AI-Designed Mini-Protein Precisely Targets Cancer Antigens Like a T Cell

Targeting peptide-MHC complexes on cancer cells offers a powerful route to tumor immunotherapy, but existing approaches face serious limitations. Natural T cell receptors bind too weakly for therapeutic use, while engineered antibody-based TCR mimics often dock too heavily on the MHC scaffold rather than the peptide itself — driving off-target toxicity that has stalled clinical development. This study from Garcia and colleagues at Stanford set out to solve that problem from scratch using computational protein design, building a compact α-helical 'mini' TCR mimic without starting from any natural protein template.

The team deployed RFdiffusion to generate 50 candidate four-helix bundle scaffolds (80–100 amino acids) and selected the most structurally compact for fold-conditioning over the NY-ESO-1/HLA-A*02 target. The NY-ESO-1 peptide's upward-facing residues Met4, Trp5, Thr7, and Gln8 were specified as design hotspots. RFdiffusion then docked 100 distinct helical bundle folds over the peptide-MHC target, and ProteinMPNN sampled six sequences per fold, yielding 600 initial designs. AlphaFold2 scored each design using the interaction predicted aligned error (iPAE) metric; four of 100 scaffolds produced at least one design below the iPAE threshold of 10.0. Expanding to 500 sequences per hit scaffold produced 2,000 designs, from which Scaffold #1 emerged as the top candidate for its TCR-like centered docking and peptide-specific contact profile.

The top five designs were tested by yeast surface display against NY-ESO-1 and MART-1 HLA-A*02 tetramers, with two out of five specifically recognizing NY-ESO-1. The lead binder, designated mini-TCRm 1.1, was expressed recombinantly in E. coli and characterized by surface plasmon resonance, yielding a dissociation constant of 9.5 nM for NY-ESO-1 HLA-A*02 with no detectable binding to MART-1 HLA-A*02 — an exceptional selectivity for a molecule recognizing the same MHC allele presenting a different peptide.

X-ray crystallography of the mini-TCRm 1.1/pMHC complex, solved at 2.05 Å resolution, confirmed a TCR-like diagonal docking mode at 40 degrees relative to the peptide groove, burying 1,216 Ų of total surface area with 31% from peptide contacts. Structurally, the A2 and A3 helices of the mini-TCRm formed two distinct hydrophobic pockets that captured the Met4 and Trp5 residues of NY-ESO-1 independently — a sharp contrast to natural TCRs and antibody Fabs, which use a single large CDR loop-formed pocket to accommodate both residues together. The rigid helical scaffold also made five hydrogen bonds and one salt bridge to MHC, helping orient the binder precisely without relying on conformational flexibility.

To assess cross-reactivity risk, the team ran alanine scanning to confirm Met4 and Trp5 as the critical anchoring residues, then performed a Hamming distance search of 14,363 HLA-A*02-presented 9-mers from mass spectrometry data. Only five peptides had a Hamming distance of four from NY-ESO-1; two additional peptides matched the Met4-Trp5 motif exactly. A ProteinMPNN-based structural compatibility score correctly identified two of these seven candidates (off-targets 3 and 5) as binders in subsequent T2 cell staining — one from PIP5K1A (a ubiquitous kinase) and one from KIRREL2 (expressed mainly in pancreatic beta cells). Binding was confirmed but weaker than NY-ESO-1. In bispecific T cell engager format, the mini-TCRm drove selective cytotoxicity against NY-ESO-1-positive target cells, validating its therapeutic potential despite identifying these two manageable off-targets.