A structure-based paralog-selectivity triage of five candidate ENPP1 inhibitors against the two closest human paralogs, ENPP2 / autotaxin and ENPP3. The analysis reverses a naïve affinity ranking: the compound that docks best into ENPP1 is the least selective, because its potency comes from chelating the catalytic zinc — the one feature every ENPP shares.
This is a computational methods demonstration, not a drug. Every number here is a prediction (ChEMBL bioactivity + docking + structural analysis). There is no wet-lab data: no measured IC50, no cellular assay, no PK, no tox. See DISCLAIMER before citing anything.
| Lead | Warhead | pChEMBL | QED | ENPP1 | ENPP2 | ENPP3 | worst margin [95% CI] | verdict |
|---|---|---|---|---|---|---|---|---|
| CHEMBL6149286 | hydroxamate | 9.20 | 0.50 | −9.2 | −10.0 | −11.0 | −1.77 [−1.82,−1.72] | anti-selective ✗ |
| CHEMBL6174154 | hydroxamate | 9.70 | 0.56 | −8.9 | −8.9 | −10.2 | −1.31 [−1.36,−1.27] | anti-selective ✗ |
| TZV (native) | sulfamide | — | — | −8.8 | −8.5 | −10.0 | −1.12 [−1.19,−1.05] | ATP-site ref |
| CHEMBL5555976 | sulfonamide | 9.28 | 0.52 | −8.2 | −8.2 | −7.7 | +0.03 [−0.06,+0.11] | indistinguishable |
| CHEMBL5915707 | sulfamide | 9.30 | 0.73 | −7.9 | −6.8 | −7.5 | +0.42 [+0.39,+0.45] | ENPP1-preferring ✓ |
| CHEMBL5826130 | sulfamide | 9.66 | 0.73 | −7.8 | −7.3 | −7.7 | +0.12 [+0.11,+0.14] | ENPP1-preferring ✓ |
Affinities are means of 42 independent docking runs (seeds 1–20, 42; smina, matched catalytic-Zn box), in kcal/mol. Margin = paralog − ENPP1; positive ⇒ prefers ENPP1. Brackets are 95% bootstrap CIs (10k resamples) on the worst-case (least-selective) paralog margin; a verdict is called only when the CI excludes 0.
Bottom line: the prior consensus-docking winner — the hydroxamate CHEMBL6149286 — is anti-selective: it binds ENPP3 (−11.0) and ENPP2 (−10.0) better than ENPP1 (−9.2), and the effect is far outside docking noise (margin −1.77, CI [−1.82,−1.72]). Both hydroxamates carry this liability. The recommended nomination is the sulfamide CHEMBL5826130 (backup CHEMBL5915707): both are statistically ENPP1-preferring once the margin is resolved against a 42-seed noise floor. Note the trade-off the replicates expose — 5915707 has the larger, cleaner selectivity margin (+0.42 vs +0.12), while 5826130 leads on potency (pChEMBL 9.66) and ties on developability (QED 0.73) and draws the strongest orthogonal structural support (see below). The two are co-nominated; if a single compound must advance, 5826130 is favored on the potency/developability/structure composite and 5915707 is the selectivity-first alternative. Both sit on the clinically validated sulfamide chemotype and avoid the hydroxamate's Ames/mutagenicity alert.
Note on margin magnitude. These margins (+0.1 to +0.4 kcal/mol) are statistically resolved but pharmacologically small — they establish direction, not fold-selectivity. The literature bar (ISM5939: >3,400× ENPP3) will require medicinal-chemistry optimization of the margin, not just selection among existing leads. See the analog-design campaign in
analysis/for a first attempt at widening it.
The hydroxamate warhead is a strong, geometry-tolerant bidentate Zn²⁺ chelator. That is exactly what makes it top the raw-affinity docking and what makes it non-selective: the catalytic bimetal Zn center is the single most conserved feature of the whole ENPP family, so a ligand whose binding is dominated by Zn chelation cannot tell ENPP1 from ENPP2/3. The sulfamide is a weaker, more directional Zn-interacting group that earns more of its affinity from the paralog-divergent pocket walls — a lower raw score, but selectivity by construction.
Structure-based selection must optimize the ENPP1-minus-paralog margin, not the absolute score.
Superposing the three catalytic pockets (anchor Cα RMSD 0.15–0.23 Å) over the 14 residues within 6 Å of the bimetal Zn:
- ENPP3 is identical to ENPP1 at all 14 core pocket positions (100% identity). This reproduces, from an independent structure, the published crystallography result that ENPP3 differs from ENPP1 by only ~2 residues within 4 Å of the ligand. ENPP3 selectivity has to be engineered, not hoped for.
- ENPP2 / autotaxin differs at 2 of 14 positions — His260→Leu214 and Ser377→Phe313 (ENPP1 numbering). These are the selectivity handles for the major circulating lysoPLD off-target.
| Compound | ENPP1 IC50 | vs ENPP2 | vs ENPP3 | note |
|---|---|---|---|---|
| ISM5939 (Insilico, clinical) | 0.63 nM | >15,000× | >3,400× | oral, IND-cleared; the bar |
| Enpp-1-IN-27 | 14.7 nM | ~410× | ~10× | ENPP3 is the weak axis even for good compounds |
| STF-1623 | <2 nM Ki | — | >1,000× | ultralong residence time |
ENPP3 selectivity is the recurring field bottleneck — consistent with the 100% core-pocket identity measured here. Any nomination must be assayed against ENPP3, not just ENPP2.
Full analysis in BENCHMARK.md (key claims cross-checked to real PMIDs/DOIs in benchmark/literature_grounding.csv) — grounded in 59 triaged PubMed
medicinal-chemistry papers, the complete ChEMBL bioactivity landscape for all three
paralogs, and 23 ENPP1 clinical trials, all queried programmatically and DOI-verified.
The single most important benchmark finding — and the reason this counter-screen is worth doing — is the data asymmetry in ChEMBL (pChEMBL ≥ 7):
| Paralog | Potent compounds | Head-to-head selectivity data |
|---|---|---|
| ENPP2 / autotaxin | 1,546 | — |
| ENPP1 | 199 | — |
| ENPP3 | 3 | — |
| Any compound tested on >1 paralog | exactly 1 in all of ChEMBL |
Public cross-paralog selectivity data barely exists. A structure-based counter-screen is a rational way to generate selectivity hypotheses where assay data is absent.
The field is ahead on potency and the clinic (ISM5939: 0.63 nM, >3,400× ENPP3, IND-cleared; four small-molecule ENPP1-inhibitor oncology programs in the clinic, two now in Phase 2; a generative-AI-designed oral inhibitor in Nat. Commun. 2025). We do not claim novel chemistry or a candidate. We build on that work — the same open docking toolchain (smina/Vina + Open Babel + RDKit) and the same 2024 biology that established ENPP3 as the second cGAMP hydrolase (Li et al., Cell Rep.) — and add replicate statistics, a 4-method consensus, mechanism analysis, and a systematic three-paralog comparison on top.
We stress-tested the central finding on 7 additional, more-potent public ENPP1 inhibitors (pChEMBL 9.18–10.05, grounded live from ChEMBL) — which all turned out to be phosphonates, a fourth warhead class beyond the original screen. Same receptors, same matched Zn boxes, same protocol. Full detail in expanded_screen/README.md.
The pattern holds and confirms the structural thesis: ENPP2 discrimination is broadly attainable (mean margin +0.80 kcal/mol, ≥0 for all 7 — the two ENPP2 switch residues are exploitable), while ENPP1-vs-ENPP3 discrimination is hard (mean −0.21, 3 of 7 anti-selective — the conserved pocket defeats metal-engaging warheads). The most potent compound in the entire program (CHEMBL5566114, pChEMBL 10.05) is anti-selective vs ENPP3 — potency alone does not solve the hard selectivity axis.
Full detail in clinical_ip/CLINICAL_IP_LANDSCAPE.md. Trials pulled live from ClinicalTrials.gov; patents from PubChem compound→PatentID cross-references (PUG-REST), each granted-patent number re-verified against the live PubChem record.
Clinical (47 ENPP1-relevant trials). The direct competitive set is four small-molecule ENPP1-inhibitor immuno-oncology programs — RBS2418 (Riboscience) and SR-8541A (Stingray) now in Phase 2, ISM5939 (InSilico) and TXN10128 (Txinno) in Phase 1. A separate, larger block of rare-disease trials (Inozyme's INZ-701 enzyme-replacement program in GACI/ARHR2/PXE) targets ENPP1 deficiency — opposite pharmacology, but it establishes ENPP1 as a clinically validated, druggable target.
IP / freedom-to-operate. The two sulfamide leads (CHEMBL5826130, CHEMBL5915707) map to a granted US composition-of-matter patent, US-11591313-B2, plus two continuation applications — validated chemotype, but occupied space. The hydroxamate leads appear only in a single 2024 application (US-2024/0383893-A1) — more white space, but the hydroxamate warhead carries the Zn-coordination liability this project's mechanism analysis flagged as the driver of anti-selectivity. CHEMBL5555976 (sulfonamide) has no PubChem patent cross-reference at all. Implication: the AN10 piperidinol analog sits on the sulfamide scaffold — it inherits the granted-patent precedent and therefore needs a composition defensibly distinct from US-11591313-B2.
Docking is a directional selectivity proxy, not an IC50 ratio, and none of this is a freedom-to-operate legal opinion — it is a programmatic reading of public patent cross-references to orient the chemistry.
The analysis/ directory stress-tests the headline ranking with
four orthogonal methods plus a rational analog-design campaign. Full write-up:
analysis/README.md. A candid self-assessment of the whole
repo — including a defect we found and fixed — is in
quality_review.md.
1. Statistical rigor (42-seed replicates + bootstrap). The margins are no longer n=1. Every ligand×target was docked 42 times; bootstrap CIs put each verdict on a significance footing. Hydroxamates are significantly anti-selective, both sulfamides significantly ENPP1-preferring, the sulfonamide indistinguishable from zero.
2. The mechanism, measured. The closest ligand-atom-to-catalytic-Zn distance per pose explains why: hydroxamates and TZV directly coordinate the ENPP3 zinc (1.7–2.2 Å) but not ENPP1's; sulfamides bind pocket walls, not metal — the tunable surface.
3. Interaction fingerprints + consensus rescore. The nominee contacts more ENPP1 pocket residues than either paralog (energy-independent), and a second scoring function (Vinardo) reproduces the hydroxamate anti-selectivity while disagreeing on the small sulfamide margins — reported honestly.
4. Rational analog design. 13 Ro5-clean analogs of CHEMBL5826130 designed to exploit the ENPP2 switch residues; the best (a 3-hydroxy-piperidine variant) improves the predicted margin while retaining potency and developability. A synthesizable next iteration — predictions to be assayed, not conclusions.
A focused 15-seed replicate validates the top analog: AN10_piperidinol improves both predicted potency (ENPP1 −9.08 vs parent −8.10) and worst-case selectivity margin (+0.70 [0.39, 0.95] vs +0.15), with the bootstrap CI excluding zero.
5. Protocol validation suite. Three controls calibrate how far the docking
protocol can be trusted on this target — a cognate redocking of the native
ligand, an actives-vs-decoys enrichment benchmark (an honest negative:
ROC-AUC 0.44, so absolute score does not rank binders — which is exactly why
only the within-ligand ENPP1-vs-paralog margin is used), and a
protonation-state robustness check (directional calls sign-robust over 8
states). Full detail in analysis/validation/ and the
referee ledger REVIEWER_GAP_ANALYSIS.md.
enpp1-selectivity-repo/
├── README.md # this file
├── LICENSE # MIT (code); data provenance noted
├── requirements.txt # Python + external-tool deps
├── enpp_lead_recommendation.md # full nomination memo
├── quality_review.md # adversarial self-assessment (grades + defect log)
├── BENCHMARK.md # positioning vs literature/competitors/open-source (verified)
├── benchmark/ # grounded field benchmark (PubMed + ChEMBL + trials)
│ ├── benchmark_literature_full.csv # 59 DOI-linked ENPP1 medchem/selectivity/structure papers
│ ├── key_benchmarks.csv # 7 landmark papers (crystal, AVA-NP-695, genAI, ENPP3-hydrolase)
│ ├── chembl_paralog_summary.csv # potent-compound counts per paralog
│ ├── chembl_landscape.csv # full ChEMBL bioactivity pull (pChEMBL>=7)
│ ├── clinical_trials.csv # ENPP1-inhibitor trials (NCT IDs, phase, status)
│ └── fig_landscape_benchmark.png # data-asymmetry + potency-distribution figure
├── clinical_ip/ # clinical competition + patent/FTO reading (verified)
│ ├── CLINICAL_IP_LANDSCAPE.md # trials + patent landscape + FTO reading
│ ├── clinical_trials_full.csv # 47 ENPP1-relevant trials (NCT, phase, status, sponsor)
│ ├── patent_landscape.csv # per-lead PubChem patent xrefs (granted + applications)
│ └── fig_clinical_patent_landscape.png # competitive phase + per-lead IP figure
├── expanded_screen/ # generalization test: 7 more-potent phosphonate inhibitors
│ ├── README.md # does the warhead->selectivity pattern hold on new chemotype?
│ ├── phosphonate_selectivity_scorecard.csv
│ ├── phosphonate_ligands.csv # 7 leads, grounded QED/potency/Ro5
│ ├── phosphonate_docking_raw.csv
│ └── fig_phosphonate_selectivity.png
├── analysis/ # extended rigor: replicates, mechanism, consensus, analogs
│ ├── README.md # methods for all four orthogonal analyses
│ ├── data/ # replicate scores, bootstrap CIs, Zn distances, IFP, analogs
│ ├── figures/ # 6 publication figures (CIs, Zn, IFP, consensus, analog design + validation)
│ └── scripts/ # run_replicates.sh, rep_analogs_focused.sh, run_analysis.py
├── writeups/ # long-form communication drafts
│ ├── substack_article.md # ~2,000-word methods narrative (limitations-first)
│ ├── x_thread.md # 10-beat thread + pinned caveat
│ └── x_article.md # condensed long-form version
├── data/
│ ├── ligands.csv # 5 leads + native TZV: SMILES, warhead, QED, pChEMBL
│ ├── docking_scores_raw.csv # long-form smina affinities (target,ligand,affinity)
│ ├── enpp_selectivity_scorecard.csv# margins + verdicts (regenerated by build_scorecard.py)
│ ├── enpp_active_site_conservation.csv # 14-residue pocket equivalence
│ ├── enpp_paralog_structures.csv # structure provenance (PDB, resolution, box centers)
│ ├── pocket_residues.json # per-structure pocket residues + distance to Zn
│ └── conservation_map.json # ENPP1↔ENPP2↔ENPP3 residue mapping
├── figures/
│ ├── enpp_selectivity.png # heatmap + margin ranking
│ └── enpp_active_site_conservation.png
├── receptors/ # prepared receptor PDBs (protonated, ligand/water stripped)
│ ├── enpp1_6wew_receptor.pdb
│ ├── enpp2_5mhp_receptor.pdb
│ └── enpp3_6c01_receptor.pdb
├── poses/ # representative docked poses (PDBQT) for the flip
│ ├── dock_enpp{1,2,3}_CHEMBL5826130.pdbqt # the nominee
│ └── dock_enpp{1,2,3}_CHEMBL6149286.pdbqt # the demoted hydroxamate
└── scripts/
├── prepare_ligands.py # SMILES → 3D → PDBQT (RDKit + obabel)
├── dock_all.sh # smina across 3 paralogs × 6 ligands
├── build_scorecard.py # raw scores → scorecard + verdicts
├── reproduce.sh # end-to-end driver
└── README.md # protocol notes
# 1. environment
mamba create -n enpp-dock -c conda-forge rdkit openbabel smina pandas numpy matplotlib seaborn
mamba activate enpp-dock
# 2. rebuild the scorecard from the committed raw scores (fast, no docking)
python scripts/build_scorecard.py
# 3. full re-dock from scratch (slower; needs receptors/*.pdbqt — see scripts/README.md)
bash scripts/reproduce.shThe scorecard step reproduces data/enpp_selectivity_scorecard.csv exactly from
data/docking_scores_raw.csv. The full re-dock regenerates the raw scores;
because docking is stochastic (mitigated here by --seed 42 --exhaustiveness 16), scores may vary by ≈0.1–0.3 kcal/mol run to run — see the disclaimer.
- Structures: ENPP1 6WEW (2.73 Å, native sulfamide TZV), ENPP2/ATX 5MHP (2.43 Å), ENPP3 6C01 (2.30 Å, apo). Receptors protonated, ligands/waters stripped, converted to PDBQT.
- Box: 24 Å cube centered on the catalytic-Zn centroid of each enzyme (paralog-agnostic, so the comparison is apples-to-apples across all three).
- Docking: smina,
--exhaustiveness 16 --num_modes 5. Best-mode affinity reported. Headline scores are means of 42 independent runs (seeds 1–20, 42) per ligand×target; the seed-to-seed SD (0.05–0.29 kcal/mol, median 0.08) is the empirical noise floor, and margin significance is assessed by 10k-resample bootstrap (seeanalysis/). - Ligands: RDKit ETKDGv3 embed (seed 42) + MMFF optimization, Gasteiger charges via obabel.
- Conservation: custom Kabsch superposition on 7 conserved anchor residues; pocket = residues within 6 Å of the Zn pair.
Full rationale and next-step design in
enpp_lead_recommendation.md.
This repository contains no experimental data and makes no therapeutic claim.
- Docking ΔΔG is a directional triage proxy, not an IC50 ratio. A +0.42
kcal/mol margin is statistically resolved (42-seed bootstrap) but is not a
selectivity fold-change. smina scoring does not model metal-coordination
quantum chemistry well. The counter-screen correctly flags direction
(hydroxamates promiscuous, sulfamides ENPP1-leaning) but the numbers cannot be
quoted as selectivity ratios. A second scoring function (Vinardo) agrees on
the hydroxamate anti-selectivity but disagrees on the sign of the small
sulfamide margins — so the developable-lead selectivity edge is directionally
suggestive, not scoring-function-robust (see
analysis/consensus_selectivity.csv). - Protonation-state sensitivity. Docking all plausible protonation/tautomer
states of one lead per warhead (8 states x 3 paralogs) leaves the directional
selectivity calls sign-robust (sulfamide ENPP1-leaning vs ENPP2; hydroxamate
anti-selective vs ENPP2 in every state; phosphonate anti-selective vs ENPP3 in
every state) while margin magnitudes move 0.4-1.0 kcal/mol — larger than the
seed-noise CIs, reinforcing that margins are directional, not quantitative.
See
analysis/validation/PROTOMER_README.md. - Retrospective enrichment (calibration). Docking 30 known ENPP1 actives
(pChEMBL >=7) against 90 property-matched decoys into 6WEW gives ROC-AUC 0.44
(below random); absolute smina affinity does not rank ENPP1 binders. This
is why no cross-chemotype absolute-affinity claim is made. The selectivity
margins survive this because they use the within-ligand ENPP1-vs-paralog
score difference (ddG), where ligand-specific scoring errors largely cancel
— a different and more forgiving quantity than the absolute ranking that
fails here. See
analysis/validation/ENRICHMENT_README.md. - Protocol validation (cognate redocking). Redocking the native 6WEW
ligand TZV with the identical protocol reproduces the correct sub-pocket and
catalytic-Zn engagement (top-scored pose 2.3 A centroid / 2.2 A Zn distance
from native) and separates pocket poses (3-4 A) from surface decoys
(10-13 A), but does not meet the strict <2 A RMSD success criterion
(best mode 3.08 A, top-scored 3.44 A). The residual error is in the
metal-chelation geometry — expected for a Vina-type function with no explicit
metal term. This bounds what the study claims: pocket-level discrimination is
resolvable; sub-2 A pose accuracy and fine ranking between similar chelators
are not. See
analysis/validation/. - ENPP3 (6C01) was docked apo; poses were not MD-relaxed; no explicit metal-coordination restraints were applied.
- pChEMBL/QED are database and cheminformatic properties, not measured potencies in this assay context.
- The decisive experiment is a wet-lab 3-enzyme ENPP1/2/3 biochemical panel. Everything here is hypothesis-generating until those IC50s exist.
Not affiliated with, or endorsed by, any company or clinical program named for comparison. Structures and bioactivities belong to their respective sources (RCSB PDB; ChEMBL, CC BY-SA 3.0).
If this workflow is useful, cite the repository and the underlying data sources
(RCSB PDB entries 6WEW/5MHP/6C01; ChEMBL). A CITATION.cff can be added on
first release.