The immune system decides whether to kill a cell by reading its surface. All nucleated cells degrade intracellular proteins and display the fragments on major histocompatibility complex (MHC) molecules, and circulating T cells inspect these peptides through their T cell receptors (TCRs). Whether a given TCR–peptide–MHC pair binds therefore determines whether a tumour is attacked, an infection cleared, or both missed — and engineered cell therapies attempt to control exactly this decision, since a therapeutic receptor must engage abnormal peptides without engaging the self peptides of healthy tissue. The obstacle is a mismatch of scale. One individual carries more than 10⁷ distinct receptor sequences and a single tumour can present thousands of peptides, while measured binding pairs number in the thousands and are severely concentrated: a handful of epitopes account for the large majority of validated pairs, and more than half of all known epitopes have exactly one reported binder. A useful model must generalize to peptides for which essentially no measurement exists.
Reported accuracies suggest current methods do so; two artefacts explain why they do not. The first is evaluation: most benchmarks allow the same peptide to appear in training and test splits, so headline numbers largely measure interpolation across seen epitopes, and independent reassessment of published predictors finds near-chance performance once unseen peptides are enforced. The second is the input. Structure-based predictors are supplied with a single docked TCR–pMHC complex, but that pose is produced by procedures parameterized on known binders — the input encodes the label it is meant to predict, and for an unmeasured peptide no such complex exists at all. Beneath both artefacts lies a representational error: a single fixed structure is the wrong object. The peptide-contacting CDR3 loops are the most conformationally labile region of the receptor, the peptide samples multiple states within the MHC groove, and the interface forms by induced fit. At the cell surface the decision is made under mechanical load, where TCR–pMHC bonds behave as catch bonds and force-dependent bond lifetime, rather than static contact area, determines signalling. A crystal or docked structure is one sample from this ensemble — the bound-state minimum — knowable only once binding is already known.
Existing work falls into three broad classes. Sequence-based models learn directly from CDR3 and epitope strings; they scale freely but, with so few epitopes supervised, tend to latch onto germline shortcuts such as V gene identity, which can account for the great majority of the variance in a model's per-receptor bias while carrying little epitope-specific signal. Structure-based models consume a docked or co-folded complex; they encode genuine physics but inherit the label leakage above and cost O(NM) structure predictions, which forecloses proteome-scale screening. Representation-transfer models import embeddings from protein language models; they generalize across sequence space but remain agnostic to the mechanism, and their advantage largely disappears under leakage-free splits.
We are building TEMPER to replace the single-pose representation with the physical object the decision actually depends on. Recognition is predicted from ensembles: the conformational distributions occupied by the free receptor and the free pMHC, and their redistribution on contact — quantities that are defined for every pair, including non-binders and pairs never co-crystallized. Because each molecule is encoded independently, cost is O(N+M) ensemble generations plus one forward pass per pair, so full-repertoire against full-proteome screening becomes feasible rather than aspirational. Supervision targets a force-observable readout, bond lifetime under load, in addition to a binary binding call, so the model is fit to the signature that distinguishes agonists from non-agonists in vivo. Evaluation is leakage-free by construction: no information about a test peptide reaches the model, through either shared peptides or a pre-docked input. Because these inductive biases are physical rather than statistical — where flexibility is localized along the receptor, which groove positions are anchored and which remain mobile, which contacts must persist long enough to trigger signalling — they are expected to transfer to peptides absent from the training distribution, the regime in which current predictors collapse. The longer-term aim is to close the loop: use the model to propose candidate pairs, resolve them with assays cheap enough to run at scale, and retrain, so that the measurement bottleneck shrinks with each cycle rather than staying fixed.