https://iubmb.onlinelibrary.wiley.com/doi/full/10.1002/iub.70108

Single-Sequence Deep Learning Delivers Crystal-Quality Models of Covalent K-Ras G12 Hotspot Complexes

Key figures

• [Figure 1]: Defines the ligand benchmark across K-Ras(G12C), K-Ras(G12D), and K-Ras(G12S), including acrylamide, beta-lactone, and noncanonical hotspot chemistries.
  
• [Figure 2]: Shows crystal-quality Chai-1 predictions for diverse K-Ras(G12C) Switch-II-pocket inhibitors, including blind Divarasib prediction at 0.56 Å pocket-aligned RMSD.
  
• [Figure 3]: Demonstrates that an explicit covalent restraint rescues K-Ras(G12D)-G12Di-1 pose prediction from ligand ejection to 0.68 Å pocket-aligned RMSD.
  
• [Figure 4]: Shows that covalent restraint positions the K-Ras(G12S)-G12Si-5 scaffold but fails to preserve piperidine stereochemistry.
  
• [Figure 5]: Quantifies Chai-1 versus AlphaFold3 performance, showing comparable pose accuracy but approximately 37-fold to 40-fold higher Chai-1 throughput.
  
• [Figure 6]: Reveals failure on low-affinity alternative-pocket fragments Cpd3 and Cpd4, which Chai-1 misplaces into Switch-II-pocket-like poses.
  
• [Table S2]: Quantifies a negative-control-like caveat: Chai-1 preserves plausible Divarasib pocket occupancy in WT K-Ras and resistance-associated K-Ras(G12C) variants without covalent restraint.
  

1) Thesis (one sentence)

To address the structural bottleneck in covalent K-Ras ligand optimization beyond readily crystallized cysteine adducts, in covalent K-Ras G12C, G12D, and G12S ligand complexes, Chai-1 single-sequence cofolding with optional covalent-bond restraints causes crystal-quality pose prediction and faster model generation than AlphaFold3 by accepting user-defined ligand SMILES, modeling protein-ligand complexes without MSA/template preprocessing, and enforcing post-reaction covalent geometry when needed, supported by pocket-aligned RMSD benchmarking against X-ray structures, pLDDT/confidence assessment, runtime comparison, and negative-control-like variant analyses.

2) Evidence card (three bullets only)

• Strongest result: (Fig. 2; Fig. 3; Fig. 5) Chai-1 predicted chemically diverse K-Ras(G12C) inhibitor poses within the 2 Å benchmark cutoff, achieved 0.56 Å RMSD for blind K-Ras(G12C)-Divarasib, restored K-Ras(G12D)-G12Di-1 from ligand ejection to 0.68 Å RMSD with covalent restraint, and generated about 1000 models per A40 GPU day versus fewer than 30 for AlphaFold3.
  
• Method enabler: (Fig. 1; Fig. 5; Table S1; computational structure-prediction benchmarking using Chai-1 web/local single-sequence mode, user-defined ligand SMILES, GDP/Mg2+ or GMPPNP/Mg2+ cofactors, optional covalent restraints, AlphaFold3 comparison, pLDDT, pocket-aligned RMSD, and runtime profiling) The workflow benchmarks a matched set of crystallographic covalent K-Ras hotspot complexes and separates restraint-free G12C pose prediction from covalently restrained G12D/G12S adduct modeling.
  
• Critical limitation: (Fig. 3c; Fig. 4b; Fig. 6; Table S2) Chai-1 can preserve plausible but misleading K-Ras pocket occupancy in WT or resistance-mutant backgrounds, misplace weak alternative-pocket fragments into the Switch-II pocket, and generate chemically imperfect covalent adducts with strained bond angles, planarized sp3 geometry, or inverted stereochemistry.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “This makes it straightforward to move from sequence and SMILES to docking.” (Introduction, page 2)
  
• Quote 2: “We therefore treat apparent covalent proximity only as a hypothesis-generating geometric feature, not as standalone evidence of binding or covalent reactivity.” (Results 2.1, page 3)
  
• Quote 3: “Thus, model confidence and pocket occupancy should not be interpreted as evidence of potency, selectivity, or resistance sensitivity.” (Discussion, page 8)
  
• Quote 4: “We therefore view Chai-1 primarily as a structural hypothesis generator rather than a standalone medicinal chemistry ranking engine.” (Discussion, page 8)
  

Key comparisons (1–3 lines)

• Compared to: X-ray co-crystallography, AlphaFold3 with MSAs/templates, conventional covalent docking, and covalent modeling without explicit adduct restraints.
  
• Win: Chai-1 is accessible, scriptable, accepts user-defined ligands, supports covalent restraints locally, and gives near-AlphaFold3 pose accuracy with much higher throughput for mutant/ligand panels.
  
• Tradeoff: It does not predict potency, selectivity, reactivity, resistance sensitivity, leaving-group chemistry, or stereochemical fidelity reliably enough to replace biochemical and structural validation.
  

Methods I might copy (protocol hooks)

• Construct design / Models: Use K-Ras complexes with GDP and Mg2+ for most G12C/G12D/G12S structures; use BBO-8520 with GDP and GMPPNP active-state structures; benchmark against PDB 5F2E, 5V9U, 6OIM, 6UT0, 9DMM, 7R0M, 8V3A, 8V39, 8T4V, 7TLG, 7A1W, and 7A47; for negative-control-like tests, model WT K-Ras and 13 K-Ras(G12C) secondary Switch-II-pocket variants at R68, M72, H95, Y96, and Q99 with Divarasib, GDP, and Mg2+.
  
• Conditions / Instruments: Submit Chai-1 web jobs in single-sequence mode without MSAs, templates, or restraints; run local Chai-1 and AlphaFold3 on UCSF Wynton HPC qb3-atgpu queue using NVIDIA A40 32 GB GPUs; prepare ligands from post-reaction SMILES with leaving group removed; model G12C acrylamides as propanamides and G12Di-1/G12Si-5 in post-reaction forms; compare AlphaFold3 with MSA/template inputs against Chai-1 single-sequence mode.
  
• Readout / Analysis: Calculate pocket-aligned RMSD by defining the pocket as all reference heavy atoms within 10 Å of ligand heavy atoms, aligning predicted and ground-truth structures by pocket Cα atoms, and computing ligand heavy-atom RMSD with RDKit/PyMOL; assess confidence with pLDDT; compare Chai-1 and AlphaFold3 by Wilcoxon matched-pairs signed-rank test; quantify restraint-free pocket occupancy using centroid distance to residues 12, 68, 72, 95, 96, and 99, ligand atom contacts within 4.0 Å of those residues, and ligand RMSD variability across ranks 0-4.
  

Open questions / Theoretical implications (2–5 bullets)

• Can single-sequence cofolding generalize from Ras-family Switch-II pockets to less characterized Rho, Rab, Arf, and Ran surfaces without memorized ligand-pocket priors?
  
• Can covalent restraints be made chemically aware enough to handle serine, aspartate, arginine, and other non-cysteine warheads without distorted adduct geometry?
  
• Can Chai-1-like outputs be integrated with reactivity, transition-state, and off-target models to predict covalent kinetics rather than only post-reaction pose plausibility?
  
• Can negative-control-like variants, resistance mutations, and alternative-pocket ligands be incorporated as routine counterscreens for model hallucination?
  
• Can fast cofolding workflows be extended from binary covalent adducts to ligand-induced ternary complexes involving a recruited presenter protein?