https://www.nature.com/articles/s41467-023-38328-5

Improving de novo protein binder design with deep learning

Key figures

• [Figure 1]: Separates de novo binder failure into monomer-folding and target-interface failure modes, then shows AF2/RF2 confidence metrics retrospectively distinguish binders from non-binders.
  
• [Figure 2]: Demonstrates prospective enrichment of functional binders on ALK, LTK, IL-10Rα, and IL-2Rα, and shows ProteinMPNN-FastRelax improves computational efficiency over Rosetta sequence design.
  

1) Thesis (one sentence)

To address the low experimental success rate of target-structure-only de novo protein binder design, in retrospective Cao et al. design libraries and prospective ALK, LTK, IL-10Rα, and IL-2Rα minibinder campaigns, deep-learning structure-prediction filtering and ProteinMPNN-based sequence design cause nearly tenfold higher binder recovery and higher compute efficiency by using AF2/RF2 pLDDT, complex RMSD, and pAE_interaction to reject monomer-folding and interface-formation failures before synthesis, supported by yeast-surface-display sorting, deep sequencing, BLI confirmation, and computational benchmarking.

2) Evidence card (three bullets only)

• Strongest result: (Fig. 2) AF2-filtered prospective libraries produced higher binder success rates for all four successful targets, including 8-fold and 30-fold gains for LTK and IL-2Rα Site 1, and recovered ALK and IL-10Rα binders where physically based filtering recovered none.
  
• Method enabler: (Fig. 1; Fig. 2; computational protein design + experimental validation using Rosetta, RifDock, AF2 initial guess, RF2, DAN, ProteinMPNN, Rosetta FastRelax, yeast surface display, FACS/deep sequencing, and BLI) AF2/RF2 monomer pLDDT and complex pAE_interaction separated folding and binding failures, while ProteinMPNN-FastRelax increased the rate of generating designs passing AF2 interface filters.
  
• Critical limitation: (Fig. 2) Even after deep-learning filtering, target-dependent success rates remained below 1%, no binders were recovered for IL-2Rα Site 2, and initial binders were still generally high-nanomolar rather than uniformly high-affinity.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “There is, however, considerable room for improvement as the overall design success rate is low.” (Abstract, page 1)
  
• Quote 2: “We show retrospectively and prospectively that this improved protocol has nearly 10-fold higher success rate than the original energy-based method.” (Results, page 2)
  
• Quote 3: “success rates among the targets remain low (<1%)” (Discussion, page 6)
  
• Quote 4: “We find further that sequence design using ProteinMPNN rather than Rosetta considerably increases computational efficiency.” (Abstract, page 1)
  

Key comparisons (1–3 lines)

• Compared to: Rosetta-only physically based binder design using monomer energy, Rosetta ddG, target_delta_sap, contact_patch, and contact_molec_sq5_apap_target filters.
  
• Win: AF2/RF2 filtering detects both wrong-fold and wrong-interface designs before synthesis, and ProteinMPNN-FastRelax gives about an 8-fold per-target efficiency improvement over Rosetta design.
  
• Tradeoff: AF2 filtering is computationally expensive, only about 2.3% of designs pass pAE_interaction filtering, and success remains strongly epitope-dependent.
  

Methods I might copy (protocol hooks)

• Construct design / Models: Run target-structure-only binder design from IL2RA structures 1Z92, 2B5I, 3NFP, and 2ERJ; IL10RA structure 1LQS; ALK structure 7NWZ; and LTK structure 7NX0. Generate approximately 10 million RifDock outputs per target, FastDesign about 500,000, extract about 6,000 motifs, graft up to 10 million docks, FastDesign another 500,000, then AF2-predict about 1 million designs per target. Select Rosetta controls by pareto-front filtering on target_delta_sap, ddG, contact_patch, and contact_molec_sq5_apap_target; select AF2-filtered designs with pAE_interaction <10 and af2_complex_rmsd <5 Å; redesign AF2-predicted interfaces with Rosetta or ProteinMPNN. For ProteinMPNN-FastRelax, mask the binder sequence in a protein complex, design the binder sequence with ProteinMPNN, thread it onto the binder backbone, relax the complex with Rosetta FastRelax, and iterate.
  
• Conditions / Instruments: Pad designs to 65 amino acids with a C-terminal (S)n linker, reverse-translate and codon-optimize for Saccharomyces cerevisiae using DNAWorks2.0, add N-terminal adapter GGTGGATCAGGAGGTTCG and C-terminal adapter GGAAGCGGTGGAAGTGG, synthesize oligo libraries through Agilent, and transform into EBY100 yeast. Grow yeast in C-Trp-Ura plus 2% glucose, induce in SGCAA plus 0.2% glucose at 1 × 10^7 cells/mL for 16-24 h at 30°C, stain with biotinylated target, anti-c-Myc FITC, and SAPE, and sort on a Sony SH800S. Express selected binders with N-terminal 8-His tags and TEV sites in modified pET-29b(+) BL21(DE3), then purify by IMAC and Superdex 75 10/300 GL SEC. For BLI, use Octet Red96 or Octet R8 with streptavidin tips, HBS-EP+ plus 1% BSA, load 30-50 nM biotinylated target, baseline 120 s, associate with 500 nM analyte or 1000 nM for IL-10Rα for 600 s, and dissociate for 1000 s.
  
• Readout / Analysis: Use yeast-surface-display sort enrichment plus deep sequencing to estimate SC50, defining designs with SC50 <4 µM as successes. Confirm yeast-display hits by single-concentration BLI. Use AF2/RF2 monomer Cα RMSD and pLDDT to flag Type I wrong-fold failures; use complex Cα RMSD and pAE_interaction to flag Type II wrong-interface failures. For computational efficiency, calculate designs passing pAE_interaction <10 per CPU-second equivalent, using 100 CPU-s as 1 GPU-s.
  

Open questions / Theoretical implications (2–5 bullets)

• Can pAE_interaction be calibrated prospectively for engineered presenter-target interfaces rather than conventional binary minibinders?
  
• Can AF2/RF2 distinguish productive ligand-dependent interfaces from high-confidence but biologically undesired constitutive binding?
  
• How much of the remaining failure rate comes from interface energetics versus inaccessible or conformationally heterogeneous epitopes?
  
• Can ProteinMPNN-FastRelax be constrained to preserve an existing functional pocket while redesigning only a target-recognition surface?
  
• Will this pipeline remain predictive for nucleotide-state-dependent targets whose switch regions move substantially between biochemical states?