing steps that occur prior to MHC binding.” So they “developed an integrated predictor of MHC class I presentation that combines new models for MHC class I binding and antigen processing.” It’s a small but important shift in framing: predict presentation, not just affinity — and ship it as a tool anyone can run, the same open-model spirit behind the lab’s BioImage Model Zoo and BioEngine.

🔑 The receptor side: recognition is partly predictable

The harder half is the receptor. There are more possible TCRs than stars in the galaxy — could recognition ever be predictable? Dash et al. (Nature, 2017) answered yes, and launched computational TCR analysis. Their title says it: “quantifiable predictive features define epitope-specific T cell receptor repertoires.” TCRs “mediate recognition of pathogen-associated epitopes through interactions with peptide and major histocompatibility complexes (pMHCs),” and though V(D)J recombination generates almost limitless diversity, the receptors that see the same epitope turn out to share measurable sequence patterns. Their TCRdist metric made those patterns computable — the seed of everything that followed.

🧠 Learn the repertoire: deep representations of TCRs

If specificity leaves a sequence signature, deep learning should be able to read it. Sidhom et al. (Nature Communications, 2021) built DeepTCR for exactly that — noting that “the ability to learn complex patterns in data has tremendous implications in immunogenomics.” DeepTCR is “a suite of unsupervised and supervised deep learning methods able to model highly complex TCR sequencing data by learning a joint representation of a TCR by its CDR3 sequences and V/D/J gene usage.” Instead of hand-crafted distances, it learns the features that matter — the same representation-learning move the lab bets on across proteins, genomes, and images.

🤝 Predict the binding: paired chains, and an honest lesson

The prize is predicting, for an arbitrary TCR and peptide, whether they bind. Montemurro et al. (Communications Biology, 2021) pushed toward it with NetTCR-2.0, which “enables accurate prediction of TCR-peptide binding by using paired TCRα and β sequence data,” showing that even “‘shallow’ convolutional neural network … architectures are adequate to deal with the problem complexity.” But their most valuable contribution may be candor: “current public bulk CDR3β-pMHC binding data overall is of low quality,” and real progress is “contingent on paired α/β TCR sequence data.” It’s a lesson the lab knows well — in this field, better data and benchmarks often matter more than a bigger model.

🎯 To the clinic: neoantigens and cancer

Why does all this matter now? Because the immune system can be taught to attack tumors — if we can find the right targets. Lu et al. (Nature Machine Intelligence, 2021) built pMTnet for that mission, starting from the clinical bottleneck: “neoantigens play a key role in the recognition of tumor cells by T cells. However, only a small proportion of neoantigens truly elicit T cell responses, and fewer clues exist as to which neoantigens are recognized by which T cell receptors.” Their “transfer learning-based model … to predict TCR-binding specificities of neoantigens, and T cell antigens in general,” borrows strength from abundant unlabeled data to fight the scarcity of labeled binding pairs — pointing straight at personalized cancer vaccines and TCR therapies.

🧫 Why it’s our kind of problem

Read across the six and the shape is familiar: immune recognition is a representation-learning problem on biological sequences, the same wager the lab makes with protein and genome language models and imaging foundation models. Two themes recur that are squarely the lab’s own. First, open models and shared data win — MHCflurry, NetTCR, and NetMHCpan are open tools trained on communal ligand databases, exactly the BioEngine idea. Second, the binding constraint is data, not architecture (NetTCR-2.0’s warning, pMTnet’s transfer-learning workaround) — which is why careful benchmarks and better datasets are as valuable as new networks. And it connects to the lab’s imaging future: recognition doesn’t happen in a spreadsheet but in tissue, where a T cell physically meets its target — spatial immunology that platforms like Agent-Lens and the REEF farm are built to watch. Antibodies were last week’s key; the TCR is this week’s lock. Learn to read what the T cell sees, and you can start to direct it.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.

neoantigens | AICell Lab