Science explainer

Why function beats binding affinity in T-cell engager discovery

A binder that binds tightly is not a drug that kills. For a T-cell engager, the question that matters is whether the molecule makes a T cell destroy a tumor cell, and that answer depends on far more than affinity. ShennonBio selects its leads on that functional answer from the first experiment.

Design is fast, proof is slow

Generative models now propose binder sequences on demand, and structure prediction ranks likely binders before anything is made. Validation has not kept up. Cell-based functional assays still run in plates, a few hundred candidates at a time, and each design, make and test cycle takes weeks to months. Public datasets are dominated by structure and binding; very few pair a sequence with a cell-based functional readout. Function, not design, is the bottleneck [1].

Binding is an input; killing is the output

A T-cell engager (TCE) works by forming an immune synapse between a T cell and a tumor cell. Whether that synapse forms, and whether it triggers killing, depends on where the binder lands on the target (membrane-proximal or distal epitope), on the distance and geometry the format imposes, on the affinity and valency of the CD3 arm, and on how much target the tumor cell carries. In multispecific TCEs, format and geometry drive killing independently of binding. Two binders with the same affinity can give very different potency once they are built into a molecule, and the strongest binder is often not the best drug.

Tens of thousands of permutations

Multiply binders by epitopes, CD3 arms, formats and valencies and a single program generates tens of thousands of candidate molecules. A plate-based cascade can test a few hundred, so most of that space is never explored and leads are chosen on what was convenient to make. A screen that tests function at the scale of the design space changes what can be found.

Selecting on what matters, from day 1

On TCELERATOR™, each TCE design is paired in its own droplet with a tumor cell and a reporter T cell. The platform reads T-cell activation and killing for more than 10 million single-cell interactions in one run, sorts the droplets that carry the functional signal, and sequences the hits. Leads are selected on activity rather than on binding, and more than 80% of post-sort clones are confirmed in bulk killing assays [1]. Where conventional discovery tests binders sequentially over months, this identifies functional leads in a single experiment.

The same logic applies to T-cell receptors. Earlier neoantigen screens asked which T cell binds a predicted peptide, for one patient. TCELERATOR asks which molecule makes a T cell kill, for an off-the-shelf drug. At SITC 2025, the lead TCR found this way, with no affinity optimization, controlled tumors and extended survival in mice [2].

Closing the loop with computational design

Function data is also what design models are missing. ShennonBio's AFFINIS™ designs binders computationally from public and in-house datasets, TCELERATOR validates them experimentally in a single run, and the validated binding and activity data retrain the model for the next design round. Each cycle returns a paired sequence and functional readout for every hit, with negatives sorted away and available on request. For AI partners, that loop turns a structure model that has learned binding into one that can learn function [1].

Related reading

• How TCELERATOR screens 10 million single-cell interactions in one run
• Why a trispecific T-cell engager for small cell lung cancer
• Why intracellular targets need TCR-based T-cell engagers
• All science explainers

References

1. ShennonBio TCELERATOR technical presentation, Oct 2026.

2. Li K, et al. J Immunother Cancer 2025;13(Suppl 2):A383 (SITC 2025).