https://www.nature.com/articles/s41586-026-10670-w

Zero-shot design of drug-binding proteins via neural iterative selection-expansion

Key figures

• [Figure 1]: Defines sequence-structure-ligand self-consistency and the NISE loop as alternating LASErMPNN sequence expansion with neural co-structure prediction-based selection.
  
• [Figure 2]: Shows the LASErMPNN architecture and validates ligand-aware sequence design on held-out binding-site recovery, streptavidin-biotin, and the rucaparib-binding PiB benchmark.
  
• [Figure 3]: Demonstrates exatecan binder design by NISE, with EPIC binding exatecan at 0.12 µM and outperforming both HSA and the best traditional COMBS-Rosetta design.
  
• [Figure 4]: Shows computation-only affinity maturation of EPIC by neural proofreading, improving exatecan affinity to 1.2 nM in the Q51N/M97L double mutant and validating EPIC/EPIC Q51N by X-ray crystallography.
  
• [Figure 5]: Demonstrates that EPIC and EPIC Q51N/M97L protect exatecan from lactone hydrolysis, with the optimized mutant preserving the ring-closed form for at least 50 h in PBS.
  
• [Figure 6]: Extends NISE to NTF2 scaffolds and apixaban, producing APEX with 80 pM affinity and no appreciable binding to exatecan.
  
• [Extended Data Figure 7]: Shows a metric caveat: Boltz-2 ligand pLDDT weakly tracks relative affinity, whereas P(bind) is better treated as a binder-versus-nonbinder classifier.
  
• [Supplementary Figure 36]: Shows a predictor-dependence caveat: experimentally verified apixaban binders would have failed RFAA-based self-consistency and ligand pLDDT filters.
  

1) Thesis (one sentence)

To address the lack of zero-shot de novo methods for designing high-affinity small-molecule-binding proteins, in de novo four-helix-bundle and NTF2 scaffolds targeting exatecan and apixaban, NISE causes high-rate nanomolar-to-picomolar binder discovery and exatecan lactone stabilization by iteratively optimizing sequence-structure-ligand self-consistency with LASErMPNN sequence generation and RFAA/Boltz co-structure prediction, supported by fluorescence anisotropy binding assays, X-ray crystallography, SEC, CD, hydrolysis spectroscopy, and computational benchmarking.

2) Evidence card (three bullets only)

• Strongest result: (Fig. 3; Fig. 4; Fig. 5; Fig. 6; Extended Data Fig. 9) NISE achieved 100% exatecan-design success and 83% apixaban-design success, produced EPIC at 0.12 µM, improved EPIC to 1.2 nM by Q51N/M97L neural proofreading, preserved >99% ring-closed exatecan for at least 50 h, and produced APEX at 80 pM.
  
• Method enabler: (Fig. 1; Fig. 2; Supplementary Fig. 11; Supplementary Fig. 32; computational protein design + tools including LASErMPNN, RFAA, Boltz-2, COMBS/CARPdock, ProteinMPNN, RFdiffusion, Rosetta, AF2/AF2-multimer, RF3, SEC, fluorescence anisotropy, absorbance spectroscopy, and X-ray crystallography) NISE couples ligand-aware autoregressive sequence/rotamer design to co-structure prediction, then selects designs by ligand confidence, protein-ligand self-consistency, pocket preorganization, BUNs, and oligomerization counterscreens.
  
• Critical limitation: (Extended Data Fig. 7; Supplementary Fig. 36; Fig. 6) NISE success depends strongly on the co-structure predictor and metric calibration: RFAA would have rejected the experimentally active apixaban binders, Boltz-2 P(bind) did not rank relative affinity well, and the picomolar APEX result lacks an experimental co-structure validating the predicted binding mode.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “Current deep-learning algorithms have struggled to navigate this landscape” (Abstract, page 1)
  
• Quote 2: “Tight coupling between two generative neural networks unlocks the zero-shot design of small-molecule binders” (Discussion, page 11)
  
• Quote 3: “NISE is agnostic to the specific networks used” (Discussion, page 11)
  
• Quote 4: “We found that positive design—focusing on moulding the binding pocket only to the target ligand—was sufficient” (Discussion, page 11)
  

Key comparisons (1–3 lines)

• Compared to: COMBS-Rosetta small-molecule binder design, LigandMPNN-Rosetta apixaban design, HSA exatecan binding, and native factor Xa apixaban binding.
  
• Win: Achieves far higher hit rate and affinity from small experimental test sets: EPIC is ~70-fold tighter than the best traditional exatecan design, and APEX is nearly 10,000-fold tighter than the prior NTF2 apixaban design.
  
• Tradeoff: Requires high-quality starting scaffolds, plausible docked ligand poses, compatible co-structure predictors, and downstream empirical validation because confidence metrics are not direct affinity, kinetics, or selectivity predictors.
  

Methods I might copy (protocol hooks)

• Construct design / Models: For exatecan, generate 40 single-chain four-helix bundles from 750 four-chain coiled coils, each helix 40 residues, connect helices with RFdiffusion loops totaling 36 residues for 148-aa proteins, design 16 sequences per backbone with ProteinMPNN v_48_020 at T=0.15 with no cysteine, predict with AF2 model 4 using 1 recycle in single-sequence mode, and filter for average Cα pLDDT >0.9, average pAE <5.0, Cα RMSD <1.0 Å, minimum nonterminal Cα pLDDT >0.85, and 0.75 Å Cα RMSD clustering. For exatecan NISE, use 1000 LASErMPNN sequences per protein-ligand pose, RFAA with 15 recycles, select by protein Cα pLDDT >0.8 and backbone/ligand self-consistency RMSD <1.5 Å, run 35 iterations, pool 100,000 designs, then filter by protein Cα RMSD <1.5 Å, ligand heavy-atom RMSD <1.0 Å, ligand pLDDT >0.8, protein Cα pLDDT >0.8, Rosetta packstat >0.55, Rosetta-minimized RMSD <1.5 Å, zero apo-pocket ligand clashes within 2 Å, no cysteines, no buried charged residues, and BUNs score = 2x ligand BUNs + 1x protein BUNs. For apixaban, start from 50 NTF2 backbones, extract two apixaban conformers from PDB 6W70 chains A and C, generate 100 CARPdock poses per conformer per backbone, use a 0.5 Å translation grid, 1000 random ligand rotations per grid point, 2.5 Å clash cutoff, and K=100 clustering, then design 3 LASErMPNN sequences per 10,000 poses and predict with Boltz-2.
  
• Conditions / Instruments: Express ordered IDT genes from peAIP32 with 6xHis-Cth Sumo in BL21 E. coli, grow 0.5 L LB/kanamycin with 34.7 mM lactose for autoinduction at 30 °C and 220 rpm for at least 16 h, lyse in PBS + 500 mM NaCl by sonication on ice, centrifuge at 16,000 rpm/11,600g for 30 min, bind to 1.5 mL magnetic Ni-NTA beads per 0.5 L culture for at least 1 h at 4 °C, wash with 15 mL PBS + 500 mM NaCl + 20 mM imidazole and 10 mL PBS, cleave with 1 µM Cth protease overnight at 4 °C, and store purified protein in PBS at 4 °C. Assess oligomeric state by BioRad NGC with BioRad ENrich 650 SEC. For EPIC CD, use 0.1 mg/mL protein in PBS, 0.1 cm quartz cuvette, Jasco J-1500, 25 °C, 195-260 nm scans, 1 nm bandwidth, 50 nm/min scan speed, 0.1 nm pitch, 1 s integration, 5 accumulations, and thermal spectra from 25 °C to 95 °C in 5 °C intervals at 10 °C/min with 3 accumulations.
  
• Readout / Analysis: For fluorescence anisotropy, use 384-well plates, 1x PBS pH 7.4 + 0.1% w/v PEG-8000, 50 nM ligand for exatecan/camptothecin-class screening, 360 nm excitation and 440 nm emission for camptothecins, 482 nm excitation with 16 nm bandwidth and 530 nm emission with 40 nm bandwidth for FITC ligands, BMG LabTech PHERAstar FSX or CLARIOstar Plus, gain adjusted so free ligand is 35 mP, and quadratic single-site binding fits. For EPIC mutant global fits, titrate protein across 25, 12.5, and 5 nM exatecan and bootstrap 1000 residual-resampled datasets. For APEX, titrate protein across 25, 10, and 5 nM apixaban-FITC, use 70-90% bound anisotropy conditions for competition with unlabeled apixaban, incubate apixaban-FITC/protein for 3 h before competitor addition, seal and equilibrate overnight at room temperature, and fit apixaban-FITC and apixaban globally. For hydrolysis, monitor 20 µM exatecan with 20 µM protein in 2 mL PBS pH 7.4 with <0.54% DMSO using an Agilent 8453 G1103A, 1 cm quartz cuvette, 190-900 nm spectra every 30 min, room temperature, and at least two biological replicates.
  

Open questions / Theoretical implications (2–5 bullets)

• Can NISE be generalized from binary protein-drug complexes to protein-ligand-protein ternary complexes, where productive geometry rather than simple ligand burial is the key design objective?
  
• Can negative design be incorporated prospectively by selecting for high on-target ligand pLDDT/P(bind) and low off-target ligand pLDDT/P(bind) across close analogs, metabolites, and endogenous binders?
  
• Can NISE handle ligands whose active binding state involves covalent reaction, metal coordination, protonation shifts, or conformationally gated protein surfaces?
  
• Can confidence metrics be calibrated to predict residence time, kinetic selectivity, intracellular stability, and functional potency rather than equilibrium binding alone?
  
• Can bespoke generative scaffolds conditioned directly on ligand and target geometry outperform precomputed four-helix-bundle or NTF2 scaffold libraries?