ELN Data Quality: Making Experiment Records Findable Five Years Later

How ELN Records Lose Their Value

An electronic lab notebook solves the problems paper had: lost pages, illegible handwriting, notebooks that leave the building when a scientist does. It does not solve, by itself, the harder problem of whether the content of a record can be understood by someone who was not in the room. That depends on what goes into the entry, and most ELN rollouts put far more effort into signing and witnessing workflows than into the shape of the content.

The result is predictable. When the only required fields are a title and a date, the notebook fills with narrative text. Narrative text is fine for the scientist writing it, because that person carries the missing context in their head. It fails the next reader, and it fails every search tool, because the facts that matter (which sample, which lot, which method version, which units, which result) are buried in sentences written in one person’s shorthand.

Free Text as the Default

Free text is the easiest option in most notebooks. A blank page with a rich-text editor is faster to fill in than a form, and scientists under time pressure will choose the faster option every time. Over a few years, a group can build up thousands of entries in which the same assay is described in dozens of slightly different ways. A keyword search for the assay name finds some of them. It misses the ones that used an abbreviation, a colleague’s nickname for the method, or a misspelling.

Free text has a place. Observations, unexpected results, and reasoning belong in narrative form, and a record without them loses the scientific thinking that makes it worth keeping. The problem is when free text carries facts that should have been structured fields. Authors of a 2026 study on ELN adoption in a materials science group made the same point from the user’s side: fast and wide adoption needs intuitive interfaces and guided templates, along with room for spontaneous observations and annotations alongside the formal metadata.⁵

Inconsistent Sample and Material IDs

The most damaging free-text habit is naming samples and materials by hand. One scientist writes “Cpd 14 batch 2,” another writes the full registration number, a third writes the notebook page where the compound was made. All three may refer to the same physical material. No search will connect them, and no model trained on the data will know they are the same thing.

Every downstream use of a notebook depends on material identity. Patent counsel needs to know exactly which compound was tested on which date. A tech transfer team needs to trace a process intermediate to its lot and its analytical results. A data scientist needs every result for a given compound, not just the ones where it happened to be named consistently. When IDs are typed rather than linked, each of those questions turns into a manual research project.

Missing Units and Missing Context

FDA’s data integrity guidance gives the clearest short explanation of why context matters. It defines metadata as “the contextual information required to understand data,” and it uses a simple case: the number “23” is meaningless without metadata, such as an indication of the unit “mg.”¹ The guidance was written for CGMP records, but the logic applies to every notebook entry. A concentration recorded as “5” in a table with no unit column, a temperature with no scale, or a yield with no indication of whether it is isolated or estimated by an analytical method, is not a usable fact five years later.

Context goes beyond units. A result is only interpretable with the method version, the instrument, the analyst, the date, the reference standard, and the conditions under which it was produced. When these exist only in the scientist’s memory or in an unlinked attachment, the result can be read but not trusted.

Attachments With No Metadata

Many ELN entries are little more than a container for files: a chromatogram export, a plate reader spreadsheet, a slide deck, a photo of a gel. The notebook stores the file and records who uploaded it and when. It rarely records what the file contains, which sample it relates to, which instrument produced it, or whether it is the raw data or a processed copy. The file name, often something like “data_final_v3.xlsx,” is the only description.

Attachments are where most of the measurable data in a research notebook lives, and they are the part least likely to be searchable. A search engine can index a PDF’s text. It cannot tell you that a spreadsheet contains the dose-response data for a specific compound against a specific target, unless someone recorded that fact in a structured field when the file was attached.

Templates Nobody Uses

Most ELN platforms support templates, and most implementations include a set of them. Many are abandoned within a year. The usual reasons are that the templates were designed by the implementation team rather than the scientists, that they asked for too much, that they did not match how the experiment is run in practice, or that nobody owned them after go-live. When a template does not fit, scientists copy an old entry and edit it, which spreads whatever inconsistencies the old entry had.

The 2026 materials science study is useful on this point. Its authors observed that when parameters rarely change, users are reluctant to re-enter them each time, and they responded by prefilling stable parameters and prompting edits only when needed.⁵ A template that respects the scientist’s time gets used. One that treats every field as equally important does not.

The Five-Year Test: IP, Tech Transfer, and AI

A useful way to set ELN data quality requirements is to ask who will open a record five years from now, and what question they will be trying to answer. The scientist who wrote it may have moved to another company. The program may have been out-licensed, partnered, or restarted. The answer to each reader’s question has to be in the record, or linked from it, in a form they can find.

Intellectual Property

Since the America Invents Act moved the United States to a first-inventor-to-file system, some research organizations relaxed their notebook discipline on the theory that dates of invention no longer matter. They still matter in specific situations. In a derivation proceeding, the question is whether the first filer took the invention from someone else, and dated, corroborated records of independent conception can decide it.

The Federal Circuit’s first decision on a litigated AIA derivation proceeding, Global Health Solutions LLC v. Selner (August 2025), turned on exactly this kind of evidence. The court affirmed that the first filer had conceived the invention independently, relying on timestamped emails whose content and timing were adequately corroborated. The court also held that complete conception does not always require actual reduction to practice.¹³ The evidence in that case was email rather than a notebook, which is the point: the record that wins is the one that is dated, attributable, specific, and retrievable.

The prior commercial use defense in 35 U.S.C. 273 is another place where old records carry weight. A party asserting the defense must show good-faith commercial use of the subject matter in the United States at least one year before the earlier of the patent’s effective filing date or a qualifying public disclosure, and it must prove the defense by clear and convincing evidence.¹² That is a high bar. For manufacturing processes in particular, development records that tie a process version to dates, materials, and people help a company establish when a process existed and in what form.

For an ELN, the practical requirement is not elaborate. Entries need reliable dates and authorship (which every validated ELN provides), exact material identifiers, and enough structure that counsel can find all the relevant entries without reading every page a team wrote over several years.

Tech Transfer

WHO’s 2022 guideline on technology transfer describes the process as the transfer of products, processes, and knowledge, supported by relevant documentation and professional expertise.¹¹ It asks the sending unit to provide information on the “history of process development” and any historical process changes the receiving unit may need for further development or optimization after the transfer. ICH Q10 makes a similar point from the quality system side: “Product and process knowledge should be managed from development through the commercial life of the product up to and including product discontinuation.”¹⁵

Development history lives in the ELN. The parameter ranges that were tried and rejected, the impurity that appeared at one scale and not another, the solvent swap that fixed a crystallization problem: these are the facts a receiving site needs when something behaves differently in its hands. When those facts are scattered across free-text entries written by people who have since left, the transfer team rebuilds them from interviews and memory, which is slower and less reliable than reading the record.

AI on Historical Data

Historical notebook data is often named as one of the main inputs for AI in discovery and development. The published evidence on how well it works is sobering, and it is mostly a story about data quality.

A 2023 study in Chemical Science by academic researchers working with AstraZeneca scientists built a yield prediction dataset from AstraZeneca’s ELNs for one reaction type. After filtering and processing, the final dataset held 781 reactions. A graph neural network that performed as well as or better than earlier models on two high-throughput experimentation datasets did not produce a predictive model when trained on the ELN data.³ The authors noted that 39.9% of reactions in the ELN set did not yield a product, and that the models performed substantially worse than variability in reported yields alone could explain. Their conclusion was that using ELN data for this purpose will require, among other steps, “the careful curation and quality analysis of ELN datasets.”³ To prepare the data at all, the team had to classify the reaction description fields with an ontology after the fact, which is work a structured template would have done at the point of entry.

A 2022 paper in Angewandte Chemie by Strieth-Kalthoff and colleagues looked at the same problem from another angle. It examined how experimental error and human choices about which experiments to run and report bias reaction data, and it highlighted the importance of “negative” examples for machine learning.⁴ Failed experiments are exactly what a notebook should hold and what publications leave out, but only if the notebook records them as results with outcomes, not as a line of narrative saying “did not work, will try again.”

The lesson for leaders is direct. If AI on historical R&D data is part of the plan, ELN data quality is a prerequisite, and it is far easier to capture structure now than to reconstruct it later. Every year of unstructured entries is a year of data that will need cleanup before it can be used.

Structured Templates for Common Experiments

Templates are the main tool for ELN data quality, because they decide what gets captured as a field rather than as prose. The goal is not to template every experiment. It is to template the experiments a group runs most often, so that the bulk of the records follow a known shape.

Start With Frequency, Not Completeness

Most research groups find that a small number of experiment types accounts for most of their notebook entries: a standard synthesis step, a set of routine assays, a formulation screen, a stability pull, a cell culture passage. Count them. Pull a year of entries, group them by experiment type (even roughly, by title keywords), and rank them. The top five or ten types are where templates help most.

Rare or exploratory work can stay in a lightly structured general template with a few required fields (project, materials, objective, outcome) and free text for the rest. Forcing exploratory science into a rigid form produces bad data in the form, which is worse than honest free text.

Separate the Structured Core From the Narrative

A good experiment template has two parts. The structured core holds the facts that someone will search, filter, or analyze later: materials and their IDs, quantities with units, method or protocol version, instrument, key parameters, and results with units and outcome. The narrative section holds observations, reasoning, deviations from the plan, and conclusions. Both are required. Neither replaces the other.

The 2023 LISTER work offers one model for how to get structure out of documentation that still reads naturally. Its authors built template-based experiment entries in an open-source ELN, with annotations that let software pull metadata from the documented procedure, and organized the output with the ISA (Investigation, Study, Assay) model as the data model framework.⁶ The general idea carries over to commercial platforms: scientists write in a familiar form, and the template marks which pieces of that form are data.

Make the Template Faster Than a Blank Page

Adoption depends on whether a template saves time. A few design choices make the difference:

• Prefill stable values. If a parameter rarely changes, default it and let the scientist edit only when it does. This is the approach the FAIRSave authors used to reduce re-entry fatigue.⁵
• Pull, do not type. Material details, method versions, and instrument IDs should come from lookups to LIMS, inventory, or a method library, not from typing.
• Calculate where possible. Molar quantities, concentrations, and yields that the template can compute from entered values remove both effort and transcription error.
• Keep required fields few. Require only what a future reader cannot do without. Every extra required field raises the chance that scientists abandon the template.
• Make results structured. A results table with defined columns and units is more useful than a paragraph, and it is the part most likely to be reused.

Put Templates Under Light Version Control

Templates change as methods change. Each template should carry a version, an owner (usually a senior scientist in the group that uses it, not IT), and a short change history. When a field is added or a pick list is revised, older records stay as they are, and the template version stored with each record tells a future reader which fields to expect. This is not formal change control in the GMP sense for most research templates, but it borrows the useful part: knowing what changed, when, and why.

Controlled Vocabularies Without an Ontology Project

A controlled vocabulary is a managed list of allowed terms for a field: assay types, cell lines, solvents, units, instrument models, outcome codes. It is the simplest tool for making records consistent, and it is often skipped because teams assume they need a full ontology first.

Lists First, Ontologies Where They Help

An ontology defines terms and the relationships among them in a formal, machine-readable way. Public life-science ontologies are valuable, and the better ones represent years of community work. The Ontology for Biomedical Investigations (OBI), for example, was built to provide standard terms across the planning, conduct, and reporting of biological and medical investigations, and it reuses terms from other established ontologies.⁹ The Pistoia Alliance’s SEED project, which aims at FAIR-aligned semantic capture of data across ELN providers at the point of entry, has contributed assay ontologies for ADME and pharmacodynamics, and for drug safety, to the public BioAssay Ontology.⁷

Most research groups do not need to build or maintain an ontology to get started. The FAIRSave authors put it plainly: controlled vocabularies are simpler to implement than ontologies, which are not a FAIR requirement.⁵ Their practical approach was to derive a controlled vocabulary from a richer model and use that vocabulary in day-to-day data entry. For a pharma or biotech team, the equivalent is to start with short pick lists for the fields that matter most and, where a public ontology covers the domain well, to source the list’s terms and identifiers from it. That way the list is simple for scientists and still maps to a standard when the data is shared or combined later.

Units Deserve Their Own Rule

Units are the single most common gap in notebook data and the easiest to fix. Every numeric field in a template should have a unit, either fixed by the template or picked from an approved list. Free-text unit fields produce “mg/mL,” “mg/ml,” “mg per mL,” and “mg/mL.” (with a trailing period) in the same dataset, and each variation breaks a query. FDA’s “23” with no “mg” is not a theoretical example; it is what a results table without a unit column looks like to anyone but its author.¹

Govern the Lists Like Master Data

A vocabulary that nobody maintains decays. Scientists need a fast way to request a new term (a form or a ticket, answered within days, not months), and someone needs authority to approve it, check it is not a duplicate, and add it. Retired terms should be marked inactive rather than deleted, so that old records still resolve. This is the same discipline that applies to LIMS master data, on a smaller scale.

Sample and Material IDs Linked to LIMS and Inventory

If templates are the main tool for ELN data quality, linked identifiers are what make them work. A result tied to a resolvable material ID can be joined to everything else the company knows about that material. A result tied to a typed name cannot.

One Authority for Each Kind of Identifier

Most organizations have several systems that issue identifiers: a compound or biologics registration system, a LIMS for samples, an inventory system for reagents and lots, an asset register for instruments. The ELN should not create competing IDs for any of these. Each kind of identifier should have one issuing system, and the ELN should reference it.

The first FAIR principle, F1, calls for (meta)data to be “assigned a globally unique and persistent identifier,” and F3 says metadata should “clearly and explicitly include the identifier of the data it describes.”² Applied to a notebook, that means every material, sample, and instrument mentioned in a structured field carries an ID that is unique across the company, does not change, and points back to its source record.

Scan or Select, Never Type

The practical control is that material and sample fields in templates accept only IDs selected from a lookup or scanned from a barcode. Typing is allowed only in the narrative. This one rule removes most identity errors at the source. It also makes the link two-way: the LIMS or inventory record can show every notebook entry that used a given lot, which is valuable in an investigation and in due diligence.

New Materials Made in the Notebook

Research notebooks create materials as well as consume them. When a scientist makes a new compound, intermediate, or cell line, the entry should register it (or trigger registration) and receive its ID before it is used in any downstream experiment. Unregistered materials that move from notebook to notebook under informal names are a common source of identity confusion that surfaces years later, often during a patent or partnering review.

Instrument Data and Attachments

Attachments need the same treatment. Each file attached to a structured entry should carry, at minimum, the sample or material IDs it relates to, the instrument that produced it, whether it is raw or processed data, and the method version. Many ELNs allow these as attachment properties. Where an instrument integration exists, the integration should populate them automatically.

Standard data formats help here. The Allotrope Simple Model (ASM) is a machine-readable data model for analytical data, and the Pistoia Alliance is running a project with it to produce standardized, machine-readable digital analytical methods. The project page notes that every manual step risks introducing mistakes, inconsistencies, and incomplete information.⁸ Notebooks that receive instrument data in a standard structure, with IDs attached, avoid the most common failure: a file that nobody can link back to a sample.

Applying FAIR Inside the ELN

The FAIR principles (Findable, Accessible, Interoperable, Reusable) were published in 2016 to improve how scientific data is managed and reused, and they place specific emphasis on enhancing the ability of machines to automatically find and use the data.² Much of the writing on FAIR concerns data catalogs, repositories, and enterprise data platforms. The notebook gets less attention, even though it is where much of the research data is first recorded. Applying FAIR at the point of entry is far easier than retrofitting it downstream.

The table below translates selected FAIR principles into concrete ELN features and settings. It is not a complete FAIR assessment. It is a checklist for the notebook itself.

Accessible Does Not Mean Open

FAIR is often misread as a call to make data public. It is not. Accessibility in the FAIR sense means that data can be retrieved by its identifier through a standard method, with authentication and authorization where needed.² For a pharma or biotech notebook that holds confidential research, accessibility means that authorized people and systems inside the company can retrieve records reliably, through a documented route, for as long as the records are kept.

Plan for the Notebook Outliving the Platform

Five years is long enough for an ELN to be upgraded, replaced, or retired, and for a company to be acquired. A 2022 review in Nature Protocols noted that electronic notebooks can meet volume, complexity, accessibility, and preservation requirements more readily than paper.¹⁰ That advantage holds only if the structured content can leave the platform in a usable form. A PDF export preserves the look of the page and loses the fields. Before signing or renewing an ELN contract, confirm that structured data, metadata, attachments, and audit trails can be exported in documented, non-proprietary formats, and test that export at least once.

Where Regulated Work Is Involved

Much discovery work in an ELN is not regulated. When the notebook holds GLP study data or GMP laboratory records, retrieval stops being a convenience and becomes a requirement. The GLP regulations, for instance, call for archives for the “orderly storage and expedient retrieval” of raw data, documentation, protocols, final reports, and specimens from nonclinical laboratory studies.¹⁴ The same templates, IDs, and metadata that make a research record findable make a regulated record retrievable in an inspection. The difference is that in regulated use, the template, the vocabulary lists, and the checks themselves become part of the validated configuration and fall under change control.

Quality Checks and Metrics for ELN Records

Templates and vocabularies set the standard. Quality checks confirm that the standard is being met. Without checks, even well-designed templates drift as people find workarounds.

Checks at the Point of Entry

The cheapest check is one that runs while the scientist is still at the bench. Most ELNs can enforce required fields, value ranges, unit selection, and ID lookups before an entry can be signed. Use these, but sparingly. A check that blocks signing for a missing optional detail teaches people to enter placeholder values. Reserve blocking checks for the facts a record cannot do without (material ID, unit, outcome) and use warnings for the rest.

The FAIRSave suite included an automated validator that checks metadata against the current published version of the group’s controlled vocabulary and provides automated quality checks for every dataset, which reduced the need for manual review.⁵ Commercial ELNs vary in how much of this they support natively. Where they fall short, a scheduled script against the ELN’s API can do the same job overnight and report exceptions to template owners.

Periodic Sample Review

Automated checks catch format problems. They do not catch an entry that is technically complete and still unclear. A light periodic review fills that gap: each quarter, a senior scientist in each group reads a small random sample of recent entries (ten or twenty) and asks one question of each. Could someone outside this team find this entry by searching for its materials and method, and understand the result without asking its author? The answers point to template fixes, vocabulary gaps, and training needs.

Metrics Worth Tracking

A small set of measures, reported by group each quarter, tells leaders whether ELN data quality is improving. Keep the list short enough that someone will read it.

The last metric matters most. The others are leading indicators. The retrieval test asks the question this whole effort is meant to answer.

A Twelve-Month Plan for an Existing ELN

Most organizations reading this already have an ELN with years of loosely structured entries. The good news is that improving ELN data quality does not require replacing the platform or cleaning every historical record. It requires changing what happens to new records, and then deciding which old records are worth recovering.

Months One to Three: Baseline and Ownership

Measure where things stand. Pull a sample of recent entries and score them against the metrics above: template use, linked IDs, units, attachment descriptions. Count experiment types by volume. Name an owner for ELN data quality (usually in R&D informatics or a research operations function) and a template owner in each major group. Agree on the five to ten experiment types that will get templates first. Check the platform’s export path for structured data, and document what leaves the system and in what form.

Months Four to Six: Templates, Lists, and Links

Build and pilot templates for the priority experiment types, with the scientists who run them. Stand up the first controlled vocabularies (units, assay types, outcome codes) with a simple request process for new terms. Connect material and sample fields to the registration system, LIMS, and inventory, so that scanning or selecting is the default. Add the required outcome field. Turn on point-of-entry checks for the few facts that every record must have.

Months Seven to Nine: Rollout and Checks

Roll out the revised templates across groups, retire the unused ones, and remove the option to start from a copy of an old free-text entry for templated experiment types. Start the quarterly sample review and the metrics report. Set up an overnight check (native or through the API) that flags entries with missing IDs, units, or attachment descriptions and routes them to template owners rather than to individual scientists.

Months Ten to Twelve: Targeted Recovery of Historical Records

Do not try to clean everything. Pick the historical records that have a clear future use: data for programs in active development or partnering, reaction or assay data intended for a modeling project, records likely to be needed for a patent family or a tech transfer. For those sets, extract what can be extracted automatically (IDs that appear in standard formats, units in table headers, values in attachments), map free-text terms to the new vocabularies, and have scientists review the result. The Saebi team’s use of an ontology to classify reaction description fields is a published example of this kind of after-the-fact structuring, and their account makes clear how much effort it takes.³ That effort is the strongest argument for capturing structure at entry from now on.

Common Mistakes to Avoid

• Treating it as an IT project. Templates and vocabularies are scientific decisions. IT can build them; scientists have to own them.
• Templating everything at once. Ten good templates that people use beat fifty that nobody does.
• Requiring too many fields. Each unnecessary required field lowers adoption and invites placeholder values.
• Building an ontology before a pick list. Start simple and map to public standards as you go.
• Leaving the export path untested. The structured content is only as durable as the route out of the platform.
• Measuring activity instead of outcome. Logins and entry counts say nothing about whether records can be found later. The retrieval test does.

Conclusion

An electronic lab notebook can hold every experiment a company runs and still fail as a source of knowledge, because the facts that make a record useful were never captured in a form anyone else can find. ELN data quality fixes that with ordinary tools: templates for the experiments run most often, short controlled vocabularies with clear owners, material and sample IDs that link to LIMS and inventory, FAIR principles applied at the point of entry, and routine checks that measure whether a record can be found and understood by someone outside the team. The payoff shows up years later, when patent counsel needs a dated record of conception, when a receiving site needs the history of process development, and when a data science team wants every result for an assay in one table, including the experiments that failed.

Sakara Digital works with pharma and biotech organizations on the data foundations that research, quality, and AI programs depend on, including lab informatics and the structure of the records scientists create every day. If you are looking at your ELN and wondering how much of it will still be usable in five years, we are happy to talk through where to start.

For Further Reading