Quantum chemistry and solvent screening
Density-functional calculations and COSMO-SAC rankings of solvents, with the full computational record and where measured solubility disagrees.
Read the record →Computational science & AI Proposal · nothing built Methods peer-reviewed Synthetic illustrations
Computational models would rank, flag and predict for scientists to test, and a larger model would follow only measured gains
How EUTECNIQ proposes to use computational chemistry, mechanistic models and machine learning on botanical data it has the right to use: provenance, testing, governance and the funding gate for each stage.
Status. EUTECNIQ has no trained model, proprietary dataset or benchmark result. Everything below is a proposed program; the evaluation charts use synthetic data to show how testing would work.
Related: why quantum chemistry for solvent screening · COSMO-RS versus COSMO-SAC (and against measured solubility) · digital lab
13 → 171PubMed records a year on machine learning for natural products, 2015 to 2025
6 → 191the same for herbal, botanical and medicinal-plant research
294papers in 17 fields affected by data leakage in one survey
37provenance fields each record would carry (proposed)
NCBI E-utilities counts, retrieved October 24, 2026 (queries on this page); Kapoor and Narayanan, Patterns 2023; provenance table on this page.
Overview
Three layers of computation would support decisions: physics-based chemistry, narrow predictive models, and later a shared multimodal model
Botanical Intelligence: models M1 to M5
A hybrid mass-transfer and machine-learning model for extraction yield, its test against simpler methods, and its month-24 stop rule.
See the program →Process engineering and scale-up
Transport, balance and scale-up models that set the physics a learned model would have to respect.
Open the calculators →Data would pass provenance and rights checks before any model; every output would go through human review, and some decisions never reach a modelProposed architecture
Text version: what EUTECNIQ and partners would do, and what EUTECNIQ does not claim
EUTECNIQ would
- Capture paired, authenticated data with provenance from the first sample
- Benchmark every model against simpler baselines on held-out botanicals
- Keep every quality and safety decision with qualified people and validated methods
Qualified partners would
- Provide independent test sets and blind evaluations
- Review data licensing and privacy terms (counsel)
- Supply compute under contract, where needed
EUTECNIQ does not claim
- Any trained model, accuracy figure or benchmark result
- A proprietary dataset or foundation model
- That any model output is accepted by FDA
- A research prediction≠a release or safety decision
- A benchmark score≠validation for a context of use
- A larger model≠a more accurate model
- Publicly visible data≠a right to train on it
- A pattern in training data≠a mechanism
Explanation, details and sources
Physics-based methods estimate molecular properties before any experiment. Narrow models, each with one target, would learn from EUTECNIQ's own designed experiments. A larger model that reads many data types together comes last, if it is justified at all.
Research activity
Machine-learning studies of natural products and botanicals are growing fast but remain under half a percent of machine-learning records in PubMed
PubMed records per year, 2010–2025: toggle a series or switch the scaleNCBI data, our query
Data table and query strings
| Year | Natural products | Herbal, botanical or medicinal plants | Dietary supplements | All machine-learning records |
|---|---|---|---|---|
| 2010 | 4 | 2 | 0 | 2,152 |
| 2011 | 7 | 2 | 0 | 2,455 |
| 2012 | 1 | 11 | 1 | 2,616 |
| 2013 | 4 | 7 | 0 | 3,093 |
| 2014 | 7 | 4 | 1 | 3,465 |
| 2015 | 13 | 6 | 4 | 4,145 |
| 2016 | 4 | 10 | 0 | 5,133 |
| 2017 | 13 | 14 | 5 | 6,891 |
| 2018 | 11 | 23 | 2 | 10,971 |
| 2019 | 26 | 14 | 4 | 16,677 |
| 2020 | 46 | 32 | 5 | 24,745 |
| 2021 | 60 | 66 | 6 | 36,434 |
| 2022 | 76 | 63 | 10 | 47,357 |
| 2023 | 87 | 65 | 11 | 51,840 |
| 2024 | 101 | 111 | 7 | 60,244 |
| 2025 | 171 | 191 | 26 | 78,401 |
- Natural products:
("machine learning"[tiab] OR "deep learning"[tiab] OR "neural network"[tiab] OR "neural networks"[tiab]) AND ("natural product"[tiab] OR "natural products"[tiab]) - Herbal, botanical or medicinal plants:
("machine learning"[tiab] OR "deep learning"[tiab] OR "neural network"[tiab] OR "neural networks"[tiab]) AND (herbal[tiab] OR botanical[tiab] OR "medicinal plant"[tiab] OR "medicinal plants"[tiab]) - Dietary supplements:
("machine learning"[tiab] OR "deep learning"[tiab] OR "neural network"[tiab] OR "neural networks"[tiab]) AND ("dietary supplement"[tiab] OR "dietary supplements"[tiab]) - All machine-learning records:
("machine learning"[tiab] OR "deep learning"[tiab] OR "neural network"[tiab] OR "neural networks"[tiab])
Natural-product records rose from 13 in 2015 to 171 in 2025 and herbal or botanical records from 6 to 191, still about 0.2% of the 78,401 machine-learning records.
Data types and provenance
Every record would carry its lot, method, instrument and license, because a model can only be as traceable as its data
37 provenance fields across eight record types: select a type to see what each record would carryProposed record schema
Spectra carry the most context: a mass spectrum without its instrument, ionization mode and collision energy cannot be compared with another.
Show the table (8 rows)
| Data type | What it describes | Provenance that would be recorded | Main pitfalls |
|---|---|---|---|
| Molecular structures | Constituents as atoms and bonds | Source database and version, or the lot and method that elucidated the structure; stereochemistry confidence | Unconfirmed annotations treated as facts; duplicate structures under different names |
| Mass spectra (MS/MS, high resolution) | Fragment masses and intensities | Lot ID, instrument, ionization mode, collision energy, acquisition method version, processing software and parameters | Spectra from different instruments and energies are not directly comparable |
| NMR spectra | Chemical shifts, couplings, integrals | Lot ID, solvent, field strength, temperature, referencing, processing parameters | Solvent and referencing differences shift peaks |
| Chromatograms (UV, MS, HPTLC) | Signal against retention time | Lot ID, column and mobile phase, method version, system suitability result, integration settings | Retention times drift between columns and laboratories; alignment is required |
| DNA barcodes | Species-level identity of the source plant | Voucher specimen, marker region, primers, sequencing run, reference library and version | Processed extracts may contain little intact DNA; DNA does not show chemistry |
| Process sensor data | Temperatures, pressures, flows over time | Batch record ID, equipment ID, sensor calibration status, time base | Uncalibrated sensors; time stamps out of sync |
| Quality and stability results | Assay, impurities, physical tests over time | Lot, storage condition, pull point, validated method ID, analyst review | Results below the quantitation limit handled inconsistently |
| Literature and regulatory text | Published findings, monographs, guidance | Citation, version and date, license of the full text | Publication bias; retracted papers; text licensed for reading but not for training |
Explanation, details and sources
Provenance is the documented origin and processing history of a record. Without it, a prediction cannot be traced back to evidence, a data error cannot be found, and a licensing problem cannot be fixed.
Records generated in regulated work would be kept under 21 CFR Part 11 controls where Part 111 requires the record (21 CFR 111.605(c)): validated systems, limited access and secure, time-stamped audit trails.
Integration
Structures, spectra, chromatograms and text would meet in one representation through records of the same authenticated lot
Proposed multimodal model: encoders, shared representation, task heads and the evidence-retrieval pipelineArchitecture schematic
Show component details
| Component | What it does | Status |
|---|---|---|
| Molecular structures → Message-passing graph network | Constituents represented as graphs: atoms are nodes, bonds are edges. Sources: COCONUT, LOTUS, PubChem, ChEMBL and EUTECNIQ's own elucidated structures. Message passing over atoms and bonds learns a vector per molecule (Gilmer et al. 2017, ICML). | Proposed |
| MS/MS spectra → Spectrum encoder | Fragment ions (m/z) and intensities from LC-MS/MS and LC-HRMS. Public sources: MassBank, GNPS/MassIVE; plus EUTECNIQ's authenticated reference lots. Encodes fragment peaks; published tools already learn structure information from MS/MS (Dührkop et al. 2019, SIRIUS 4). | Proposed |
| NMR spectra → NMR peak-list encoder | ¹H and ¹³C chemical shifts, multiplicities and couplings; used for structure confirmation and whole-extract profiles. Encodes peak lists or binned spectra so they can be compared with structures and other measurements of the same lot. | Proposed |
| Chromatograms → 1-D signal encoder | Retention-time signals (UV, MS, densitometry) used for fingerprints, marker assays and batch comparison. One-dimensional convolutions or attention over chromatographic traces; aligned retention times are a prerequisite. | Proposed |
| DNA barcodes → Sequence encoder | Short standardized gene regions for species identification; ITS2 was validated for medicinal plants by Chen et al. 2010. Reference sequences: GenBank, BOLD. DNA identifies the source species, not the chemistry of a processed extract. Encodes barcode sequences as k-mer profiles or token sequences for species-level identity. | Proposed |
| Process sensor data → Time-series encoder | Sensor histories from extraction, evaporation and drying runs, linked to batch records and lot results. Summarizes run histories (heating, solvent ratio, drying) for yield and quality prediction. | Proposed |
| Literature and rules → Language-model text encoder | PubMed citations and licensed full texts, pharmacopoeial monographs, FDA guidance and regulations; also the source of the knowledge graph. Turns passages into vectors for search and for linking entities (species, constituents, studies) into the knowledge graph. | Proposed |
| Shared representation | One vector per lot; paired records of the same lot aligned by contrastive training; applicability-domain check before any answer. | Proposed |
| Identity and adulteration screening | Flags lots that look inconsistent with authenticated references. A flag triggers confirmatory testing; it never releases or rejects a lot. | Proposed |
| Property prediction | Predicted physicochemical properties with prediction intervals, or an abstention outside the applicability domain. | Proposed |
| Extraction-yield prediction | Learned features passed to the mass-transfer hybrid model on the Botanical Intelligence tab, which remains the baseline to beat. | Proposed |
| Formulation-stability prediction | Ranks formulations for real stability studies. A prediction is not stability data and never sets a shelf life. | Proposed |
| Interaction flags | Highlights constituents with reported CYP or transporter effects for expert review. Flags are hypotheses for testing, not conclusions. | Proposed |
| Evidence retrieval for dossiers | Starts the retrieval pipeline: find, draft, verify citations, then human review. | Proposed |
| Scientific knowledge graph | Links species, plant parts, constituents, studies, hazards and regulations, built from LOTUS/Wikidata, PubMed, monographs and FDA documents, with the source of every link stored. | Proposed |
| Retriever | Selects passages and graph facts relevant to the question (retrieval-augmented generation, Lewis et al. 2020). | Proposed |
| Language model | Drafts an answer conditioned on the retrieved passages; it is not asked to answer from memory. | Proposed |
| Citation verifier | Matches every sentence to the passage it cites; sentences without support are removed and logged. | Proposed |
| Cited draft | A draft with sources attached, handed to a qualified scientist. It is never filed or published without review. | Proposed |
| Human review and QA sign-off | Every output reviewed by a qualified scientist; regulated records also need quality-unit approval; audit trail of inputs, model version and output. | Proposed policy |
Explanation, details and sources
Each data type has its own encoder. Paired records of one lot are trained to land close together (contrastive alignment), so a spectrum can retrieve its likely structure and a chromatogram its likely species. Text enters through retrieval: the model answers only from retrieved passages, and every sentence is checked against its source.
Why one shared representation. The same sample can be measured many ways. If a spectrum, a chromatogram and a barcode from one authenticated lot are aligned, a model trained on abundant data of one type may help where another type is scarce. That benefit is plausible but unproven for botanicals, which is why the plan tests it against narrow models before scaling.
Training data and rights
Large public resources exist, but each has its own terms, and none contains paired, authenticated measurements of the same lots
Record counts were checked on October 22, 2026 against each resource's own page or latest database paper. The units differ, so the bars show scale, not comparable quantities.
Public scientific resources by number of records (log scale)Verified counts · Oct 22, 2026
Terms from: NCBI website and data usage policies; GenBank overview; ChEMBL interface documentation (general questions); COCONUT About page; Rutz et al. 2022 (LOTUS); NPASS home page; Neumann et al. 2025 (MassBank); BOLD citation page. All accessed October 22, 2026. This is a summary, not legal advice.
Own data, customer data and datasheets (also shown in the data-architecture figure above)
Most valuable data
EUTECNIQ's own paired measurements
Spectra, chromatograms, barcodes, process histories and stability results from the same authenticated lots, captured in LIMS and electronic notebooks from day one. Public data cannot supply these pairings.
PlannedCustomer data
Only under contract, never by default
Customer samples and results would be used for model training only with written permission, segregated by contract, and excluded from shared models unless the agreement says otherwise.
Policy proposalDocumentation
A datasheet for every training set
Each dataset would record its origin, collection method, license, known gaps and intended uses, following the datasheets-for-datasets practice (Gebru et al. 2021).
Practice peer-reviewedExplanation, details and sources
What each source would contribute, and the terms that govern it
Terms are quoted from each provider's own pages. Where a provider's page does not state terms, the table says so; those sources would be reviewed by counsel before any commercial training use.
| Source | Contents | Terms of use (as stated by the provider) | Proposed use in the model |
|---|---|---|---|
| PubChem | Chemical structures, identifiers and deposited bioassay results | NCBI places no restrictions on its own data, but PubChem incorporates depositor material that may be protected by copyright; reproduction and commercial use must follow the holder's terms | Structure pretraining; identifier mapping |
| ChEMBL | Bioactivity data curated from the literature | Creative Commons Attribution-ShareAlike 3.0 Unported. Property calculations derived from commercial software should not be extracted in isolation to train models intended to replicate the commercial process | Activity and target context for interaction flags |
| COCONUT | Open collection of natural-product structures with names, organisms and organism parts | Curated dataset under CC0 (public-domain dedication); software under MIT license | Natural-product chemical space |
| LOTUS | Structure–organism pairs with literature references, hosted on Wikidata | CC0, per the LOTUS paper | Species–constituent links for the knowledge graph |
| NPASS | Natural-product activities, species sources and composition data | Not stated on the home page; confirm before commercial use | Candidate activity context, after terms are confirmed |
| GNPS / MassIVE | Public MS/MS datasets, spectral libraries and molecular networks | Not verified here; confirm per library and dataset | Spectrum-encoder pretraining, after terms are confirmed |
| MassBank | Curated reference mass spectra | Every record carries explicit license information | Reference spectra filtered by record license |
| BOLD | DNA barcode records and specimen data | Site states "© BOLD Systems. All Rights Reserved" and gives record-level citation formats; confirm before use | Barcode reference set, after terms are confirmed |
| GenBank | Nucleotide sequences, including ITS2, rbcL and matK markers | NCBI places no restrictions on use or distribution; some submitters may claim patent, copyright or other rights | Barcode reference sequences |
| PubMed | Citations and abstracts of biomedical literature | NLM does not claim copyright on abstracts; journal publishers or authors may | Retrieval index with links to sources; full texts only where licensed |
Evaluation
A model is judged on botanicals it has never seen, against independent benchmarks, and it must say when it does not know
Random split versus held-out botanicalsIllustration
Applicability domain: predict inside it, abstain outside itSynthetic data
Calibration: does a stated 80% confidence come true 80% of the time?Synthetic data
Coverage: do prediction intervals contain the true value as often as promised?Synthetic data
Method, assumptions and limits: applicability-domain illustration
- What it calculates
- For 14 synthetic query points, whether each lies inside the applicability domain, defined as a distance to its fifth-nearest training sample below the threshold.
- Equations
- Euclidean distance in the two-dimensional projection; threshold = 0.25 + 1.35 × slider/100 (synthetic units).
- Assumptions
- 108 synthetic training points from three Gaussian clusters (fixed seed); a real model would measure distance in its own learned representation.
- What it does not establish
- It shows how abstention works; it says nothing about any real model's domain or accuracy.
- Verification
- Independent Python reference (scipy k-d tree nearest neighbors) in
tests/numerics/ref/technologies-computational-science-ai.py.
Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns. 2023;4(9):100804. Schymanski EL, et al. Critical Assessment of Small Molecule Identification 2016: automated methods. J Cheminform. 2017;9:22. Sheridan RP. Time-split cross-validation as a method for estimating the goodness of prospective prediction. J Chem Inf Model. 2013;53(4):783–790. Wu Z, et al. MoleculeNet: a benchmark for molecular machine learning. Chem Sci. 2018;9:513–530. Gebru T, et al. Datasheets for datasets. Commun ACM. 2021;64(12):86–92. Mitchell M, et al. Model cards for model reporting. FAT* 2019:220–229.
Explanation, details and sources
Random splits let the same species, lot or replicate appear in training and test sets, which inflates scores. Held-out botanicals, independent test sets, a defined applicability domain and calibrated uncertainty are the minimum before any output informs a decision.
| Risk | What goes wrong | Proposed control |
|---|---|---|
| Data leakage | Information from the test set reaches training: duplicates, replicate injections of one lot, the same species, preprocessing fitted on all data, or future data used to predict the past | Split by species, lot and time; fit preprocessing on training data only; leakage checklist before every evaluation (Kapoor and Narayanan 2023) |
| Weak benchmarks | A model is compared only with itself or with a straw-man baseline | Compare with the strongest simple baseline (library search, PLS chemometrics, gradient boosting, the mechanistic model) and, where they exist, public challenges such as CASMI for compound identification |
| No independent test | The team that trained the model also chose the test set | Blind test sets assembled or held by a partner laboratory; prospective tests on lots that arrive after the model is frozen (time-split validation, Sheridan 2013) |
| Bias | Training data over-represent some species, regions, suppliers, instrument vendors or published results | Report coverage by subgroup in the datasheet; report errors by subgroup in the model card; abstain outside covered groups |
| Drift | New suppliers, harvest seasons, method changes or instrument replacements move inputs away from training data | Monitor input distributions and performance on control samples; retrain or restrict under change control |
| Overconfidence | Stated probabilities or intervals do not match observed frequencies | Calibration and coverage checked on held-out botanicals, as in the charts above |
| Automation bias | People accept an output because a model produced it | Human scientific review of every output; disagreement logged; no output enters a regulated record without quality-unit approval |
Regulated decisions
Predictive research and regulated decisions would stay separate: a model can propose a test, never replace one
Risk-based credibility assessment: seven steps, with model risk set by influence and consequenceSteps verified · draft guidance
Explanation, details and sources
FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs and biologics sets out a seven-step credibility assessment; it excludes AI used in drug discovery or for operational efficiency. EUTECNIQ would adopt its steps as a design standard; adopting them does not make any output FDA-accepted.
| Decision | Who decides | The model's allowed role |
|---|---|---|
| Release or reject a lot | Quality unit, on validated test results | None; a flag can trigger an extra confirmatory test |
| Identity of a raw material | Qualified analyst, on validated identity methods | Screening flag only |
| Safety conclusion | Qualified toxicologists and experts | Evidence retrieval with citations, for expert review |
| Which experiments to run next | Scientists | Rank candidates with stated uncertainty |
| Shelf life | Stability data and the quality unit | Rank formulations for real stability studies |
Risk management
NIST AI Risk Management Framework 1.0
Published January 26, 2023 for voluntary use; organized in four functions: Govern, Map, Measure and Manage. A Generative AI Profile (NIST AI 600-1) followed on July 26, 2024. NIST states the framework is being revised.
VerifiedRecords
21 CFR Part 11 wherever FDA requires the record
Electronic records required under Part 111 must comply with Part 11 (21 CFR 111.605(c)): validated systems, limited access and secure, time-stamped audit trails. Model inputs, versions and outputs that enter such records would be kept the same way.
Regulation verifiedReporting
A model card for every released model
Intended use, out-of-scope uses, training and evaluation data, metrics by subgroup and known limitations, following Mitchell et al. 2019. Datasheets describe the data; model cards describe the model.
Practice peer-reviewed
Simple versus large
A larger model would be justified only if it beats the strongest simple method on held-out botanicals at equal data
Simple, well-understood methods often perform well on small scientific datasets and are easier to validate. A larger model earns its cost only by a measured gain.
All four conditions must hold before a larger model is built; any failure keeps the strongest simple methodProposed decision rule
Text version: when a larger model makes sense, and when it does not
When a larger model makes sense
- Several data types describe the same lots, and transfer between them helps where one type is scarce
- The gain over the best baseline holds as data grow and across held-out botanicals
- Uncertainty is calibrated and the applicability domain is defined
- Every training source has documented rights
When it does not
- A single data type and a simple baseline already meet the need
- Too few paired, authenticated records for held-out evaluation
- The decision it would inform cannot tolerate an unvalidated output
- Compute and governance cost exceed the value of the decision
Wold S, Sjöström M, Eriksson L. PLS-regression: a basic tool of chemometrics. Chemom Intell Lab Syst. 2001;58(2):109–130. No performance figures are claimed for any model on this page.
Explanation, details and sources
| Task | Strong simple baseline | What a larger multimodal model would have to show |
|---|---|---|
| Spectral identity screening | Library search; PLS discriminant analysis on spectra | Fewer misses on species absent from the library, with calibrated confidence |
| Property prediction | Gradient boosting or Gaussian processes on molecular descriptors | Lower error on held-out scaffolds and natural-product classes |
| Extraction yield | The mass-transfer model plus classical design of experiments | A measured gain over that hybrid on an unseen botanical (gate G1) |
| Evidence retrieval | Keyword and citation search by a trained scientist | Equal or better recall, with every sentence verified to its source |
Roadmap and funding gates
Each stage would be funded only after the previous one passes its gate on measured data; failing a gate stops the spending and keeps the data
Gates use measured results, not dates. Thresholds would be proposed to the board before each stage starts; none is set yet.
From one narrow model to a shared multimodal model, gate by gateProposal
What models would and would not do (also shown in the data-architecture figure above)
Would
Decision support with stated uncertainty
- Rank candidate botanicals, solvents and formulations for testing
- Flag possible adulteration for confirmation by a validated method
- Predict properties and yields with intervals, or abstain outside its domain
- Propose the next experiment for scientists to approve
- Retrieve evidence and draft dossier sections in which every sentence cites a source, for expert editing
Would not
Anything that replaces accountable people or validated methods
- Reach a safety conclusion or a "generally recognized as safe" judgment
- Release or reject a lot, or replace a validated identity or potency test
- Generate health, structure–function or marketing claims
- Decide a product's regulatory category or pathway
- Train on customer data without a written agreement
Stop rules carry over from Botanical Intelligence. If the first model shows no measured gain by month 24, AI spending stops. The same logic would apply at every later gate: no measured gain on held-out botanicals, no next stage.
Explanation, details and sources
| Stage | Spending released when | Evidence reviewed | Stop rule |
|---|---|---|---|
| M1 hybrid extraction model | Designed-experiment data from the laboratory exist | Gain over physics alone and classical design on an unseen botanical | No gain by month 24: AI spending stops; data kept |
| M2–M5 narrow models | M1 passes G1 and data rights are documented (G2) | Each model against its strongest baseline | Unclear rights: the source is excluded |
| Shared-representation pilot | Enough paired, authenticated records for held-out evaluation (G3) | Pilot against the narrow models, with calibration | No gain: keep the narrow models (G4) |
| Larger multimodal model | The pilot gain holds as data grow | Credibility plan for one defined context of use; QA approval; model card (G5) | Inadequate for that use: not used for that decision |
Discuss a data or modeling collaboration
Partners with authenticated botanical data, independent test sets or a defined prediction problem can describe it; EUTECNIQ would respond with a proposed evaluation plan and its limits.