This collection contains the experimental data used to validate BoltzGen in its initial preprint release.

It includes four distinct classes of computational binder designs:

Nanobodies Binding to Novel Proteins

• Synthetic nanobody sequences and structures designed to bind protein targets with low homology (<30% sequence identity) to any previously structurally characterized complexes.

• Designed to assess generalization and innovation beyond known binding motifs, as these targets have not been observed in any prior training data.

Proteins Binding to Novel Proteins

• Designs of general protein scaffolds (outside of nanobody architecture) targeted against the same set of novel protein targets as above, providing a point of comparison between different binding modalities for unexplored target space.

Nanobodies Binding to Benchmark (BenchBB) Proteins

• Binders designed to target benchmark (BenchBB) targets with well-characterized structures and many known binders present in public datasets (TNFα, PD-L1, PDGFR, IL-7Rα, InsulinR).

• These cases allow for direct performance benchmarking, as the availability of known binders makes binding efficacy evaluation more rigorous but less reflective of novel discovery scenarios.

Proteins Binding to Benchmark Proteins

• General protein binder designs targeting the same benchmark (BenchBB) targets, enabling comparison with nanobody designs and with literature standards.