RBX1 is deeply conserved across eukaryotes. The RING-H2 domain (residues ~27–104) shows uniformly high conservation (>0.7 for nearly all positions), with zinc-coordinating cysteines and histidines reaching near-perfect conservation (0.91–1.0). The N-terminal tail (residues 1–20) is more variable, consistent with it being unstructured in the NMR ensemble (2LGV).

The Cullin face is significantly more conserved than the E2 face (mean 0.938 vs 0.839). This makes sense: the Cullin interaction is constitutive (RBX1 is always bound to a Cullin scaffold in vivo), whereas the E2 interface cycles through multiple E2 partners. The higher conservation at the Cullin face means binders targeting it are more likely to disrupt a functionally critical interaction, but the surface may also be harder to compete with due to the tight, conserved binding.

Three perfectly conserved residues stand out: W33, W35 (both Cullin face), G73 (Cullin face), and R46 (E2 face). These are absolutely invariant across all 165 sequences in the MSA. W33 and W35 form a tryptophan pair that likely stacks against the Cullin surface—a classic hot-spot motif. R46 is the catalytic arginine critical for E2 activation.

Gap fraction is low (<2%) for the core domain (residues 36–105), meaning the MSA is well-aligned in the structured region. The N/C-terminal tails show higher gaps (10–27%), reflecting length variation among homologs. This gives confidence that the conservation scores in the core are reliable.

Design implications: The 60/40 E2/Cullin split is well-justified. The Cullin face offers a tighter, more conserved target with multiple tryptophan hot-spots, favoring high-affinity designs. The E2 face is more diverse, potentially offering more epitope options but requiring designs that can outcompete the native E2 partners. The zinc sites should be included in the target structure but not directly targeted—they are buried and structurally critical, not surface-accessible.

ESM-2 and evolution strongly agree on which residues are critical. The Spearman correlation of 0.581 (p < 10^-10) and Pearson of 0.714 (p < 10^-17) between ESM-2 sensitivity and Shannon entropy conservation confirm that the protein language model has learned genuine structural and functional constraints from sequence alone.

Zinc-coordinating cysteines dominate the sensitivity landscape. C42, C56, C83, C75, C53, C68, and C94 are all among the top 15 most sensitive positions. Any mutation at these sites is catastrophic (dLLR < -10), consistent with their role as structural zinc ligands that maintain the RING-H2 fold. The most deleterious single mutation in the entire protein is I54W (dLLR = -18.25), a massive tryptophan insertion into the hydrophobic core adjacent to C53 and C56.

The Cullin face has the single most sensitive residue (D36, sensitivity = 13.0), which also has perfect evolutionary conservation (0.945). This aspartate likely forms critical salt bridges in the Cullin interaction. For binder design, targeting residues around D36 could be highly effective at disrupting the complex.

ESM-2 identifies some positions as sensitive that conservation misses. I54 (ESM sensitivity rank 5, conservation only 0.786) and I84 (rank 11, conservation 0.802) are moderately conserved but ESM predicts they are among the most intolerant of mutations — likely because they play critical roles in hydrophobic packing that the MSA alone doesn't fully capture.

Design implications: Binders should maximize contacts with high-sensitivity residues (especially D36, C42, R46, F79, W87) since these positions cannot easily mutate to escape binding. The DMS data also suggests that the N-terminal tail (residues 1–20) has low mutation sensitivity, confirming it is a poor target for binder design.

1. BoltzGen designs validate significantly better than RFdiffusion+MPNN designs in AF3, despite having lower Boltz-2 scores. This is the single strongest predictor of AF3 success.

2. Sequence diversity matters: designs with higher Shannon entropy (more diverse AA usage) and more unique amino acids tend to validate better. Low-complexity sequences rich in alanine and glycine consistently fail.

3. Aromatic residues (F, W, Y) are enriched in successful designs — these contribute to specific hydrophobic contacts at the interface that AF3 can validate.

4. High Boltz-2 ipTM is NOT predictive of AF3 success — in fact, it may be slightly anti-correlated. Designs with Boltz ipTM > 0.95 mostly fail AF3 validation (delta > 0.7).

5. Cullin-face designs show better AF3 agreement than E2-face designs, though the sample size is small. The one Cullin design tested (design_0357) has AF3 ipTM = 0.78 with low delta.

6. For future campaigns: prioritize BoltzGen-style generation, increase aromatic content in ProteinMPNN sampling, and reduce alanine/glycine bias. Consider Cullin-face targeting which appears more AF3-compatible.