We engineer RBX-1 mini-binder candidates with a TRAIRC-style pipeline: Physics-informed regularization during generation, foldability screening, and language-model–based sequence plausibility, followed by deep-learning structural validation on the full target panel.

Hypothesis: the most defensible designs are those that remain structurally plausible under more than one independent complex predictor and MSA strategy. A second model is not expected to reproduce the first model’s scores numerically; instead, it stress-tests whether the same sequences still support a coherent heteromeric fold when the architecture and priors change.

Our primary ranking used Boltz-2 on four RBX-1-related contexts per binder, with ensemble-mean ipTM, model confidence, and interface iPDE to reduce dependence on any single structural template. From that Boltz-scored pool we selected fifteen elite binders (submission ranks 1–15), drawn from the top Boltz-ensemble-structural cohort of our hybrid portfolio.

This submission is the full 100-design portfolio (30 cores: 22 Boltz-ensemble-structural + 8 pipeline-forward under a mean-ipTM floor; plus 35 single-point variants each of rbx1_v2_1390 and rbx1_v2_2336). Separately, we ran ColabFold (AlphaFold2-multimer v3, MMseqs2) on the same fifteen elite cores corresponding to ranks 1–15 in this file, as an orthogonal structural cross-check. ColabFold iptm/ptm are reported for that panel only and are not numerically equivalent to Boltz ipTM.

We are sharing this panel publicly so expert and community reviewers can assess both sequence choices and an independent folding pass on the same elite subset—consistent with a multi-angle, reproducible design story rather than reliance on a single score from a single network.