https://pubs.acs.org/doi/10.1021/jacs.5c22222

Discovery of Covalent Ligands with AlphaFold3

Key figures

• [Figure 1]: Defines COValid as a covalent virtual-screening benchmark with acrylamide actives, property-matched decoys, and docking-parameter optimization.
  
• [Table 1]: Quantifies the central benchmark result: AF3 ranked by mPAE achieved 71.8 ± 5.9% adjusted LogAUC versus 12.1 ± 9.9% for DOCK6, 9.9 ± 4.8% for DOCKovalent, and 5.7 ± 7.9% for AutoDock.
  
• [Figure 2]: Shows that AF3 mPAE strongly enriches true covalent binders across COValid and also improves noncovalent enrichment across DUDE-Z.
  
• [Figure 3]: Demonstrates prospective BTK hit validation, including covalent labeling, GSH reactivity control, DSF stabilization, kinase inhibition, cellular pBTK inhibition, kinome profiling, and chemoproteomic selectivity.
  
• [Figure 4]: Validates AF3-predicted YS1 and YS2 binding poses by X-ray crystallography with ligand heavy-atom RMSDs of 0.50 Å and 0.41 Å.
  
• [Figure S11]: Shows that mPAE does not function as an affinity predictor, because it does not meaningfully correlate with IC50 values.
  
• [Figure S12]: Reveals a selectivity caveat by showing cross-docking enrichment for ligand sets against proteins with similarly positioned cysteines.
  
• [Figure S28]: Reveals a geometric caveat: AF3 covalent cofolding can generate malformed covalent bond lengths and angles.
  

1) Thesis (one sentence)

To address the scarcity of computational methods for covalent virtual screening, in cysteine-targeting acrylamide libraries against kinases and K-Ras plus a prospective BTK screen, AlphaFold3 covalent cofolding ranked by minimal predicted aligned error causes strong enrichment of true covalent binders and discovery of a novel selective BTK inhibitor by modeling covalent protein-ligand complexes from sequence and ligand inputs and using mPAE as a structural-confidence classifier, supported by COValid/DUDE-Z benchmarking, LC-MS labeling, kinase assays, cellular Western blotting, chemoproteomics, kinome profiling, and X-ray crystallography.

2) Evidence card (three bullets only)

• Strongest result: (Table 1; Fig. 2; Fig. 3; Fig. 4) AF3 mPAE achieved 71.8 ± 5.9% average adjusted LogAUC on COValid, then prospectively discovered YS1, a covalent BTK inhibitor with 30 nM biochemical IC50, low-hundreds-of-nM cellular pBTK inhibition, and an X-ray-validated ligand pose at 0.50 Å RMSD.
  
• Method enabler: (Fig. 1; Fig. 2; Fig. S30; research type + tools) Computational virtual screening plus experimental validation used COValid property-matched covalent decoys, AF3 bondedAtomPairs covalent cofolding, mPAE ranking, RDKit ligand preparation, Rosetta minimization/rescoring, LC-MS, kinase assays, DSF, GSH reactivity assays, chemoproteomics, kinome profiling, and X-ray crystallography.
  
• Critical limitation: (Fig. S11; Fig. S12; Fig. S25; Fig. S28) mPAE is a structural-confidence classifier rather than an affinity or selectivity model; it did not meaningfully correlate with IC50, cross-enriched ligands at analogous cysteine sites, mispredicted the YS3 binding orientation, and can output distorted covalent bond geometry.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “Virtual screening of large ligand libraries is a cornerstone of modern drug discovery.” (Discussion, page 8)
  
• Quote 2: “Our key finding was that AF3 vastly outperformed traditional methods.” (Discussion, page 8)
  
• Quote 3: “AF3 confidence metrics were trained to quantify modeling accuracy.” (Discussion, page 8)
  
• Quote 4: “YS1 is not a “back pocket” binder nor can it be characterized as a front pocket binder.” (Results/Fig. 4, page 7)
  

Key comparisons (1–3 lines)

• Compared to: DOCKovalent, DOCK6 attach-and-grow, AutoDock flexible-side-chain covalent docking, and traditional static-structure virtual screening.
  
• Win: AF3 mPAE was far stronger for early enrichment, enabled a prospective BTK screen from ~906K acrylamides, and recovered chemically novel YS1 with potent biochemical, cellular, structural, and selectivity validation.
  
• Tradeoff: AF3 was much slower per ligand, modeled the covalent adduct rather than the pre-reactive intermediate, and mPAE did not directly encode electrophile reactivity, kinact/KI, binding affinity, or proteome-wide selectivity.
  

Methods I might copy (protocol hooks)

• Construct design / Models: For AF3 covalent cofolding, reduce the acrylamide double bond to the adduct, generate a 3D conformer with RDKit ETKDGv3, minimize with MMFFOptimizeMolecule, convert to CCD mmCIF as userCCD, and specify the covalent bond using bondedAtomPairs between cysteine Sγ and acrylamide Cβ. Run one AF3 seed to generate five samples and select the top-ranked prediction. For COValid-style benchmarking, curate cysteine acrylamide actives with exactly one acrylamide, molecular weight ≤500 Da, rotatable bonds ≤12, activity ≤100 nM, protonation states from Dimorphite-DL at pH 7.0-7.8, and decoys from ZINC20 free-amine precursors with matched molecular weight, hydrogen-bond donors/acceptors, charge, LogP, and rotatable bonds. For prospective BTK screening, filter Enamine REAL from 6.71B compounds to MW 250-500 Da, rotatable bonds ≤12, and at least one primary/secondary aliphatic amine; diversity-filter to 887,569 precursors; convert to 906,510 acrylamides; keep AF3 predictions with mPAE ≤0.9 Å; remove compounds with Tanimoto similarity ≥0.35 to known submicromolar BTK inhibitors; manually inspect hinge-bonding poses.
  
• Conditions / Instruments: For intact protein LC-MS screening, use 10 mM compound stocks in DMSO; mix 1 µL compound with 49 µL of 1 µM BTK kinase domain in 25 mM HEPES pH 7.5 and 50 mM NaCl to give 200 µM compound; incubate 2 h or overnight at 25 °C; stop 20 µL sample with equal volume 20% acetonitrile + 0.25% TFA; analyze on Waters ACQUITY UPLC class H with C4-BEH column, 40 °C column, 10 °C autosampler, 0.4 mL/min flow, 0.1% formic acid mobile phases, Waters SQD2 detector, and MaxEnt deconvolution. For kinase inhibition, use 3 nM BTK, 20 mM HEPES pH 7.5, 10 mM MgCl2, 1 mM EGTA, 0.01% Brij35, 0.02 mg/mL BSA, 0.1 mM Na3VO4, 2 mM DTT, 1% DMSO, 60 min compound preincubation at room temperature, 2 h reaction with 0.2 mg/mL pEY substrate and 33P-ATP, with 20 µM ATP for WT BTK and 50 µM ATP for C481S BTK.
  
• Readout / Analysis: Rank virtual libraries by mPAE and benchmark enrichment with adjusted LogAUC, where 0% corresponds to random enrichment and 85.5% to all actives above all decoys. Confirm covalent engagement by intact protein LC-MS, non-promiscuous reactivity by GSH assay using 5 mM GSH and 0.2 mM compound at pH 8 and 25 °C, target stabilization by DSF using 5 µM BTK-KD and fourfold compound excess overnight at 25 °C, cellular BTK pathway inhibition in Mino cells after 1 h compound treatment and 10 µg/mL anti-human IgM stimulation for 10 min at 37 °C, and proteomic selectivity using 1 µM YS1 competition followed by 1 µM Probe 4 or 2 µM XO44 for 45 min with n = 4 and log2 difference >2 plus p < 0.01 filtering. For covalent kinetics, use 100 nM His6-BTK labeled with 25 nM RED-tris-NTA dye, 24-point 1.5-fold compound dilution from 2 mM DMSO stock, 25 nM final BTK, 2.5% DMSO, Dianthus-X spectral-shift acquisition for 100 min at 4 min intervals, and fit kobs versus ligand concentration to estimate kinact, KI, and kinact/KI.
  

Open questions / Theoretical implications (2–5 bullets)

• Can mPAE-ranked covalent cofolding generalize from cysteine acrylamides to serine-, lysine-, tyrosine-, or histidine-reactive warheads?
  
• Would modeling the pre-reactive encounter complex improve prediction of kinact/KI and selectivity relative to adduct-state modeling?
  
• Can AF3 confidence metrics be recalibrated for flexible, shallow, or protein-protein-interface surfaces rather than kinase ATP pockets?
  
• How should AF3 mPAE be combined with physics-based scoring, covalent geometry filters, and chemoproteomic priors to avoid false positives with plausible poses but poor kinetics?
  
• Can the same cofolding-confidence logic rank ternary or chemically induced proximity complexes rather than binary covalent adducts?