Elix, CBI2026, Oral Presentation, Synthesizer: Extending Vina-GPU with Core-Constrained and Covalent Docking

Transcript

1. Any unauthorized reproduction of this material is strictly prohibited. Extending Vina-GPU with Core Constrained and Covalent Docking CBI Annual Meeting 2026 Loic Beheshti, Research Engineer @ Elix, Inc. Nazim Medzhidov, PhD, Research Engineer @ Elix, Inc. Tasuku Ishida, PhD, CSO @ Elix, Inc. Masakazu Atobe, PhD, Medicinal Chemist @ Elix, Inc. Copyright © Elix, Inc. All rights reserved
2. Two Capability Gaps in Vina-GPU Workflows Goal: Bring core-constrained and covalent docking to GPU-parallel Vina search while retaining high screening throughput. Standard GPU Docking Baseline Capability Fast GPU-parallel search Current Status Fully functional baseline for standard virtual screening. 1. Core-Constrained 2. Covalent Docking Objective Objective Keep a known substructure fixed in space during sampling Model considers ligand-residue attachment during sampling. Capability Gap 1 Not natively supported in stock AutoDock Vina / Vina-GPU[1]. Capability Gap 2 No standard Vina-GPU workflow available. [1] AutoDock Vina / Vina-GPU 2.1 documentation Copyright © Elix, Inc. All rights reserved
3. Our approach: Extend QuickVina2-GPU 2.1 Why this engine? • • • What this enables High-throughput virtual screening Iterative molecular design Fastest average speed among Vina-GPU 2.1 variants[4] 1. 2. 3. Retain GPU-parallel search Support practical medchem workflows Build a unified screening workflow Core-constrained docking Anchor a known ligand substructure while sampling the remainder. QuickVina2-GPU 2.1 Covalent docking Model explicit ligand-residue attachment. Existing Covalent alternatives Contribution GNINA: CPU-based covalent Vina-family docking[2] AutoDock-GPU: GPU covalent AutoDock4-based[3] A fast GPU docking workflow expanded to two important use cases while preserving high screening throughput. [2] GNINA 1.3 (2025); [3] AutoDock-GPU / Meeko documentation; [4] Tang et al. (2024), Vina-GPU 2.1 Copyright © Elix, Inc. All rights reserved
4. FIRST CAPABILITY Core-constrained docking Implementation in QuickVina2-GPU 2.1 Core-constrained (template-guided) docking restrains a selected ligand substructure to predefined 3D coordinates while allowing the remaining ligand atoms to be sampled within the binding pocket. Copyright © Elix, Inc. All rights reserved
5. Vina Optimization: Moveset Global translation Global rotation Bond torsion R t Move τj ROOT Origin Atoms affected Geometric effect Translation t Entire ligand Rigid-body displacement in the receptor frame Rotation R Entire ligand Rigid-body orientation about the ROOT frame Torsion τj Descendant subtree Rotation around a bond axis; upstream group unchanged Copyright © Elix, Inc. All rights reserved
6. Core-Constrained Docking: Reroot and Restrain Pre-align the core → Reroot its rigid group Before: ROOT in group A Group A Group B τ₁ ROOT = A After: ROOT contains selected atoms Group C Group A τ₂ Group B τ₁ C B τ₁ moves B + C, including the selected atoms A Group C τ₂ ROOT = B τ₁ moves A; τ₂ moves C—not the core Teal = two restrained atoms; blue boundary = full ROOT; cross = its origin; orange = torsion axis. → Apply positional restraints Er = ½ k ∑ᵢ ‖rᵢ − rᵢ⁰‖² rᵢ⁰ = reference coordinates; k = restraint stiffness. Copyright © Elix, Inc. All rights reserved C
7. Two benchmarks: Pose Recovery Focus BACE1[5]: 22 native complexes Astex Diverse Set[6]: 83 supported complexes Scaffold-guided test: 14-heavy-atom scaffold target. Minimal-guidance test: one selected atom. CCD: constrained RDKit embedding + UFF refinement. Conventional: independent ETKDGv3 + MMFF preparation. Both workflows share the same freshly generated ETKDGv3 + MMFF conformer. Search / evaluation Global QuickVina2-GPU 2.1 Output and recovery Harmonic CCD variants Shared protocol 5,000 lanes · depth 20 · seed 42 · final host refinement Up to 10 distinct poses; evaluate rank 1 and best of first ≤3 k=1000 and k=5000; selected atoms only. Native receptor and reference pose per complex. BACE1 retains its original box and receptor preparation. Results are evaluated separately. [5] Berman et al., Nucleic Acids Res. 2000, [6] Hartshorn et al., J. Med. Chem. 2007, Copyright © Elix, Inc. All rights reserved
8. BACE1: Larger-Scaffold Pose Recovery Top pose ≤2 Å Best of up to 3 ≤2 Å CCD top pose closer 50% → 100% 68.2% → 100% 21 / 22 cases Workflow Conventional CCD k=1000 Top pose Best of ≤3 Top RMSD Top RMSD Core RMSD Core RMSD vs ref. ≤2 Å ≤2 Å median (Å) mean ± SD (Å) median (Å) mean ± SD (Å) 11/22 (50.0%) 15/22 (68.2%) 2.132 3.237 ± 2.701 1.921 2.690 ± 2.132 22/22 22/22 (100.0%) (100.0%) 0.726 0.708 ± 0.456 0.048 0.048 ± 0.009 Rank-1 restraint-target drift: k=1000, 0.0015 ± 0.0011 Å; k=5000, 0.0003 ± 0.0003 Å. Core RMSD in the table is measured against the crystal reference. Copyright © Elix, Inc. All rights reserved
9. Astex Diverse Set: Single-Atom Pose Recovery Top pose ≤2 Å Best of up to 3 ≤2 Å CCD top pose closer 62.7% → 86.7% 74.7% → 91.6% 81.9% cases Median Top pose Best ≤3 ≤2 Å ≤2 Å RMSD (Å) Conventional 52/83 (62.7%) 62/83 (74.7%) CCD k=5000 72/83 (86.7%) 76/83 (91.6%) Workflow Top RMSD (Å) Atom displacement mean ± SD (Å), mean ± SD 1.022 2.421 ± 2.430 3.2129 ± 2.9420 0.699 1.208 ± 1.460 0.0004 ± 0.0007 Best ≤3: minimum RMSD in the first 3 outputs (or fewer). SD across 83 complexes; symmetry-corrected RMSD without fitting. Max CCD atom drift: 0.0117 Å. Copyright © Elix, Inc. All rights reserved
10. Visual Inspection Copyright © Elix, Inc. All rights reserved
11. SECOND CAPABILITY Covalent docking Integration to QuickVina2-GPU 2.1 Covalent docking predicts binding poses for ligands that form a covalent bond with a target protein residue. Here, we focus on cysteine-targeting acrylamides and their Michaeladdition products. Copyright © Elix, Inc. All rights reserved
12. Covalent Docking: Tethered and Reactive Approaches Free docking protein S Cys Sγ Tethered docking C ligand reactive C Unreacted ligand and residue. No covalent attachment or reactive constraint. protein S Cys Sγ C Reactive docking ligand reactive C Explicit covalent product bond. Optimize the remaining flexible degrees of freedom. [2] GNINA covalent docking documentation; [3] AutoDock-GPU / Meeko tethered and reactive docking tutorials. Copyright © Elix, Inc. All rights reserved protein S Cys Sγ C ligand reactive C Unbonded ligand and residue. Specialized reactive-pair interaction potential.
13. Covalent Docking: Fixed-ROOT Strategy Fresh product Attach known Cys Generate a product conformer with RDKit. Construct the Cys–ligand adduct; best results with Open Babel. Reroot and place GPU search Align and fix the Cys backbone / Cβ ROOT. Optimize downstream ligand torsions. O C N O Fixed ROOT ≠ strong harmonic spring CA CB EXACT FIXED ROOT C S C Cys Sγ ligand reactive C C R N We avoid duplicate atoms and artificial S– C repulsion without masking in this framework. Downstream torsion search Schematic torsion tree, not molecular coordinates or a bond-length scale. Invariant Confromer quality matters as much as docking engine Copyright © Elix, Inc. All rights reserved
14. Two benchmarks: Pose recovery & Enrichment C207: 92 reference complexes [7] COValid: 10 cysteine sites / 9 targets [8] Cys–acrylamide Michael-addition pose recovery. Generated vs. deposited ligand/adduct geometries. Compare predicted poses with crystal references. 38,793 protomers: 874 active, 37,919 decoy. 496 active-parent entries across site libraries. Fresh conformers; rank actives against decoys. Evaluation Primary readout Pose recovery · C207 Top-1 and best-of-3 RMSD ≤2 Å against the crystal pose Enrichment · COValid ROC-AUC · adjusted LogAUC · EF1% (means across 10 sites) Reactive geometry Explicit attachment or C–S ≤2 Å contact assessed alongside ranking Distinct preparation and evaluation protocols; results are reported separately. [7] Scarpino et al., J. Chem. Inf. Model. (2018). [8] Shamir et al., J. Am. Chem. Soc. (2026). Copyright © Elix, Inc. All rights reserved
15. Pose Recovery: C207 (Scarpino) Benchmark With generated conformers, QuickVinaGPU achieves comparable pose recovery to GNINA (57/92 vs. 59/92 top-1) at lower recorded computational time (2.98 vs. 8.78 s) Protocol Scoring functions Confromer Preparation Top pose RMSD ≤2 Å Within first 3 RMSD ≤2 Å Mean measured time, s (binary) Contact model QuickVinaGPU Vina Generated + OB 57/92 71/92 2.98 (2.91) Explicit S–C bond GNINA Vina Generated + OB 59/92 72/92 8.78 (8.78) Explicit S–C bond QuickVinaGPU Vina Deposited + OB 65/92 78/92 2.70 (2.64) Explicit S–C bond GNINA Vina Deposited + OB 63/92 76/92 8.79 (8.79) Explicit S–C bond QuickVinaGPU Vina Deposited-adduct 82/92 88/92 2.70 (2.63) Explicit S–C bond AutoDock-GPU tether AD4 Deposited-adduct 56/92 62/92 5.74 (1.10) Explicit S–C bond AutoDock-GPU global reactive Reactive AD4 Generated 47/92 55/92 5.79 (1.42) Reactive potential Free QuickVinaGPU Vina Generated 13/92 24/92 3.14 (2.99) Unconstrained Compare methods within matched input conditions. Deposited: Crystal-derived adduct geometry is an optimistic pose-recovery diagnostic, not a realistic virtual-screening input. OB: open babel attachment. QQV: 5,000 lanes, depth 1–20; GNINA: exhaustiveness 8 (best), CNN disabled. Mean recorded times; search budgets, attachment models differs. AutoDock-GPU timings are search-only. [7] Scarpino et al., J. Chem. Inf. Model. 2018, 58, 1441–1458. DOI: 10.1021/acs.jcim.8b00228. Copyright © Elix, Inc. All rights reserved
16. COValid Enrichment Benchmark Native covalent workflows show weaker enrichment than conventional docking, despite explicitly enforcing or favoring reactive attachment geometry. Native search workflow Scoring Function ROC AUC Adjusted LogAUC[8] EF1% (fold) Mean (s)† Reactive contact ≤2 Å all / active Free QuickVinaGPU / original pre-reaction ligand Vina 0.773 22.18% 12.91 2.85 0.0% / 0.0% Free QuickVinaGPU / saturated H-capped ligand Vina 0.774 21.77% 13.16 2.87 0.0% / 0.0% Free GNINA / saturated H-capped ligand Vina 0.794 24.68% 16.37 14.02 0.0% / 0.0% QuickVinaGPU covalent Vina 0.663 10.94% 5.13 3.12 Bond imposed GNINA covalent Vina 0.686 11.23% 4.27 20.45 Bond imposed AutoDock-GPU tethered AD4 0.558 1.35% 0.20 5.55 Bond imposed AutoDock-GPU global Reactive AD4 0.729 14.10% 5.98 4.62 60.6% / 66.5% ROC-AUC: Probability that an active outranks a decoy; random = 0.5. Adjusted LogAUC: ROC area on a logarithmic false-positive-rate axis (0.1–100%), normalized and corrected for random ranking. Emphasizes early enrichment. EF1%: Active fraction in the top 1% divided by the overall active fraction (fold enrichment). Mean recorded times; search budgets, attachment models differs Best protomer per active parent; decoy protomers retained separately; failures bottom-ranked. Reactive contact indicates C–S ≤2 Å, not confirmed covalent reactivity. Mean recorded stage times; preparation scope differs across engines. [8] Shamir et al., J. Am. Chem. Soc. 2026, 148, 13043–13054. DOI: 10.1021/jacs.5c22222. Copyright © Elix, Inc. All rights reserved
17. Combining Tethered Poses with Reactive AD4 Scoring The hybrid workflow offers a compelling balance of enrichment, covalent pose reliability, and throughput: approaching free-docking enrichment while maintaining C–S proximity. Covalent endpoint Detach and H-cap Restore intact Cys AD4 reactive score Explicit Cys–ligand bond Saturated product-like ligand No explicit ligand–Cys bond Fixed score or local refinement Pose source Reactive refinement Selection pool Ranking score ROC AUC Adjusted LogAUC EF1% (fold) Mean timed (s) Selected C–S ≤2 Å all / active QuickVinaGPU covalent No Top-1 Vina 0.663 10.94% 5.13 3.12 98.2% / 99.4% QuickVinaGPU covalent No Top-1 Reactive AD4 0.735 18.31% 10.96 3.12 + n/t¹ 98.2% / 99.4% QuickVinaGPU covalent Yes + Fallback Top-1 Reactive AD4 0.735 18.22% 9.12 7.41 98.2% / 99.4% QuickVinaGPU covalent Yes + Fallback Pool-10 Reactive AD4 0.749 19.01% 11.15 7.41 98.2% / 99.4% GNINA covalent Yes + Fallback Pool-10 Reactive AD4 0.752 18.94% 9.61 24.67 98.5% / 99.8% Autodock-GPU tether Yes + Fallback Pool-10 Reactive AD4 0.721 13.99% 6.29 9.77 98.3% / 99.8% Free QuickVinaGPU No Top-1 Vina 0.773 22.18% 12.91 2.85 0.0% / 0.0% Autodock-GPU global reactive Default Top-1 Reactive AD4 0.729 14.10% 5.98 4.62 60.6% / 66.5% Ten-site means. Pool-10 selects among up to 10 poses; fallback retains the original seed when needed. Selected C–S ≤2 Å includes fallbacks (all / active). Times are recorded stage means, not end-to-end; fixed rescoring untimed, top-1 refinement extracted from pool-10. Fresh preparation and QV attachment excluded. Scoring and representation both change. Copyright © Elix, Inc. All rights reserved
18. Visual Inspection Copyright © Elix, Inc. All rights reserved
19. Covalent Docking: Conclusions & Next Steps Benchmark Conclusion Recap 10.94% 18.31% 22.18% Native QuickVinaGPU covalent Fixed reactive AD4 Free QuickVinaGPU No selected C-S contact within 2.0 A Fixed AD4 rescoring uses the same tethered poses; only the representation and ranking score change. Reactive scoring inside QuickVinaGPU Current hybrid: two engines Treat the selected reactive C-S contact differently, QuickVinaGPU covalent poses while retaining ordinary Vina interactions elsewhere. -> detach / H-cap + restore Cys AD4-inspired[9]: 13-7 well; assess neighbor contacts Vina-native option: adjust distance / repulsion terms -> external AutoDock4 reactive score Added molecular conversion and runtime overhead Goal: recover reactive-aware ranking in QuickVinaGPU without relying on external AD4 rescoring. [9] Bianco et al., J. Chem. Inf. Model. 2023, 63, 5631-5640. DOI: 10.1021/acs.jcim.3c00832. Copyright © Elix, Inc. All rights reserved
20. References METHODS & SOFTWARE [1] BENCHMARKS & REACTIVE DOCKING Trott & Olson (2010). AutoDock Vina: Improving the Speed and Accuracy of Docking with a New Scoring Function, Efficient Optimization, and Multithreading. J. Comput. Chem. 31, 455–461 | DOI: 10.1002/jcc.21334 [2] McNutt et al. (2025). GNINA 1.3: The Next Increment in Molecular Docking with Deep Learning. [4] [7] IEEE/ACM Trans. Comput. Biol. Bioinform. 21, 2382–2393 DOI: 10.1109/TCBB.2024.3467127 Scarpino et al. (2018). Comparative Evaluation of Covalent Docking Tools. J. Chem. Inf. Model. 58, 1441–1458 | DOI: 10.1021/acs.jcim.8b00228 [8] Tang et al. (2024). Vina-GPU 2.1: Towards Further Optimizing Docking Speed and Precision of AutoDock Vina and Its Derivatives. Hartshorn et al. (2007). Diverse, High-Quality Test Set for the Validation of Protein–Ligand Docking Performance. J. Med. Chem. 50, 726–741 | DOI: 10.1021/jm061277y Santos-Martins et al. (2021). Accelerating AutoDock4 with GPUs and Gradient-Based Local Search. J. Chem. Theory Comput. 17, 1060–1073 | DOI: 10.1021/acs.jctc.0c01006 Berman et al. (2000). The Protein Data Bank. Nucleic Acids Res. 28, 235–242 | DOI: 10.1093/nar/28.1.235 [6] J. Cheminform. 17, 28 | DOI: 10.1186/s13321-025-00973-x [3] [5] Shamir et al. (2026). Discovery of Covalent Ligands with AlphaFold3. J. Am. Chem. Soc. 148, 13043–13054 | DOI: 10.1021/jacs.5c22222 [9] Bianco et al. (2023). Reactive Docking: A Computational Method for High-Throughput Virtual Screenings of Reactive Species. J. Chem. Inf. Model. 63, 5631–5640 | DOI: 10.1021/acs.jcim.3c00832 Copyright © Elix, Inc. All rights reserved