CRO data integration that catches a changed column or unit before a single value loads
CRO deliverables become clean, cited records within hours, never a wrong unit.
See one case, screen by screen ↓Getting the results your contract research organisations send — PDF reports, Excel tables, instrument exports — into your own research data warehouse as clean, consistent records. CRO Data Integration does it with seven agents that classify each deliverable, check its layout, extract and convert every value, test it against the contract and load it with a citation, while the study monitor decides every exception.
How to choose AI solutions for life sciences →
Updated
A discovery sponsor sends its ADME, DMPK, pharmacology, safety and screening work to six or more CROs, and each sends back its own PDFs, Excel workbooks and instrument exports, in its own layout and its own units. Scientists and data stewards open every file and re-key the values into the house schema — a 24-compound microsomal stability study alone is 96 values — and check controls against the statement of work by eye.
The risk is in what nobody sees. A CRO marks its verapamil control "Pass" against its own SOP range, which is wider than the contract. A CRO upgrades its instrument software and a header changes from ng/mL to µg/mL; a copy-paste or a fixed import script loads every value a thousand times too low, and the clearance, exposure and ranking decisions built on it are wrong with it. The sponsor stays accountable for the data either way.
Estimated split for a typical in-vitro ADME deliverable, from receipt to records in the warehouse, by hand and with the solution.
Seven agents classify, compare layouts, extract, normalise and check every file against its contract; the study monitor decides each stop.
Tuesday 7 October, 09:40. Quillon Bioanalytical drops QB-1180_PK_plasma_conc.xlsx on SFTP: LC-MS/MS concentrations of NWB-5104 in mouse plasma after an IV and an oral dose, 9 samples, under contract NWB-SOW-047 §2. Here is what happens next, screen by screen, in the working solution.
Priya monitors the studies Northwind Bio runs at six CROs. Her intake conveyor has a lane for each — Halden, Corrib, Pellucid, Sorrel, Quillon, Larkspur — and four stations: Received, Mapped, QC-checked, Loaded. This week: 38 deliverables, 30 loaded, 1 ready to load, 4 stopped for a decision, 1 waiting on the CRO, 2 on the way in. Quillon's new file sits at Received.
Drift watch: “Before a single value is mapped, each file is compared with its CRO's last accepted layout — headers, column order, units and value ranges.”
The deliverable classifier matches the workbook to Quillon, study QB-1180 and contract NWB-SOW-047 §2. The template drift checker compares its header with Quillon template v3, last used on QB-1174 on Sep 29, and finds four changes. Sample became Sample ID and Flag became BLQ — matched by meaning. Route and Time (h) swapped places — matched by name. Column E went from “Conc (ng/mL)” to “Conc (µg/mL)” — and that stops the file.
The template drift checker never accepts a unit change on its own.
The diff shows what an import would have done: M1-0.25 IV reads 1.284 and would load as 1.284 ng/mL instead of 1,284 ng/mL; M4-2 PO would load as 0.451 instead of 451. The last three Quillon runs at 0.25 h read 1,190–1,410 ng/mL, so the values agree with the new header. Priya chooses “Convert × 1,000 and save template v4” — the rule is applied, template v4 is saved and Sam Patel, the template owner, is informed.
Her other choices on the same screen: “Query Quillon” or “Hold the file”.
The rest of the intake runs. The result extractor reads 9 samples, IV and oral, each with its cell in column E; the unit and identity normaliser converts the values to ng/mL and matches NWB-5104 through the shipped manifest. The study-spec checker stops at two samples, M1-24 and M4-24 at 24 h, that the export writes as 0 with flag BLQ. Loading 0 would pull exposure down.
Suggested: “Load both as "below LLOQ (1.0 ng/mL)" — never as zero.”
Priya loads both as below LLOQ. The inspector now reads: “Every check passed or is decided. 9 values mapped to house units, each citing its cell or page.” Both decisions sit on the file with her name — the schema drift converted × 1,000 with template v4 saved, and the two samples loaded as below LLOQ.
One click on “Load 9 values”. The warehouse loader checks the records against house schema v4, writes them with house units, qualifiers and flags, holds back no values, links each to its page or cell and refreshes the SAR view for the Cedar program. QB-1180 moves to Loaded — eleven minutes after it arrived.
The warehouse loader loads only when every exception on the deliverable is decided.
In the review, the workbook sits beside the record grid. Click NWB-5104, IV at 0.25 h, and a line runs to cell E5: as reported 1.284 µg/mL, in house units 1,284 ng/mL, confidence 97, “µg/mL converted to ng/mL (× 1,000) by the rule approved today”. The two 24 h samples read <LLOQ (1.0 ng/mL), not zero.
On File templates, Quillon's plasma concentration export is now v4: column E reads µg/mL → ng/mL. The drift log for the quarter gains its newest row — Oct 7, QB-1180, unit changed, converted × 1,000, template v4 — Dr. Priya Raman. Earlier rows show how harmless changes were handled: a sheet renamed at Halden matched by content, a column moved at Pellucid matched by name.
Renamed or moved columns are matched above 95% confidence; a unit change in a header always stops the file.
Halden's metabolic stability study HB-26-0417 — 96 values from human and mouse liver microsomes, day 8 of 10 — is read with a page or cell for every value and 46 converted to house units. On page 6 the verapamil control on MLM plate 2 ran at 17.9 min. Halden's SOP accepts 3–20 min and calls it Pass; contract NWB-SOW-031 §4.2 allows 3–15 min, and the report notes a second freeze–thaw of microsome lot ML-2291 on that plate.
The study-spec checker never marks a failed control as passed — the contract range wins over the CRO's SOP.
Each exception arrives with its evidence and a proposed decision. For the plate 2 control: hold the 22 MLM values and ask Halden to repeat the plate with a fresh microsome lot. For NWB-4488, CV 31% against the contracted 20%: load with a “high variability” flag. NWB-4490, not incubated because it precipitated at 1 µM: record as “not tested — solubility”. Corrib's “NB-4493” is not on the shipped manifest, so Sam decides whether it is NWB-4439, digits swapped on the tube label.
“The agents propose — a named person decides.”
The same deliverables, seen by the four people who carry them — what their week looked like, and what it looks like now.
Seven agents classify, compare layouts, extract, normalise and check every file against its contract; the study monitor decides each stop.
Reads every file in a CRO package — report PDF, results workbook, instrument export, cover email — and identifies the CRO, study, assay and contract schedule, matched to the open study in the study register.
Before anything is mapped, compares each file with the CRO's last accepted template: header names, column order, units written in headers, sheet names and each numeric column's range against the last three runs. Renamed or moved columns are matched and recorded; a unit change shows a before/after header diff.
Turns report tables, workbook cells and instrument exports into result records in the house schema — compound as reported, assay, matrix, condition, parameter, value, unit, qualifier, replicates — with the file and page or cell for every value.
Converts every value to the house unit by the approved unit rules and matches each compound as reported to the registry through the manifest shipped for the study, proposing the likeliest match when a name does not fit.
Tests every acceptance criterion in the contract schedule — positive and reference controls, replicate variability, recovery, integrity markers, time points, compounds delivered against the manifest and turnaround — and raises an exception with the evidence for each failure.
Drafts a short, courteous query to the CRO study director: what was found, the report page or workbook cell, the contract clause, which values are held and what is requested.
Loads decided deliverables to the research data warehouse, each record with its house unit, qualifier, flags and citation, keeps held values out until they are released and refreshes the SAR view.
Converts, holds, queries or loads. The agents propose; a named person decides.
Ask in plain words about a file, a value and its source, a contract's acceptance criteria or a CRO's performance — or tell it to add a rule. Answers point to the page or cell.
Why did the Quillon file stop?
QB-1180 stopped at Mapped because its layout no longer matches Quillon template v3. Column E now reads “Conc (µg/mL)” where the template and the contract say ng/mL. The values agree with the new unit: 1.284 at 0.25 h, against 1,190–1,410 ng/mL in the last three runs. Loaded as-is, all 9 values would be 1,000× too low. Route and Time also swapped places and two headers were renamed — those matched on their own.
What is blocking the Halden stability load?
Three decisions are open on HB-26-0417 (D-2611). The verapamil control on MLM plate 2 ran at 17.9 min — the contract allows 3–15 min; Halden marked it “Pass” against its own SOP and notes a second freeze–thaw of the microsome lot. 22 values wait on this. NWB-4488 replicates differ, CV 31%. NWB-4490 was not tested — solubility.
Which CRO fails quality checks most often?
Over the last 8 weeks, Halden Biosciences: 6.1% QC failures, 83% on time. Then Corrib Labs 4.2%, Quillon Bioanalytical 3.8%, Sorrel Bioservices 3.1%, Pellucid Safety 2.0% and Larkspur Screening 1.6%. Most of Halden's failures are controls — its SOP accepts a wider verapamil range in mouse microsomes than the contract.
Add a rule: remind the CRO when a deliverable is 2 days late
Added “Send the CRO a reminder when a deliverable is 2 working days past the contracted turnaround” to Settings. Today it would remind Corrib Labs about CL-7719 (day 12 of 10). Every reminder is on the audit trail.
13 screens from the working solution, on its sample data. Pick one to see it large.
Who decides, which rules the agents apply and where deliverables come from are settings, changed by your team and on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| ADME and safety exceptions (Halden, Pellucid) go to | Dr. Priya Raman | Dr. Priya Raman · Dr. Lena Ortiz · Sam Patel · Omar Haddad · Dana Okafor |
| Compound identity — any name not on the shipped manifest | Sam Patel | Dr. Priya Raman · Dr. Lena Ortiz · Sam Patel · Omar Haddad · Dana Okafor |
| Hold values when a control is outside the contracted range — even when the CRO marks it “Pass” against its own SOP | On | On · Off |
| Flag replicate variability above | CV 20% | Step up or down |
| Minimum recovery for permeability | 70% | Step up or down — mass balance below it raises an exception |
| Stop the file when a unit in a header changes | Always on | A person accepts the conversion and a new template version is saved |
| Match renamed or moved columns automatically | On | On · Off — above 95% confidence, recorded on the file and in the drift log |
| Compare value ranges with the last 3 runs | On | A column that jumps 10× or more stops the file even if the header did not change |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (TetraScience customer story on drug transporter assays — 2+ hours per scientist per week handling ADME CRO data; Benchling — up to 4 hours per scientist per week). The sponsor, CROs, compounds, studies and people named on this page are fictional — characters and sample data in the working solution.
Getting the results your contract research organisations send — PDF reports, Excel tables, instrument exports — into your own research data warehouse as clean, consistent records. CRO Data Integration does it with seven agents that classify each deliverable, check its layout, extract and convert every value, test it against the contract and load it with a citation, while the study monitor decides every exception.
Before a single value is mapped, the template drift checker compares each file with that CRO's last accepted template: header names, column order, units in headers and value ranges against the last three runs. Renamed or moved columns are matched and recorded; a unit change or an unmatched column stops the file with a before/after header diff and what would have loaded. A person accepts the conversion, and it is saved as a new template version.
Against your contract. The study-spec checker tests controls, replicates, recovery, integrity markers, time points, compounds delivered and turnaround against the statement of work, and the contract range wins when the CRO's own SOP is looser — as with a verapamil control the CRO marked “Pass” at 17.9 min against a contracted 3–15 min.
Qualified values such as > 60 min or < LLOQ are kept as reported and never turned into plain numbers. When an export writes 0 for a sample below the quantification limit, the study-spec checker raises it and the value loads as below LLOQ — zero is never written.
Yes. Every record cites its file and page and table, or sheet and cell, and any conversion names the unit rule that made it. In the review, clicking a value draws a line to the exact page or cell it came from.
Yes. The agents propose and a named person decides every exception and every drift; compound identity goes to the data steward. Values load only when every exception on the deliverable is decided, held values never reach the warehouse, and a query goes to the CRO only after the study monitor approves it.
GLP study data loads as a labelled convenience copy for research use, after a person confirms it. The signed report and the test facility's archived dataset stay the record, and every value cites the signed report page.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your statements of work, each CRO's file layouts, your unit rules and a sample of past deliverables — and it goes live once every quality gate has passed. We will walk you through it on your own CRO data first.
Sourced answers from papers, patents and reports in seconds, our own disagreements first.
1-min demoCompetitive Pipeline TrackingEvery rival trial, readout and approval in a sourced brief each Monday
1-min demoPreclinical Study Report AutomationA draft GLP report in hours, every number traced to raw data.
1-min demoWe'll run CRO Data Integration on a sample of your own past CRO deliverables and contracts.