EUTECNIQ Sciences

See every page on one mapArrow keys move through results · Enter opens · Esc closes

The platform · proposed

The Large Botanical Model: scientific AI for botanical ingredients.

A family of specialized, physics-informed models on one shared botanical data layer. It is designed to return testable predictions with their uncertainty, choose the experiments worth running, and assemble traceable evidence for safety, regulatory, process and formulation decisions. It is not a chatbot, and it is not one giant model.

In developmentNo Large Botanical Model has been trained. What exists: computed research on published data and illustrative simulators, one benchmark, and a rules-based NDI demonstration. The register below shows the status of each part.

01 · The architecture

Data in, predictions with uncertainty out, experiments back in.

Choose an application to see which inputs, models, outputs, experiments and decisions it uses, what evidence it must earn, and what exists today.

02 · Seven kinds of data

Seven kinds of scientific data, kept with their source and rights.

Each record keeps its source, its license or ownership, the method that produced it and its uncertainty. Customer data stay the customer’s and are not used to train shared models unless the customer agrees in writing.

The seven data classes of the Large Botanical Model
Data classWhat it coversWhere it would come fromStatus today
Chemical and molecularStructures, conformers, properties, solubility, thermodynamics, stability, interactionsPubChem, published measurements (for example NIST ThermoML), computedComputed for 3 molecules; screen of 13
BotanicalSpecies, cultivar, chemotype, origin, cultivation, harvest, processing, natural variabilityTaxonomic checklists, literature, supplier recordsPlanned
AnalyticalHPLC, LC-MS, GC-MS, spectroscopy, fingerprints, purity, identity, contaminants, batch consistencyPartner laboratories; published datasetsPublished data reanalyzed (ginkgo PCA)
Extraction and manufacturingSolvents, mass transfer, thermal behavior, kinetics, yield, selectivity, equipment, scale-up, mass and energy balances, exergyPartner pilot runs; literature; computationIllustrative simulator and 2D CFD
Formulation and deliverySNEDDS and SEDDS, dispersions, emulsions, encapsulation, complexation, precipitation, stability, dissolution, bioaccessibilityPartner laboratories; literaturePublished designs reanalyzed
Biological and safetyToxicology, ADME, PK/PD, PBPK, dose–response, transporters, interactions, human clinical evidenceLiterature; CRO studies where requiredPK calibrated on published human data
RegulatoryIdentity, manufacturing history, safety substantiation, intended use, jurisdiction, NDI and NDIN recordsStatute, regulations, FDA guidance and public NDIN recordsRules run in the NDI demonstration

03 · How the pieces combine

Specialized models on a shared data layer, not one model for everything.

Botanical data are scarce, mixed in kind and expensive to measure. A single large transformer trained on them would have little to learn from and would not respect mass balances or give calibrated uncertainty. The proposed design combines five kinds of component.

Representation

Molecular representations

Molecular graphs, descriptors and surface-charge (σ) profiles that let models compare compounds across plants.

Structure

Knowledge graph

Species, plant part, compound, lot, process, study and regulatory record as linked entities, so a question can follow the links.

Learning

Specialized machine learning

Narrow models for one property or decision each, trained only where enough rights-cleared data exist, and judged on held-out data.

Physics

First-principles simulation

COSMO-type solvation, molecular dynamics, mass transfer, CFD and PBPK, which hold where data are thin.

Decision

Uncertainty-aware optimization

Bayesian inference, multi-fidelity methods and active learning choose the next experiment by the information it adds.

Methods: Gilmer et al., Neural message passing for quantum chemistry, ICML 2017; Hogan et al., Knowledge graphs, ACM Computing Surveys 2021;54(4):71; Klamt, J Phys Chem 1995;99:2224 (COSMO-RS); Kennedy and O’Hagan, J R Stat Soc B 2001;63:425 (Bayesian calibration of computer models); Peherstorfer, Willcox and Gunzburger, SIAM Review 2018;60:550 (multifidelity methods); Jones, Schonlau and Welch, J Global Optim 1998;13:455 (expected improvement).

04 · One benchmark, with its uncertainty

With 15 runs, the model’s worst-case error was half that of the classic method.

A computed benchmark on the illustrative extraction simulator behind FIG 047: 1,296 settings, 50 random repeats, the same 15-run budget for each method. It tests the method, not a real process, and no laboratory data are involved.

Heat map of extraction recovery over temperature (40 to 80 °C, across) and hold time (15 to 240 minutes, up) from the illustrative simulator; the best setting is a star in the top-left corner; the 15 runs are marked as rings (first five) and dots (ten chosen by the model).
The simulator’s answer, with the 15 runs: rings, 5 space-filling starts; dots, 10 chosen by expected improvement; star, the best setting.
Parity plot of model prediction against the simulator for the 1,281 settings the model never saw, from 0 to 70% recovery on both axes; points lie close to the diagonal, with 95% intervals as faint bars.
Predicted against true recovery (0–70%) for the 1,281 settings the model never saw, with 95% intervals.

Typical error, recovery points

1.4 · this model
2.3 · classic fit

Worst error, recovery points

6.4 · this model
13.7 · classic fit

Uncertainty check

The 95% band covered a median 91% of unseen settings: somewhat overconfident, and stated as such.

Benchmark results, medians of 50 repeats
Method, 15 runsTypical error (RMSE)Worst errorFound the best setting95% band coverage
5 space-filling + 10 by expected improvement, Gaussian process1.4 points6.450 of 5091%
15 space-filling, Gaussian process1.07.450 of 5098%
15 space-filling, quadratic response surface (classic)2.313.7Within 0.02 pointsNone given

All three methods found the best setting, so the gain is a more accurate map and a usable uncertainty, not finding the peak. EUTECNIQ computation, November 15, 2026 (scikit-learn Gaussian process, Matérn 5/2 kernel); scripts kept for review. Recovery points are percentage points of intact recovery. Box and Wilson, J R Stat Soc B 1951;13:1 (response surfaces); Rasmussen and Williams, Gaussian Processes for Machine Learning, MIT Press 2006.

05 · Platform maturity register

What exists, what is being built, and what has not been proven.

A drawn architecture is not an operating platform. This register is updated at each stage gate; a level is marked only when the evidence for it exists.

Platform maturity register, November 19, 2026
ComponentExisting research or toolUnder developmentTrained and benchmarkedValidated on independent experimental dataUsed by customersRegulatory support capabilityEvidence today and next gate
Scientific research libraryReachedNot yetNot yetNot yetNot yetNot yet60 figures with data and scripts, each labeled by evidence type
Next gate: Kept current; every new figure labeled
Molecular modeling toolsReachedNot yetNot yetNot yetNot yetNot yetOpen-source conformers and force fields; COSMO-SAC screen of 13 molecules; MD set-up
Next gate: Benchmark against measured solubility (Phase 2)
Extraction and process modelsReachedNot yetPartlyNot yetNot yetNot yetSurrogate benchmarked on an illustrative simulator (50 repeats); Pareto and Monte Carlo design space; 2D CFD calculation
Next gate: Fit to partner-measured extraction runs (Phase 3)
Exposure models (PK)ReachedNot yetPartlyNot yetNot yetNot yetBayesian PK for curcumin, quercetin and berberine calibrated on published human data
Next gate: Test on held-out published datasets (Phase 2)
Pareto solvent screenReachedNot yetNot yetNot yetNot yetNot yetLive interactive screen on published NIST ThermoML measurements
Next gate: Extend to new botanical actives as data allow
NDI regulatory intelligence workflowReachedPartlyNot yetNot yetNot yetPartlyRules-based demonstration on this site: seven screening questions, FDA Appendix A table, safety calculations
Next gate: Test against public FDA response letters; first paid feasibility work (Phase 1)
Botanical knowledge graph and data pipelinesNot yetNot yetNot yetNot yetNot yetNot yetSchema and provenance rules designed; not built
Next gate: First release with source provenance (Phase 1)
Specialized predictive modelsNot yetNot yetNot yetNot yetNot yetNot yetNone trained on EUTECNIQ data
Next gate: Each must beat a stated baseline on held-out data (Phase 2)
Larger multimodal botanical modelsNot yetNot yetNot yetNot yetNot yetNot yetFuture research objective
Next gate: Only when data quality, validation, cost and demand justify it (Phase 4)
Owned ingredients and licensingNot yetNot yetNot yetNot yetNot yetNot yetFuture objective
Next gate: After validated client work and an IP assessment (Phase 5)

Reached Partly: on an illustrative simulator, published data or rules only Not yet

Register as of November 19, 2026. No component has been validated against independent experimental data or used by a customer. Regulatory support capability means a working, cited workflow, not a filed notification.

06 · The roadmap

Five phases, each with a gate to pass and a reason to stop.

Evidence comes before infrastructure: data rights, models, external validation and paying customers first; a laboratory only after demand and technical milestones are proven.

Phase 1

Months 0–6

Scientific data infrastructure and the first product

The botanical knowledge graph and source-data pipelines with recorded provenance and rights; the existing computational models connected to them; one focused regulatory-intelligence use case.

Stage gateSigned, deposit-paid feasibility work from at least two unrelated clients; the screening logic reproduces every cited rule on a test set of public FDA responses to notifications.
Validation requiredEvery record traceable to its source and license; rules reviewed by regulatory counsel.
Stop criterionNo paid demand by month 6: re-scope the offer; no laboratory spending.

Phase 2

Months 4–12

Specialized predictive models

Narrow models for selected molecular properties, safety-gap identification, extraction conditions and formulation-relevant properties, each benchmarked before use.

Stage gateEach model beats a stated simple baseline on held-out data, with its error and calibration published internally.
Validation requiredHeld-out and blinded test sets; error and interval coverage reported for every model.
Stop criterionA model that does not beat its baseline stays research and is not offered to clients.

Phase 3

Months 9–18

Commercial pilots with partners

The platform used on real client problems with qualified CROs and analytical laboratories, compared with the conventional workflow on documented metrics.

Stage gateRepeat paid work; at least one decision-ready dossier-support package; a measured, independently checked benefit in time, cost or experiments.
Validation requiredPredictions compared with blinded external measurements; time and cost logged against a conventional baseline.
Stop criterionNo measurable benefit: continue as a services business and do not scale platform spending.

Phase 4

After month 18, if justified

Larger multimodal botanical models

Shared representations and larger models across data types, built only when dataset quality, validation evidence, computing economics and customer demand justify them.

Stage gateA shared model beats the specialized models on held-out tasks at an acceptable cost per prediction.
Validation requiredData volume and rights criteria met before training; held-out comparison with the Phase 2 models.
Stop criterionIf it does not beat the specialized models, keep them; no larger model for its own sake.

Phase 5

When capabilities are validated

Proprietary ingredient development and licensing

Validated capabilities applied to selected EUTECNIQ-owned botanical ingredient programs, commercialized by license, qualified manufacturing partners or supply.

Stage gateValidated capability on client programs; intellectual-property and freedom-to-operate assessment; evidence of buyer interest.
Validation requiredThe same regulatory, safety and quality evidence any ingredient needs.
Stop criterionNo freedom to operate or no buyer interest: the program stops before spending on scale-up.

Phases overlap. Timing is the earliest proposed and may slip; each phase is funded only when the previous gate is met.

07 · Benchmarks to earn

No benefit is claimed until it is measured.

EUTECNIQ does not claim shorter studies, higher success rates, FDA acceptance, better bioavailability or lower cost. These are the studies designed to establish whether the platform delivers any of them.

Regulatory

Gap detection on public letters

Share of the identity and safety deficiencies cited in public FDA responses to notifications that the gap analysis flags in advance.

Prediction

Error against blinded measurements

Solubility, extraction recovery and droplet size predicted before partner laboratories measure them; error and interval coverage reported.

Experiments

Runs to an operating window

Experiments needed to locate a specified window, against a classic design of experiments on the same problem.

Time and cost

Candidate to decision

Elapsed time and cost from intake to a documented pathway decision, logged against the client’s conventional baseline.

08 · Where it fits

One platform, three ways it would earn.

Business 1

Enterprise platform

Subscriptions, research modules and API access for ingredient makers, brands, CROs and CDMOs. A future product.

Business 2 · first

Regulatory and development solutions

Model-assisted NDI assessment, dossier support, ingredient qualification and development programs, with partner laboratories and CROs.

Business 3

Owned ingredients and IP

EUTECNIQ-funded ingredient programs, licensed or supplied through qualified partners.

The Large Botanical Model is the core; each business uses it and, with permission, improves it.

EUTECNIQ Atlas, the free public guide on this site, is a separate program with its own governance. It is not the Large Botanical Model and makes no predictions. Information only; not an offer of securities.

Health information notice. These statements have not been evaluated by the Food and Drug Administration. Nothing on this site is intended to diagnose, treat, cure or prevent any disease, and nothing here is medical advice.

Read the full notice

The information is for general education only. It does not replace advice, diagnosis or treatment from a qualified health professional, or any medication a clinician has prescribed. Do not start, stop or change any medicine, supplement or herbal product because of something you read here. Talk to your doctor or pharmacist first, especially if you are pregnant or breastfeeding, have a medical condition, take any medicines, are preparing for surgery, or are caring for a child. Research findings, traditional uses and safety data are reported as their sources state them; they are not recommendations or claims about any product. If you think you may have a medical emergency, call your doctor or 911 at once.

EUTECNIQ Sciences sells no supplements or health products today and receives nothing from any product or ingredient mentioned on this site; any commercial interest it later has in an ingredient will be stated on that ingredient’s page.

Ready to start a development project?

Tell us about an ingredient you want to specify, verify or license. Questions? Email [email protected]