Cryo-EM Validates an AI-Designed Antibody: A Collaboration with Nabla Bio

Our partners at Nabla Bio recently published a report showcasing the fully computational design of T-cell-engaging antibodies against peptide-MHC class I targets, including a dual-variant anti-KRAS antibody that engages both G12V and G12C mutants while sparing wild-type. The work was generated by JAM-2, their generative biomolecular design model, from target sequence alone.

Our team at NIS contributed the cryo-EM validation for the lead anti-KRAS G12V design.

‍

What the cryo-EM analysis showed

The complex we solved was the JAM-2–designed VHH bound to KRAS G12V:HLA-A*03:01:β2m, with an anti-β2-microglobulin Fab included as a fiducial marker for particle alignment. From 566,000 particles, the full map reached 3.30 Å nominal resolution, improving to 3.25 Å across the VHH–pMHC interface after local refinement.

The headline result was the agreement between the experimental structure and the design model. The two superposed to 0.93 Å Cα RMSD across the entire VHH–pMHC complex. When the structures were aligned on the MHC alone, the designed VHH still superimposed to 1.26 Å on the experimental coordinates, meaning the model captured not just the antibody fold, but its docking geometry, approach angle, and register on the pMHC surface.

The interface itself also matched the design intent. The single residue distinguishing KRAS G12V from wild-type (the Val12 side chain) inserted into a complementary hydrophobic pocket on the VHH, fully buried on complex formation. This packing arrangement explains, structurally, why the antibody discriminates G12V from wild-type KRAS at the level of one amino acid.

‍

Why this matters

Experimental validation of computationally designed molecules is critical, regardless of the outcome. It is the fundamental step that connects design to reality. In this case, the experimental structure also returned a notably strong result. The designed VHH–pMHC complex superposed to 0.93 Å Cα RMSD against the experimental coordinates, on a design generated from target sequence alone with no prior structural data on this target. That level of agreement, for a fully computational design, reflects how far generative antibody design has come.

We work on this kind of validation regularly. The pairing of generative AI design with structural confirmation is becoming a standard part of how the strongest AI-driven discovery programs operate, and cryo-EM, alongside other structural methods, is well positioned to contribute. We wrote about how structural biology fits in AI-driven drug discovery loop in our recent piece. This collaboration is a concrete recent example of the argument made there.

Read the full report

The full Nabla Bio report is available here. Congratulations to the Nabla team on a remarkable piece of work.

If you are running an AI-driven discovery program and want to talk about how cryo-EM validation fits into your workflow, please get in touch with our team.

‍