Designing AI-Powered Molecular Patrols for Next-Gen Immunotherapy

A recent study published in Science, titled “Design of high-specificity binders for peptide–MHC-I complexes”, introduces a revolutionary approach: using generative AI to design novel proteins from scratch that can selectively recognize and bind to specific peptide–MHC-I (pMHC-I) complexes on the surface of cancer or virus-infected cells. These engineered “molecular patrols” pave the way for safer and more personalized protein and cell therapies.

The Cellular “Display Window”: pMHC-I

Every cell displays fragments of internal proteins (peptides) on its surface via pMHC-I complexes, acting as a cellular status display. Healthy cells show “self” peptides, while cancerous or infected cells display abnormal ones. T cells monitor these peptides through their T-cell receptors (TCRs) and attack when they detect anomalies.

The Challenge of Specificity

Designing proteins that can accurately bind to specific pMHC-I targets is incredibly difficult:

• Peptide differences can be as small as a single amino acid, making precise recognition essential to avoid harmful off-target effects.

• MHC molecules vary across individuals due to genetic diversity (HLA alleles), requiring highly personalized designs.

Traditional methods, like screening natural TCRs or engineering antibodies, often fail due to low specificity, high cost, and long timelines.

Enter Generative AI: RFdiffusion

The research team turned to RFdiffusion, a generative AI model that creates protein structures de novo. They provided the model with the structure of target pMHC-I complexes and guided it to generate protein “binders” that:

• Arch over the peptide-binding groove

• Focus interactions on the peptide, avoiding conserved MHC regions

These designs were refined using:

• ProteinMPNN for optimal amino acid sequences

• AlphaFold2 to validate folding and binding accuracy

Out of thousands of AI-generated designs, the team selected binders for 11 diverse pMHC-I targets, including antigens from cancer (e.g., WT1, MAGE-A3, PRAME) and viruses (HIV, SARS, YFV).

From Design to Reality: Yeast Display & Functional Testing

The binders were displayed on yeast cells and tested for specificity using dual-color flow cytometry, distinguishing true targets from similar off-targets. Example: Binder yfv-2 showed 88.8% on-target binding and only 6.96% off-target, demonstrating high specificity. A co-crystal structure of mart1-3 with its target peptide showed a near-perfect match with the AI design (RMSD of 0.4Å).

Activating T Cells with CAR Constructs

The most promising binders were incorporated into CAR constructs and expressed on Jurkat T cells. Example: mage-513 triggered strong activation (CD69+) only in the presence of its target MAGE-A3 peptide, and not a nearly identical Titin-derived peptide. Mutational analyses confirmed the precise interactions responsible for specificity—validating AI's ability to design effective “molecular keys” for specific “locks.”

Beyond Structure: Redesign and Prediction-Based Engineering

Two powerful expansions:

1. Partial Diffusion for Rapid Redesign – Starting from an existing binder, only the peptide-interacting regions were re-optimized to target new peptides, achieving high efficiency with minimal redesign.

2. Design from Predicted Structures – The team used AlphaFold-predicted pMHC-I structures (e.g., for PRAME) as design templates and still achieved high specificity, eliminating the need for experimentally determined structures.

Toward Personalized Immunotherapy

Finally, CAR-T cells made with binders targeting WT1 and PRAME demonstrated >80% killing of target cells in vitro, with minimal off-target effects. This study marks a new era of rational, scalable, and AI-driven drug design. With rapid sequencing and HLA typing, personalized immunotherapies could be created in days instead of years, using AI to craft patient-specific binders from scratch.