FoodProt Help & Guides

Browse Data, Ask FoodProt, Analytics and default exports contain plant-protein records only. Non-plant records that entered the database earlier are kept for provenance, marked out of scope and excluded from every workspace page by default.

Sign in with your e-mail address and password. Creating an account sends an activation link to that address; opening it finishes the account. Passwords are handled by the authentication service and are never stored or logged by FoodProt.

New accounts start as a Viewer. A Viewer account sees the 'Research workspace under development' status on Browse Data, Analytics and Export; this is not an error, and there is nothing to request at this stage.

A Contributor, Editor or Reviewer who sees an approval message is verified but has not been granted research access yet, and can ask a FoodProt administrator for it.

Administrators grant or revoke access, suspend or reinstate an account and change roles, always with a written reason. They cannot change their own account or remove the last administrator. Every access change is recorded permanently in an append-only audit trail that survives account deletion.

Access is enforced by database row-level security, not only by the pages: a Viewer account reads no research record through any page, export, download or direct database request. Contributors see approved records and their own submissions; pending, rejected and draft records are visible to the roles that review or manage them. The edit history is admin-only, and private reviewer notes never appear on any page or export.

Public pages (About, Help & Guides, News and a Dashboard that describes the property taxonomy) never show research row counts or aggregate totals.

For authorized members, Ask FoodProt scans the complete set of matching records in a deterministic order, paged to completion up to a documented 20,000-row safety cap. Statistics cover the entire matched set, not only the cards displayed, and mixed-unit records are never averaged together.

If the scan reaches the cap and more rows exist (checked by looking for row 20,001), the answer says the scan is incomplete and withholds counts, means, ranges, rankings and comparisons rather than summarizing a hidden subset.

Answers quote approved, in-scope records only. Records still in review, returned for correction, rejected or quarantined are never quoted. When no approved record matches, the answer says so instead of estimating a value; when only another concentration matches, it shows the closest one and says so.

One observation is counted once. When the database holds more than one stored copy of the same observation, the answer quotes it once and says how many copies were left out.

Each quoted value shows the result of its source's citation check, for example 'DOI verified' or 'DOI mismatch'. See Checking a citation's DOI below.

Ask FoodProt never reads documents you uploaded to My Sources. See the My Sources section below.

Hide fields groups the catalog as Source, Core sample, Treatment, Additives, pH and ionic environment, Gel preparation and storage, Measurement result, Statistics and extraction, and Review and metadata. Search the catalog, toggle whole sections, or press Reset view.

Protein is pinned by default so it stays visible while you scroll sideways. Pin more columns from a header menu; pinned columns sit side by side without overlapping, and pinning is stored in your browser only.

Row height can be Short, Medium, Tall or Extra Tall, and headers can wrap. Full screen expands the grid. Print view produces a page-break-friendly table with a timestamp and the active filters.

The Colour button tints the category or status pill and adds a row accent by property category, data status, quality flag, scope status or protein source. Reset colours restores the default.

Every measurement has two identities. The measurement name is the wording the source or contributor used. The canonical metric key is the stable identity that known synonyms resolve to, so 'Foam stability' and 'Foaming stability' are analysed as one property.

A name that cannot be mapped confidently is flagged for mapping review and left out of property-bound analysis rather than guessed.

Ask FoodProt answers and every export carry the provenance fields, so a number can be traced to its place in the source.

When you pick a property, the unit list shows its recommended units. FoodProt never replaces a reported unit: % w/v is not % w/w, and g water/g sample is not %.

Record the basis whenever the source states it. Dry basis and wet basis, w/v and w/w, and per g sample and per g protein are different quantities.

Original and stored values are kept apart. The original reported value and unit hold the source's exact wording (for example '3.2 ± 0.2' and 'g H2O/g protein'). The stored value, unit and basis are what analysis uses. FoodProt converts nothing itself: when a contributor converted a value, unit_converted is true and the original fields still hold the source's numbers.

Amino-acid composition is multi-analyte data: one row per amino acid, named in analyte_or_band, never a whole profile in one number. Free and protein-bound amino acids are never pooled, and an amino-acid score is only interpretable with its reference pattern and calculation stated in the method record. Nitrogen content, the nitrogen-to-protein factor, moisture and ash can be recorded alongside so a protein-basis value can be traced.

Example: one water holding capacity value may come from 10,000 ×g for 20 min at 4 °C, another from 6,000 ×g for 20 min, and a third from '3,000 rpm' with no rotor stated. They may also be reported per g sample or per g protein.

Each measurement can carry a method record: a controlled method family plus the conditions that decide comparability, such as protein source and extraction route, concentration and basis, buffer, salt, ionic strength and pH, temperature, time and mixing, centrifugation, instrument category and settings, replicate structure, and the exact calculation. Everything is stored as reported; FoodProt never fills in a default for a laboratory.

A method is 'reported' by default. 'Recommended' and 'validated consensus' are protocol statuses that only a reviewer or administrator can set on a FoodProt protocol record, and 'validated consensus' is refused until the protocol carries its version, scope, approver, validation design, participating laboratories, acceptance criteria and reproducibility evidence.

This method model is original to FoodProt. It does not reproduce any organisation's procedures, and no brand or instrument is required or recommended.

FoodProt is designed for traceable, reviewable, and reproducible scientific data management. Verify source passages, method conditions, units and statistical assumptions before drawing conclusions.

Quarantine is for data whose source evidence is not verified, such as a value no page, table or figure of the cited paper shows. A quarantined record cannot be approved by anyone. Releasing a quarantine approves nothing: the record then goes through the review queue like any other.

Rows are never deleted to correct the database. A wrong record is quarantined or returned for correction, and every change stays in the audit history.

Each observation appears once by default. The manifest (JSON, and the Export Notes sheet in Excel) records how many stored copies were left out, why, and which row was exported in their place.

Choose 'Export every stored row' to keep the copies; each is flagged with duplicate_of_measurement_id and must not be counted as a separate observation.

Every export includes observation_key and citation_verification_status, described in the data dictionary.

The identity is built from the article, record, property, value type, reported value, unit exactly as reported, basis and measurement condition, the analyte, the method, the sample, experimental unit, condition and series identifiers, the replicate number, the table, figure or page the value came from, and the study.

Analytics: every chart, count, mean, range, ranking, statistical test and saved analysis uses one row per observation. The dataset bar says how many stored copies were left out, and chart exports record the number in their manifest.

Model-ready datasets: copies are left out before any filter or split, so one observation can never be in both training and test. The manifest lists each copy left out.

Browse Data shows every stored row so reviewers can find the copies; a later copy is labelled 'Duplicate copy — in review'. The database refuses a new copy of an observation it already holds.

DOI verified: the registered title is identical, the year matches, the authors or journal match, and nothing else disagrees; or a reviewer confirmed the match.

DOI probable match - review: most fields agree but at least one differs, for example a word of the title, an author or the year. Records on the article cannot be approved until a reviewer confirms the citation.

DOI mismatch: the DOI is registered to a different work. Records on it cannot be approved until the citation is corrected and checked again, or a reviewer records a decision.

DOI not verifiable: the metadata service had no record for the DOI, could not be reached, or the stored DOI is not a valid DOI. Check the citation by hand. An outage never replaces an earlier result. 'DOI not yet checked' and 'No DOI' mean what they say.

A check never changes the stored citation. The registered metadata is kept separately and every check stays in the history. The DOI button always opens https://doi.org/ followed by the DOI; a stored publisher link is shown separately as 'Publisher page (stored link)', and a stored doi.org link that names a different DOI is flagged and not offered.

Reviewers and administrators can check a citation again, or record a decision with a note saying what they compared, under Data integrity review, Citations. A daily check re-examines the 25 least recently checked articles and never replaces a reviewer's decision.

The file goes straight to private quarantine storage. The server inspects the container and rejects legacy .xls, macro-enabled, encrypted, external-link and embedded-object files, then parses the stored bytes in an isolated process with time and memory limits.

The browser only proposes the column mapping. When the import is created, the server re-reads the quarantined file and builds every row itself.

Reviewers decide through the review queue and never edit scientific fields.

FoodProt looks for possible duplicates: the same file, DOI or normalized citation, the same record, the same property for the same material, treatment, unit, basis and method, and terms that differ only by case, punctuation, abbreviation or spelling. Nothing is merged automatically.

Reviewers see each candidate pair with field-by-field evidence and decide Keep, Link, Merge or Reject with a written reason on the Data integrity review page. Merges are administrator-only and reversible, and every decision is kept in an append-only log.

A reviewer or administrator decides each duplicate item with a reason kept in the audit trail: Keep both, Link (the rows are related, for example the same study reported twice), Not duplicates, or Merge (administrators only: nothing is deleted, references move, and the merge can be reverted).

Keep does not make two identical observations count twice. If two rows are separate observations, give one of them a distinguishing replicate number, sample id or experimental-unit id, and it will be counted separately.

The quarantine register shows what was quarantined and why, the evidence, the row as it stood, and who quarantined or released it. Reviewers quarantine a record, or its whole article, from the record page with Report or quarantine this record.

A reviewer releases a quarantine with a reason and the source evidence: the page, table or figure that reports the value. Releasing approves nothing; the record still goes through the review queue.

The record page has a Report link that fills in the record for you. Say what looks wrong and, when you can, where the source states the correct value.

Add record is the form path for a single measurement: source, protein system, treatments, additives, pH and ionic environment, gel preparation, the result with unit and basis, and the method.

Every correction is re-validated on the server and recorded in the audit history with the user and time.

Fixed charts such as solubility against pH are bound to their exact property, whatever the global filters say.

Incompatible units are never plotted on one axis. When a property's records span several units, each chart plots one unit group, states how many records were left out, and offers a Unit control. Statistics follow the same rule.

If no records match, every chart shows an empty state and its exports are disabled. The summary bar states how many observations, proteins and studies are in play.

Edit figure changes titles, axis ranges and scale, precision, size, fonts, markers, lines, legend, gridlines, background and colour palettes (including colour-blind-safe and greyscale options). Settings are remembered in your browser, never the data.

SVG is recommended for line graphs and scientific plots. PNG export asks for a physical width and a target resolution of 150, 300 or 600 ppi and shows the resulting pixel size. The PNG file has no embedded ppi/dpi metadata, so set the print size in your layout software.

SD describes the spread of values; SE = SD/√n describes the uncertainty of the mean; the 95% CI uses the Student-t critical value for n − 1 degrees of freedom (n = 3 uses t = 4.30). n counts database records, so a literature mean reported with n = 3 is one record unless the raw values exist as separate rows.

Group comparisons offer one-way ANOVA, Welch ANOVA and Kruskal-Wallis, and run only when their requirements are met. A test on identical values is shown as not estimable, and a pairwise comparison with an undefined standard error is shown as not estimable with the reason.

Independence is checked with experimental_unit_id. The same unit in different groups, or repeated within a group, withholds the independent-group tests; when unit IDs are missing, independence is reported as unverified and the tests are withheld. Literature data pooled across laboratories still violates independence, so treat p-values as descriptive screening.

Regression shows a confidence band only when the data are eligible; otherwise the fit is labelled descriptive. Cross-property correlations pair values within the same experimental record, and a figure built from protein-level averages says so.

Each observation is counted once: stored copies of an observation already counted are left out, and the dataset bar says how many. See How duplicates are counted.

Stored copies of an observation are left out before any filter or split, and the split refuses input that still holds one observation twice, so the same observation can never be in both training and test.

The export is not preprocessed: nothing is imputed, normalized or converted. The manifest records the filters, the split, the copies left out and a dataset fingerprint.

Training in FoodProt: on Model training, choose a property, its unit and basis, the features, the split by study, an algorithm (linear, ridge, lasso, elastic net, PCR, PLS, random forest, gradient boosted trees, SVR or a multilayer perceptron) and the hyperparameter values to search. Every combination is cross-validated by study; the result gives the test accuracy with bootstrap intervals, split-conformal prediction intervals, permutation importance, residual plots and the exact dataset (SHA-256), seed and settings. A job may use at most four minutes of one CPU thread, and its predictions are model outputs, never measurements.

Training elsewhere: export a model-ready dataset with a group-aware split, train in your own environment, and record the run in the model run registry. FoodProt stores that run's metadata, reported metrics and predictions; it never re-runs or re-scores it.

3D views and PyMOL: the structure lookup reads an entry's metadata from the RCSB PDB Data API. Show the 3D structure here draws it in FoodProt's own Mol* viewer (cartoon, ball-and-stick or surface), with the coordinates fetched by FoodProt's server from RCSB PDB and not stored. A browser without WebGL shows a message instead. The PyMOL script (.pml) runs in your own PyMOL installation, which downloads the entry itself.

GROMACS runs in FoodProt: when an administrator has connected a simulation gateway (for example on an SDSU cluster), an editor or administrator can request a run on Simulation runs from a PDB entry or a PDB file. The structure is checked first: water, ions, ligands and non-standard residues are removed, and a structure with a gap or a missing heavy atom is refused, never rebuilt. The run minimizes, equilibrates (NVT, then NPT, with restraints) and runs production on the gateway's CPUs. Its energies, temperature, pressure, density, backbone RMSD and radius of gyration are read by the database from the files GROMACS wrote, every file is listed with its size and SHA-256, and the final structure opens in the 3D viewer. One run per member at a time, ten a day.

GROMACS elsewhere, and AMS: describe the setup under Record simulation metadata and export the job manifest (JSON); whoever runs the job does so on their own hardware or cluster. AMS calculations always run elsewhere, because FoodProt has no AMS protocol and no licensed AMS installation. Values the record does not state are listed as unspecified, never filled in. Then import the results file: every output needs an explicit unit and its analysis method, and simulated outputs are never combined with experimental measurements.

An invitation is a one-time link, valid for a limited time, that only a verified account with exactly the invited e-mail address can accept.

Owners can change member roles, revoke access, transfer ownership, rename or delete the workspace. Every membership change is recorded with who made it, why, and the previous and new value.

Press Enter to ask and Shift+Enter for a new line. Find exact evidence and Extract data candidates read the passages that contain all of your words (write or between two words to accept either); put a phrase in double quotes to require it as written. Ask a question, Compare sources, Explain a table or sheet and Check my text rank the passages that match any of your words, and Summarise and Outline or synthesis start from representative passages of each selected source.

The Tools menu offers Ask a question, Summarise, Compare sources (two or more sources), Explain a table or sheet, Check my text and Outline or synthesis, plus Find exact evidence and Extract data candidates, which read only your sources and quote them exactly.

With an answer service configured by an administrator (named on the page, with where it runs and how long it keeps request text), the first six tools write their answer through it. It receives your question, up to three earlier exchanges of the same conversation and the passages retrieved for this question (with, in the FoodProt Database modes, the approved observations found for it), never whole documents; every statement it writes must cite one of them, and FoodProt removes any statement whose quotation, number or unit is not in the cited evidence. When the evidence is insufficient or contradictory, the answer says so instead of choosing.

Without an answer service, My Sources answers with exact quoted passages, extractive summaries built only from those passages, side-by-side comparisons and data candidates for your review; it writes no answer in its own words. Scanned pages without a text layer are read only when an OCR service is configured.

Follow-up questions keep the conversation's source selection; changing it is always your own action. A saved answer whose evidence was replaced or deleted is marked out of date.

Your questions and results are saved privately for you in the workspace and reappear when you return. New conversation starts an empty one; earlier ones stay in the Saved conversations list.

If a request fails, the page shows the reason and a reference you can quote when reporting it, with Try again.

The text counts against the same per-file size limit as an upload; the dialog shows the size as you type.

Paste only text you may keep in the workspace. Deleting the source removes it like any other source.

A ZIP with a nested archive, an encrypted entry, a program, a macro-enabled Office file, a symlink or a path that escapes the archive is rejected outright rather than partly imported; the upload report lists each entry's outcome.

Executables, encrypted or password-protected documents, macro-enabled files and file types not listed above are refused.

Charts, equations and images are not interpreted.

A source entered only as a title, DOI, citation or abstract is marked ABSTRACT ONLY, and every result states its evidence scope.

Documents are parsed by FoodProt's own server code and indexed with PostgreSQL full-text search. A passage or question leaves FoodProt only for a service an administrator configured - the answer, similarity-search or OCR service named on the page - and only what that request needs; never another workspace's text.

When sources disagree or state values in different units, the passages are shown side by side exactly as written. FoodProt does not convert, combine or average values across sources.

Search and extraction are text matching over what FoodProt could read: they can be wrong or incomplete, they are not medical or clinical advice, and they are not a substitute for reading the original article.

FoodProt is designed for traceable, reviewable, and reproducible scientific data management. Verify source passages, method conditions, units and statistical assumptions before drawing conclusions.

Open DOI/publisher page appears only for a source entered with a valid DOI. It opens the publisher's record of the article, not the file FoodProt read.

Values extracted from the text are candidate drafts for your review, not FoodProt data. Propose for review places them in the workspace's review queue; a candidate value becomes FoodProt data only after you accept it and take it through the normal template upload and review workflow.

Deleting a source removes its stored original, parsed text and index, and marks any earlier answer that cited it as out of date.

One question reads at most 24 passages, chosen from every selected source; the list of sources is paged, and "select all ready" is resolved by the database, so a collection of 1000 sources is never loaded into your browser at once.

Each account may process 1000 processed uploads per day and ask 200 questions per day.