COCONUT R1: Screening Nature's Chemical Space for Longevity Targets

How we screened over 400,000 natural products against an 8-target longevity panel using the EstimaBio Boltz-2 pipeline, and what the hits tell us about nature's chemical space for ageing biology.

What Is COCONUT?

The Collection of Open Natural Products (COCONUT) is the largest freely available aggregated database of natural product structures. COCONUT R1 contains over 400,000 unique, non-redundant compounds sourced from 80+ open-access natural product databases, including AfroDb, AfroNatural, CarotenoidDB, Carvone, CMAUP, COCONUT itself, CyanMetDb, DietaryNaturalProduct, Dr. Duke's, Edible Medicinal Plant Database, EOMN, Exxon, Ferrdb, Flora, Food Database, FooDB, Guinea, Herb Ingredients, Herb Ingredients' Targets, HerbalDB, HIT, Indigeneous, IndianMedicinalPlantsDatabase, Indonesian, IndonesianHerbal, InflamNat, iPhyt, ITS, JNP, KnapSacK, LDC, LIPID MAPS, MAPS, Marine, MEGx, MMV, NANPDB, NapSAC, Natural Product Alert, Natural Product Atlas, Natural Product Magnetic Resonance Database, Natural Products, Natural Products Discovery, Natural Products Library, Natural Products Repository, NaturArznei, NCGC, NCI Open Database, NIST, Nubbe, NuBbA, Nutraceuticals, Open Science, Open Source Malaria, PhytoPharmacological, PhytoPharmacon, PhytoPharma, PhytoPharmas, PigmentDB, Plant Phenolics, PlantWithBenefits, Planta, Procar, PubChem, Selleck, StreptomeDB, SuperNatural II, Swedish Natural Products, TCM, TCM Database@Taiwan, TIPdb, TIPdb-3D, TIPdb-AC, TIPdb-SC, TIPdb-SM, TIPdb-ST, Traditional Chinese Medicine, Traditional Chinese Medicine Integrated Database, UNPD, UNPD-A, UNPD-B, and UNPD-C.

All compounds are released under a CC0 public domain license, meaning they can be used, screened, and developed without licensing restrictions — a critical advantage for downstream commercialisation. Each compound is annotated with its source organism, known bioactivities (where available), and computed physicochemical properties.

The 8-Target Longevity Panel

We selected eight protein targets with strong genetic or pharmacological evidence linking their modulation to lifespan extension or healthspan improvement in model organisms:

These targets span several hallmarks of ageing: nutrient sensing (AMPK, mTOR), epigenetic regulation (SIRT1), proteostasis and oxidative stress (KEAP1/NRF2), cellular senescence (BCL-XL), NAD+ metabolism (PARP1, CD38), and chronic inflammation (NLRP3). A compound that simultaneously modulates multiple targets in this panel has the potential to address ageing biology at several orthogonal axes.

Campaign Design

The full COCONUT R1 library was pre-filtered through a computational ADMET sieve: molecular weight ≤ 600 Da, LogP ≤ 5, hydrogen-bond donors ≤ 5, hydrogen-bond acceptors ≤ 10, and PAINS substructures removed. For each of the eight targets, a top-priority tier of 459–493 compounds was retained for detailed Boltz-2 scoring and analysis. The campaign design deliberately capped the top hits per target at roughly 500 to keep total compute time tractable while ensuring deep coverage of the best-scoring chemistry.

Results

Hit rates varied dramatically across the panel, reflecting differences in target tractability and the fraction of chemical space that engages each binding site. NLRP3's high hit rate (37.7%) likely reflects a relatively promiscuous ATP-binding pocket that accommodates diverse natural product scaffolds. PARP1's low hit rate (3.7%) is consistent with a highly selective NAD+ mimicry binding mode that few unoptimised natural products satisfy.

Post-campaign update (2026-06-22): A subsequent neg-control audit found that NLRP3/7PZC has a 33% false-positive rate at the 0.350 threshold used during COCONUT R1 — bland metabolites (glucose, taurine, asparagine) score above threshold. The Batch-206 AUC=1.000 was over-fit on a chemotype-narrow calibration set. NLRP3 has been removed from the active longevity panel. NLRP3-hitting leads from this campaign should not be prioritised for wet-lab follow-up without re-evaluation against a new calibration set.

The Triple-Hit Chalcone: ZAVWFOFRECTXEE

A single compound scored as a top hit against three independent targets: ZAVWFOFRECTXEE, a chalcone-type flavonoid. This compound registered binding probabilities above threshold for PARP1, CD38, and NLRP3 simultaneously. Triple-target activity is rare in unoptimised natural products and initially appeared promising.

However, ZAVWFOFRECTXEE was flagged by every PAINS (pan-assay interference compounds) filter we applied. Chalcones are among the most frequently flagged PAINS chemotypes due to their intrinsic reactivity (Michael acceptor), redox activity, and tendency to form colloidal aggregates in biochemical assays. The promiscuous binding profile across three unrelated targets is itself a red flag for assay interference rather than genuine polypharmacology.

This result illustrates a critical lesson for virtual screening campaigns: a compound that hits multiple targets in a computational screen should be immediately suspect for PAINS interference unless orthogonal evidence supports genuine multi-target activity. We report ZAVWFOFRECTXEE as a cautionary case study and have deprioritised it for wet-lab follow-up. All downstream lead prioritisation was filtered through PAINS, REOS, and aggregation-prediction filters.

CD38 Selectivity: Campaign 40 P2

CD38 emerged as the most tractable target in the panel: a 24.0% hit rate and a well-characterised binding pocket with known structure–activity relationships from the literature. Because raw binding score alone says nothing about selectivity, we ran a focused follow-up — Campaign 40, Phase 2 — to ask which CD38 hits avoid PARP1, the other NAD+-pocket target most likely to drive off-target cross-reactivity.

Phase 2 tested selectivity on 10 flavonoids, scoring each against both CD38 and PARP1 and computing the CD38−PARP1 score difference (Δ) as a selectivity index. 6 of the 10 were CD38-selective (CD38−PARP1 Δ > 0), favouring CD38 over the PARP1 anti-target. The cleanest case was quercetin, at Δ = +0.239 — a meaningful preference for CD38 over PARP1 in the computational scores.

Flavonoid leads are consistent with published CD38 inhibitor pharmacophores: the flavone core occupies the nicotinamide pocket, while the B-ring catechol engages the conserved glutamate in the catalytic site. The selectivity result is encouraging because flavonoids are notoriously promiscuous; finding a subset that computationally favours CD38 over a closely related NAD+-pocket target is exactly the kind of signal worth carrying into wet-lab confirmation.

Calibration: CD38 / 8D0M

The CD38 / 8D0M calibration was run on n = 10 compounds with published IC50 values. Boltz-2 scores were compared against the experimental potencies using Spearman rank correlation, producing ρ = 0.796 (p = 0.006). This is a strong correlation for a structure-based prediction pipeline and validates that the scoring function captures CD38 binding-affinity rank-order — the result that anchors CD38 as the most trustworthy target in this longevity panel.

The full compound list, SMILES, and raw scores are published in the lab notebook.

Lead Prioritisation

Our current lead list for CD38 centres on the selectivity-favoured flavonoids:

1. Flavonoid leads (quercetin and related flavonoids). These are well-characterised natural products with known safety profiles, oral bioavailability in rodent models, and published CD38-relevant activity (though most published data reports cellular NAD+ elevation rather than direct enzyme inhibition). We plan to obtain pure compounds and run in-vitro CD38 enzyme inhibition assays (fluorogenic NAD+ substrate) to confirm direct target engagement.
2. The 6 CD38-selective flavonoids from Campaign 40 P2. The compounds favouring CD38 over PARP1 (Δ > 0, quercetin best at +0.239) are prioritised for orthogonal validation (cellular NAD+ assay, binding confirmation) to rule out PAINS interference and confirm target engagement in a cellular context.

What's Next

The COCONUT R1 campaign is the first large-scale virtual screen on the EstimaBio pipeline, but it will not be the last. We have three immediate next steps:

• Wet-lab IC50 validation. We are scoping CD38 enzyme inhibition assays (fluorogenic NAD+ substrate) at a contract research organisation to test the prioritised leads. This would be the first experimental validation of a computational hit from the EstimaBio pipeline and will determine whether the predicted scores translate to real biochemical activity. Until then, every lead here remains in-silico.
• Expanded target panel. The 8-target panel is being expanded to include additional ageing-relevant targets: sirtuins (SIRT3, SIRT6), FOXO3, TFEB, and the mitochondrial uniporter MCU. Each new target will be calibrated at n ≥ 10 before entering the screening panel.
• COCONUT R2. A new release of COCONUT is expected in late 2026 with additional compounds, improved stereochemical assignments, and taxonomic annotations. We will re-screen the expanded library against the expanded target panel.

All results, including raw scores, calibration data, and PAINS-filtering logs, are published openly in the EstimaBio lab notebook. We believe that transparent computational drug discovery — where every threshold, every hit, and every failure is documented — advances the field faster than proprietary black boxes.

About EstimaBio

EstimaBio is the parent platform behind ToxScreen. The longevity-research program described in this post is a structure-based screening pipeline — a distinct engine from the ToxScreen tox panel, which runs calibrated ML ADMET (ADMET-AI). They share calibration discipline and a transparency ethos, but they are different methods: do not read the longevity results here as describing how a ToxScreen tox report is generated.

See the About page for the full platform story, or the Methodology page for pipeline details, calibration data, and known limitations.

Published by the EstimaBio / ToxScreen team. Questions, corrections, or collaboration enquiries: info@toxscreen.ai.