AI-designed minibinders arm T cells to attack cancer in weeks

AI-designed minibinders train T cells to attack cancer in weeks

Scientists have built an AI platform that designs small proteins-minibinders-that guide T cells to cancer cells. Instead of searching for rare, natural T-cell receptors, the system engineers binders that lock onto peptide-MHC (pMHC) targets on tumor cells and does it in 4-6 weeks from concept to lab prototype.

The result: faster prototyping, tighter control over specificity, and a clearer path to personalized cell therapies.

How it works

• In-silico design: Generative models propose minibinders for a defined pMHC target.
• Virtual screening: Simulations assess affinity and filter out likely off-targets across healthy tissue pMHCs.
• Build-and-test: Selected designs are synthesized and inserted into immune cells (IMPAC-T cells) for functional testing.
• Validation: Engineered cells are checked for cytotoxicity against target-positive cancer cells.

Proof of concept: NY-ESO-1

In work published in Science, teams at the Technical University of Denmark (DTU) and Scripps Research targeted the well-characterized cancer antigen NY-ESO-1. The AI designed minibinders that bind the NY-ESO-1 pMHC with high affinity. When expressed in T cells (IMPAC-T cells), these constructs recognized and killed NY-ESO-1-positive cancer cells in vitro.

"It was incredibly exciting to take these minibinders, which were created entirely on a computer, and see them work so effectively in the laboratory," said Kristoffer Haurum Johansen of DTU.

Generalizing to new targets

The platform also handled a previously unmapped pMHC from a metastatic melanoma patient (RVTDESILSY/HLA-A*01:01). Despite the lack of a solved structure, the system produced binders that matched the new target. This supports use on patient-unique neoantigens where structural data is scarce.

"We are essentially creating a new set of eyes for the immune system," said Timothy P. Jenkins of DTU. "Our platform designs molecular keys to target cancer cells using the AI platform, and it does so at incredible speed."

Engineering safety in from the start

Off-target toxicity remains the critical risk for pMHC binders. The team added a preclinical "virtual safety check," screening designs against a broad panel of healthy-tissue pMHCs to identify likely cross-reactivity before wet-lab work.

"By predicting and ruling out cross-reactions already in the design phase, we were able to reduce the risk associated with the designed proteins and increase the likelihood of designing a safe and effective therapy," said Sine Reker Hadrup of DTU.

From bench to bedside

The envisioned workflow mirrors current autologous cell therapies: collect blood, isolate immune cells, insert the validated minibinder, expand, and reinfuse. The difference is speed and target scope-digital design avoids waiting for patient- or donor-derived receptors and can address solid tumors and rare mutations that are hard to tackle today.

First-in-human trials are estimated to be about five years out, pending further safety and manufacturing studies.

Why this matters for researchers

• Faster hypothesis-to-assay: 4-6 weeks to functional prototypes compresses antigen validation cycles.
• Patient specificity: Feasible path to individualized pMHC targets, including low-frequency neoantigens.
• Built-in liability management: Computational cross-reactivity screens reduce dead-ends and unsafe candidates.
• Manufacturing alignment: Constructs plug into established cell engineering pipelines.

Technical notes to watch

• Model fidelity: Affinity prediction for pMHC interfaces, especially for unseen HLA contexts, will drive success rates.
• Coverage of safety panels: Breadth and representativeness of healthy-tissue pMHC libraries are pivotal.
• Functional equivalence: Head-to-head comparisons with natural TCRs across avidity, persistence, and exhaustion.
• Solid tumor microenvironment: Pairing with trafficking/persistence strategies (e.g., chemokine receptors, cytokine support) may be necessary.

What you can do now

• Audit your antigen pipeline for pMHC targets amenable to binder design and in-silico triage.
• Integrate structural modeling and off-target prediction early to reduce wet-lab burden.
• Stand up standardized cytotoxicity and cross-reactivity assays to benchmark designed binders versus natural TCRs.

For institutional teams building AI capacity in biomedicine, see curated upskilling paths at Complete AI Training: Courses by Job.

Learn more about the institutions behind this research: Technical University of Denmark (DTU).