RBX1 RING Domain Binder Design Report

Computational design of protein binders targeting RBX1 using PC, G, AR Monte Carlo, and GA pipelines

Target: RBX1 RING Domain

PDB: 2LGV

UniProt: P62877

Residues: 12–108 (97 aa)

Function: E3 ubiquitin ligase (Cullin-RING)

Hotspot residues highlighted in red: D40, R46, D51, E55, E66, E67, W72, R86, K89

Method Comparison Summary

PC (Method 1)

Best AF3 ipTM: 0.510

Best AF3 Ranking: 0.620

Best iPSAE: 0.309 (binder_4)

Design type: General protein binders (60–120 aa)

G (Method 2)

Best Cofold ipTM (C1): 0.759

Best Halluc ipTM (AF-M): 0.767

Design type: VHH nanobodies (~130 aa)

AR (Method 3) — Best Result

Best iPSAE: 0.650 (77aa binder)

Best ipTM: 0.870

Method: Monte Carlo + AF3 oracle

Compute: 30 rounds × 8 mutations

Starting from PC binder_4 (iPSAE=0.309)

GA (Method 4)

Best iPSAE: 0.594 (131aa VHH)

Best ipTM: 0.810

Method: Monte Carlo + AF3 oracle on VHH

Starting from G top VHH (iPSAE=0.594)

Note: G designs are VHH nanobodies (~130 aa), PC designs are general protein binders (60–120 aa). Validation methods differ (AF3 vs C1). iPSAE for PC binders was computed from C1 cofold, not AF3. AR used Monte Carlo optimization with AF3 as oracle starting from the best PC binder.

Method 1: PC Pipeline

PC

Flow matching generative model produces backbone structures conditioned on the RBX1 target. 200 samples generated with AF2 reward scoring.

200 backbones

MPNN

Inverse folding model designs sequences for top backbones. 8 sequences sampled per backbone.

80 sequences

AF3 Validation

Predict complex structures with AF3. Evaluate via ipTM, pLDDT, iPAE, and ranking score.

9 validated

Method 2: G Pipeline

AF-M Hallucination

AF-M hallucination generates VHH nanobody backbones and sequences targeting RBX1 hotspot residues.

Hallucinated nanobodies

AbMPNN Redesign

AbMPNN redesigns CDR loops and framework residues for improved stability and binding.

5 variants per seed

C1 Cofold Validation

Independent structure prediction with C1 cofold. Evaluate binding via ipTM, pLDDT, iPAE, and hotspot contacts.

20 validated

AR (Method 3) Monte Carlo Optimization — Top 15 Results

Starting from the best PC binder (iPSAE=0.309), we ran Monte Carlo sequence optimization using AF3 as oracle. 30 rounds × 8 mutations = 240 AF3 evaluations.

	Rank	Sequence	Length	iPSAE	ipTM	pLDDT	iPAE	Actions	Structure

iPSAE > 0.6 / ipTM > 0.8 iPSAE > 0.5 / ipTM > 0.7 Below threshold

Binder Candidates — AF3 Validation Results

Click any column header to sort. Binder_4 (highlighted) shows the best AF3 metrics overall.

	Rank	Binder Name	Length	PC Reward	PC ipTM	PC pLDDT	MPNN Score	AF3 ipTM	AF3 pTM	AF3 pLDDT	AF3 iPAE	iPSAE	AF3 Ranking	Actions

Good (ipTM>0.5, pLDDT>70, iPAE<5) Moderate Poor

G (Method 2) VHH Nanobody Candidates — C1 Cofold Validation

Click any column header to sort. Top candidate (highlighted) shows the best C1 cofold ipTM. All designs are VHH nanobodies (~130 aa).

	Rank	Name	Length	Halluc ipTM	Halluc pLDDT	Cofold ipTM	Cofold pLDDT	Cofold iPAE	iPSAE	Hotspot Contacts	Actions

Cofold ipTM > 0.7 / pLDDT > 80 Cofold ipTM > 0.5 / pLDDT > 70 Below threshold

GA (Method 4) Optimization — Top 8 Results

Starting from the best G VHH nanobody (iPSAE=0.594), we ran Monte Carlo sequence optimization using AF3 as oracle. All designs are 131 aa VHH nanobodies.

	Rank	Sequence	Length	iPSAE	ipTM	pLDDT	iPAE	Actions	Structure