henhaobio
Sign in

Science AdvancesBiophysicsPeer-reviewed journal article

Extensive exploration of structure activity relationships for the SARS-CoV-2 macrodomain from shape-based fragment merging and active learning

G. J. CorreyiD, Moira M. RachmaniD, Takaya TogoiD, Stefan GahbaueriD, Yagmur Umay DorukiD, Maisie GV StevensiD, Priyadarshini JaishankariD, Brian P. Kelley, Brian B. GoldmaniD, Molly SchmidtiD, Trevor KrameriD, Dmytro S. RadchenkoiD, Yurii S. MoroziD, Alan AshworthiD, Patrick RileyiD, Brian K. ShoichetiD, Adam R. RensloiD, W. Patrick WaltersiD, James S. FraseriD

PDFDOIPreprintCode

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (26 Aug 2024) · CC BY 4.0

Shape-based fragment merging (FrankenROCS) with Thompson sampling over 22 billion molecules found neutral SARS-CoV-2 macrodomain inhibitors; analogs reached single-digit micromolar IC50, backed by 137 crystal structures.

Earlier fragment linking gave sub-micromolar Mac1 inhibitors that depended on a carboxylic acid, with poor membrane permeability and other liabilities. FrankenROCS with Thompson sampling searches ultra-large synthesis-on-demand libraries for neutral scaffolds that preserve fragment shape and pharmacophore.

Key results

  1. ShownExhaustive FrankenROCS search of the Enamine HTS collection gave seven new adenosine-site binders, including two neutral ones. The best hit was AVI-313. Carboxylic acids were hit far more often than non-acids.

    2.1 million compounds searched; 39 purchased; structures for 10 (seven adenosine site). AVI-313 IC50 130 μM. Acid hit rate 5/11 (45%) vs 2/28 (7%) for non-acids. · FrankenROCS identifies alternative scaffolds in…

  2. ShownThompson sampling made a search of the 22 billion molecule REAL library tractable by scoring only a small fraction of products. It yielded six crystallographic adenosine-site ligands but no gain in affinity over the HTS search.

    97 fragment-pair queries; about 22 million molecules scored per query (~0.1%); 36 selected, 32 synthesized; structures for six. Best AVI-328 IC50 220 μM vs 130 μM for AVI-313 from the HTS search. · Using Thompson sampling (TS) to search Billions…; TS-FrankenROCS identifies sulfonamides binding…; Using TS to search REAL

  3. ShownAnaloging AVI-313 across aromatic, aliphatic, core and rigid series produced several low-micromolar neutral compounds with diverse binding poses. This added to the large structure and binding dataset.

    148 compounds screened, 121 crystal structures. IC50: AVI-3669 8.7 μM (>10-fold better than AVI-313), AVI-1504 9.8 μM, AVI-607 5.1 μM, AVI-3716 4.5 μM, AVI-3731 3.7 μM. · Extensive exploration of the chemical space…; Aromatic series explore hydrophobic packing in…; Aliphatic series reveals phosphate tunnel and…; Rigidification identifies ligand-induced…

  4. SuggestedAmong the five most potent 2-aminoethanol analogs, four hydrogen-bond the oxyanion subsite through the hydroxyl, and methyl or alkyl groups in the hydrophobic patch were linked to potency. The Phe156 peptide flip may favor binding, based on one matched pair.

    Four of the five most potent compounds hydrogen-bond the oxyanion subsite (AVI-3669 does not). Removing the hydrophobic-patch methyl cost 5-20-fold in affinity. One pair: AVI-607 (5.1 μM) vs AVI-4062 (55 μM), which lacks one methyl and does not induce the Phe156 peptide flip. · Rigidification identifies ligand-induced…; Extensive analoging reveals affinity…; Multiple scales of protein flexibility revealed…

  5. SuggestedThe lead-like AVI-313 analogs are reported to have good solubility and permeability and excellent mouse microsomal stability (AVI-1504 given as example). The five most potent scaffolds did not bind human MacroD2 or Targ1.

    Solubility, permeability and stability values are in Table S3 and not quantified in the text. HTRF selectivity data for the five top scaffolds against MacroD2 and Targ1 are in Fig. S3B,C; no binding detected. · Lead-like AVI-313 analogs have promising…; Discussion

Methods, data and reproducibility

Crystallographic fragment pairs were merged into ROCS queries ranked by Tanimoto Combo. Enamine HTS was searched exhaustively and REAL by Thompson sampling with strain, minimization and interaction-fingerprint filters. Hits were tested by crystal soaking with PanDDA and an HTRF peptide-displacement assay.

  • Mac1 X-ray fragment screen binders used to build merged ROCS queries (5RS7 as alignment reference) · SARS-CoV-2 · n = 214 unique binders; 7181 fragment pairs · PDB 5RS7 (reference structure)
  • Enamine HTS collection searched exhaustively with FastROCS · n = 2.1 million compounds
  • Enamine REAL synthesis-on-demand database searched by Thompson sampling · n = more than 22 billion molecules; 97 fragment-pair queries
  • Mac1-ligand crystal structures from FrankenROCS hits and AVI-313 analogs, with HTRF binding data · SARS-CoV-2 · n = 137 structures; 148 analogs screened, 121 structures
  • Selectivity counter-screens of top scaffolds against human macrodomains MacroD2 and Targ1 (HTRF) · human · n = five scaffolds tested
Methods & tools
FrankenROCS, OpenEye FastROCS (ROCS, Tanimoto Combo), Thompson sampling (TS), RDKit (CombineMols, Taylor-Butina clustering), MayaChemTools torsion strain filter, OpenFF ff14sb_parsley minimization (OEForceField), ProLif interaction fingerprints, PanDDA, XDS, Aimless, Dimple, phenix.refine, Refmac, COOT, HTRF peptide-displacement assay, GraphPad Prism v.10.0.2, SmallWorld and Arthor
Software & environment
CCP4 version 7.0 (PanDDA), COOT 0.9.6 and 0.8.9.2, phenix.refine 1.21.1-5286, LigPrep 2022-1, GraphPad Prism v.10.0.2; PDB 5RS7 and 5RSG used as reference structures.
Assay conditions & compute
FastROCS runs on a GPU at more than 1 million conformers per second; the 7181 queries were searched in a couple of hours. Thompson sampling is described as working on commodity computing hardware.
Reproducibility gaps
The text gives no deposition accessions or data URLs for the 137 structures. Per-compound assay and ADME values are in supplementary tables not included here.
Limitations (6)
  • The limited affinity of the hits likely reflects the ROCS-based scoring, even though Thompson sampling is expected to have found the best-scoring REAL molecules. Scoring that correlates better with affinity may do better. · stated by the authors
  • Thompson sampling requires a library built from reagents and reactions, can over-focus on single building blocks, and top scores can reflect artifacts such as wrong tautomers and torsional strain. · stated by the authors
  • Testing racemates and diastereomer mixtures lowers the effective concentration of each isomer and makes the bound isomer(s) ambiguous in density maps. AVI-3669 IC50 could be up to four-fold more potent than reported. · stated by the authors
  • Two pyridines (AVI-290, AVI-295) had IC50 values of 56 and 52 μM but never showed binding in crystal soaks; the cause (crystal form or assay interference) is unresolved. · stated by the authors
  • Only 10 of 39 and 6 of 32 purchased compounds gave structures, and potency rests on HTRF IC50 from three technical replicates. No orthogonal biophysics or cellular antiviral data are reported. · noted in this summary
  • No head-to-head comparison against other fragment-merging methods (e.g. Fragmenstein) or against exhaustive scoring is shown in this campaign, so the efficiency gain of Thompson sampling rests on cited earlier work. · noted in this summary

Read this if you design fragment-linking or ultra-large library searches and want worked structure-activity data and binding poses for neutral SARS-CoV-2 macrodomain inhibitors.

Citations
28
FWCI
8.61
Field percentile
top 1%

Abstract

The macrodomain contained in the SARS-CoV-2 non-structural protein 3 (NSP3) is required for viral pathogenesis and lethality. Inhibitors that block the macrodomain could be a new therapeutic strategy for viral suppression. We previously performed a large-scale X-ray crystallography-based fragment screen and discovered a sub-micromolar inhibitor by fragment linking. However, this carboxylic acid-containing lead had poor membrane permeability and other liabilities that made optimization difficult. Here, we developed a shape- based virtual screening pipeline - FrankenROCS - to identify new macrodomain inhibitors using fragment X-ray crystal structures. We used FrankenROCS to exhaustively screen the Enamine high-throughput screening (HTS) collection of 2.1 million compounds and selected 39 compounds for testing, with the most potent compound having an IC50 value equal to 130 M. We then paired FrankenROCS with an active learning algorithm (Thompson sampling) to efficiently search the Enamine REAL database of 22 billion molecules, testing 32 compounds with the most potent having an IC50 equal to 220 M. Further optimization led to analogs with IC50 values better than 10 M, with X-ray crystal structures revealing diverse binding modes despite conserved chemical features. These analogs represent a new lead series with improved membrane permeability that is poised for optimization. In addition, the collection of 137 X-ray crystal structures with associated binding data will serve as a resource for the development of structure-based drug discovery methods. FrankenROCS may be a scalable method for fragment linking to exploit ever-growing synthesis-on- demand libraries.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSchuller '20Gahbauer '22Ban '26Sherrill '22Virdi '20Ni '21
79% semantic matchBiochemistry
M. Schuller, G. J. Correy, S. Gahbauer+38 more·bioRxiv·Nov 24, 2020