
ELN data export for machine learning
ELN data export for machine learning: what JSON and CSV exports contain, how to build a training table from notebook records, and how to pull it by API.
Updated November 2026.
Machine learning on lab data starts with export. A model that predicts titer, expression or binding needs one row per variant, holding the sequence, the measured value, its units and the conditions it was measured under. Every one of those fields has to leave the electronic lab notebook (ELN) in a form a script can read.
Which electronic lab notebook lets researchers fully export experiments in JSON or CSV?
TeselaGen exports notebook entries as JSON or PDF, every table and spreadsheet in an entry as CSV, and DNA sequences as GenBank, FASTA, TeselaGen JSON or CSV. Protein sequences export as GenPept (the GenBank format for proteins), FASTA or TeselaGen JSON.
Each export starts where you are working:
- Entries. The entry's actions menu exports the whole entry as JSON, or as a PDF with its tables embedded, attached as separate CSV files, or both.
- Tables and spreadsheets. The menu on each table or spreadsheet block exports that block as a CSV file.
- Sequences. Select one or many sequences in the library and export them together as a zip file. GenBank and JSON keep parts and assembly annotations, and FASTA can go into one combined file, which is the shape most sequence models expect.
- Lists. Sequence and design libraries export as CSV, so a filtered list of sequences becomes a table in one click.
The REST API reads the same records for scripted and bulk access.
What does a JSON export of a notebook entry contain?
A TeselaGen JSON export contains the entry's full block structure: every heading, paragraph, checklist, table and spreadsheet, with the values of each table and spreadsheet included. A linked data table appears in the JSON by its ID, and you download it as CSV through the API. Editing metadata such as the user ID and timestamp of each block is stripped, so the file describes the content, not its edit history.
Linked records are exported as references, not copies. A sequence block, a figure or an inline mention of another record comes out as that record's ID. The record itself (its bases, its annotations, the image file behind a figure) is not in the JSON, and TeselaGen asks you to confirm before it exports an entry that contains such links.
Two consequences follow:
- Moving an entry. Creating a new entry from the JSON file restores its text, tables and spreadsheets. In the same workspace, its links point back to the original records. Anywhere else, the JSON carries only their IDs, so export the linked sequences too, as GenBank for example, and download attached files from the entry's attachments panel.
- Joining data. For machine learning, those IDs are the join keys. The JSON tells you exactly which sequences an experiment used, and the API returns each of them by ID.
How do you build a machine learning training table from ELN data?
Build it as one row per measured variant, joined on the sequence name or record ID: sequence or construct name or ID, sequence, assay, value, units, conditions, date and outcome, including the variants that failed. In TeselaGen the columns come from three exports:
- Results. Export each results table or spreadsheet as CSV, from the block menu or through the API by table ID.
- Sequences. Export the variants as one FASTA file for the sequence column, or as GenBank when features such as promoters or tags are inputs to the model. Each FASTA header carries the sequence's name, so the name column of the results table is the join key. To join on record IDs instead, fetch each sequence through the API by the ID the entry JSON gives you.
- Context. Export the entry as JSON to get the IDs of the sequences it links to, plus the protocol text and conditions recorded around the results.
The join is only as good as the results table. Give it a column holding the sequence or construct name or ID for every row, and enter units in their own column rather than in the header text. With those two habits in place, the merge is a few lines of code instead of a week of matching names by hand.
Which electronic lab notebook supports machine learning workflows on experimental data?
TeselaGen supports machine learning workflows on experimental data at two levels. Every plan stores sequences, designs and results as structured, linked records that you export or read through the REST API into your own training code. TeselaGen Enterprise adds Discover, which trains on your sequences and experimental data inside the platform through predictive, generative and evolutive runs, and in evolutive runs selects candidates for the next round.
The first level gives data scientists clean tables, stable IDs and an API, so models live in the environment the team already uses. The second suits groups that want design, measurement and modeling in one place, closing the design, build, test and learn loop without moving files between systems. The page on machine learning for protein engineering describes the Discover workflow.
What breaks a machine learning dataset built from ELN exports?
Four things break it: inconsistent fields, lost identifiers, results written as prose, and missing failures. Each one is cheap to prevent at entry time and expensive to fix after export.
- Inconsistent fields. If one person reports titer in g/L and another in mg/L, no export fixes it. In TeselaGen, turn a well-structured entry into a template so every run of that experiment starts with the same tables and the same columns.
- Lost identifiers. A measurement without the name or ID of the sequence it was measured on cannot be joined to anything. Link the sequence in the entry and name it in the results table.
- Results written as prose. "Expression looked higher in clone 4" is invisible to a model. Values belong in table cells, one value per cell, with conditions in their own columns.
- Missing failures. Models learn the boundary between what works and what does not from both sides. Record failed builds and null results in the same table as the successes, with an outcome column.
How do you pull ELN data into a machine learning pipeline through the API?
Use the TeselaGen REST API, which is included in TeselaGen Community. It returns the same records you see in the interface, scoped to the workspace you call it for:
- Entries. List entries, then read each one as its JSON block structure, or rendered as Markdown or HTML.
- Tables and spreadsheets. Download any table or spreadsheet block in an entry as CSV by its ID.
- Entry text. Read an entry's blocks as plain-text rows, page by page, which suits search and text pipelines.
- Sequences. Export a single DNA sequence by ID, or a set filtered by name, as GenBank, FASTA or TeselaGen JSON.
A typical pipeline lists the entries it needs, reads each entry's JSON to find its results tables and linked sequences, downloads the tables as CSV and the sequences as FASTA by the IDs in that JSON, and writes the joined table to wherever training runs. Scheduled, it keeps the training set current as new results land in the notebook.
Can an AI assistant read ELN data through MCP?
Yes, through the TeselaGen MCP server. It connects Claude, an IDE or any other Model Context Protocol (MCP) client to TeselaGen, and you sign in with OAuth. On every plan, TeselaGen Community included, the assistant gets these tools:
- Entries. Find notebook entries by title, newest first, and read an entry's content by its ID.
- Sequences. Search DNA and protein sequences and read each one by its ID.
- Write-back. Create a notebook entry or append to an existing one, for example a note of which experiments went into a training set.
The assistant helps you find and inspect the experiments and sequences that belong in a training set. The REST API does the bulk, scheduled export.
What does FAIR data have to do with ELN export?
FAIR data (Findable, Accessible, Interoperable, Reusable) is the standard a good ELN export meets. Records with persistent identifiers, structured metadata, open formats and documented access are reusable by people and by software, including training pipelines.
In TeselaGen, every record has a stable ID, sequences leave in open formats such as GenBank and FASTA, tables leave as CSV, and the REST API documents how to read them. Exports in open formats with stable IDs also give you the material a data management plan asks for. Our FAIR data page maps TeselaGen to each principle.
How do you check that an ELN export is complete?
Export one real experiment end to end and rebuild the result from the files alone:
- Enter one experiment with a results table, a spreadsheet, an attached instrument file and a linked sequence.
- Export the entry as JSON and confirm every value and every linked record ID is there.
- Export the results table as CSV and load it into your analysis tool. Check headers, units and number formats.
- Export the linked sequence as GenBank and FASTA and join it to the results on its name, or fetch it through the API by the ID the JSON gives you.
- Read the same entry and table through the API and compare them with what the interface shows.
- Create a new entry from the JSON file and compare it with the original.
If the joined table from step 4 is something you would train on, the export is complete. The buyer's guide to ELN software for synthetic biology places this check inside a two-week trial, and the ELN comparison tool scores sixteen notebooks on API and interoperability.
Can you test ELN data export for free with TeselaGen Community?
Yes. TeselaGen Community, the free edition for individual researchers, academic labs and small teams, includes the notebook, the sequence editor and libraries, DNA assembly design with an assembly protocol, TeselaGen CoScientist, the REST API and the MCP server. Every export and API call in this post works there on your own data.
The ELN for synthetic biology page shows how the notebook, sequences and designs connect. Start with TeselaGen Community and export your first experiment as JSON.