We designed nanobodies with variable CDR lengths using a combination of template-binder hallucination and partial-hallucination protocols implemented in ColabDesign. Custom loss functions were incorporated to ensure that i) binding occurs primarily through the CDRs, ii) CDRs achieve high pLDDT confidence, and iii) the sequences retain antibody-like characteristics according to the AbMPNN score.

For binder generation, we used a nanobody framework without CDR loops as the fixed template. Only the CDR regions were designed, with loop lengths systematically varied. Designs were targeted toward three known epitopes on the Nipah glycoprotein: Trp504, Met267, and Ile517. Yet, the Met267-directed designs achieving the strongest in silico performance metrics.

Model quality was evaluated by refolding each complex using all AlphaFold2-Multimer models trained with templates (models 1 and 2). Templates were provided for both the target and the binder while removing all inter-chain information. We collected quality metrics as pLDDT, CDRs pLDDT, ipAE, ipTM and Calpha-RMSD respect to the design. When specified in de candidate, we used these metrics with to filter with the following criteria: i) all ipTMs >0.8, ii) all ipAEs < 8 Angstroms, iii) CDR pLDDT>0.8, iv) Calpha-RMSD respect to the design < 1.5 Angstroms.

All designs were refolded with Boltz-2 without MSAs to avoid introducing evolutionary biases. To ensure that the Nipah glycoprotein G was accurately folded under no-MSA conditions, we provided a crystal structure as a template. In addition, we excluded the flexible C-terminal region of glycoprotein G from the predictions, as this segment is unlikely to fold reliably and could introduce noise into the results.

The designed nanobody binders were predicted in single-sequence mode and without templates. From these predictions, we computed ipSAE values and VoroMQA quality scores to assess structural reliability and evaluated the RMSD relative to the original design as a measure of self-consistency.

All the selected designs had ipSAE values ranging between 0.80 and 0.88 according to our ipSAE scores.