Nonclinical written summaries drafted in hours, every number checked against its report page
IND nonclinical summaries drafted in hours, each number checked
See one case, screen by screen ↓They are Module 2.6 of the Common Technical Document under ICH M4S: factual written summaries of the pharmacology (2.6.2), pharmacokinetics (2.6.4) and toxicology (2.6.6) studies, with tabulated summaries (2.6.3, 2.6.5, 2.6.7). The 2.4 nonclinical overview interprets them. Nonclinical Written Summaries drafts all of them from the study reports, for the writer to review and sign.
How to choose AI solutions for life sciences →
Updated
Before a first-in-human IND goes in, a nonclinical regulatory writer turns 20–60 study reports into the Module 2.6 written and tabulated summaries for pharmacology, pharmacokinetics and toxicology, and then the 2.4 nonclinical overview. The reports run to thousands of pages, and the writer chases every study owner for the NOAEL, the toxicokinetics and the conclusion that each section needs.
The risk sits in the numbers. Tabulated summaries are retyped from report tables, and a transposed digit in a toxicokinetic value quietly changes the exposure margin that justifies the starting dose. Text and tables drift apart over review rounds, and every report amendment means re-checking each sentence that cites it — by hand.
Hours to a first draft, by section. The by-hand split is the working solution's own estimate: its 2.6.2, 2.6.4 and 2.6.6 rows add up to the published figure of about 100 hours, and 2.6.7 comes on top. The agent hours are from its sample IND. Read the split as an estimate.
Six agents read the reports, draft, build the tables and check every number; the writer and section owners sign.
NWB-4417 is an oral kinase inhibitor from Northwind Bio heading into its first-in-human IND. Module 2 sign-off is due on 30 October and the IND goes to FDA on 20 November. Here is how the toxicology summaries get to signature, screen by screen, in the working solution.
Lena's home screen is the Module 2 nonclinical binder: a tab for 2.4 and each of 2.6.1–2.6.7, every section with its approver, due date and a five-step sign-off track. 27 headings need her, 4 numbers are to resolve and there are 2 coverage gaps. She selects 2.6.6.3 Repeat-dose toxicity: it draws on 5 reports and cites 46 of 48 extracted findings.
Coverage gap: “TX-311 · Urinalysis, Day 29 — no test-item effect” — with “Add a sentence” or “Not needed”.
Seven pharmacology, five pharmacokinetics and eleven toxicology reports, each with its test system, GLP status, pages, findings and the sections it feeds. The 28-day rat study TX-311 alone runs to 2,914 pages and gives 286 findings. The reader records each finding with study, species, sex, dose, day, value, unit, page and table.
The reader never infers a NOAEL the study director did not state.
Lena opens 2.6.7.7 Repeat-dose toxicity: pivotal studies. The table builder fills the house template from TX-311, TX-312, BA-511, BA-512 and HP-311, and the right-hand panel shows the report page it is reading — here BA-512 Table 3, page 11, the dog toxicokinetics on Day 28.
“Every cell comes from a finding the reader took from a report page. No value is typed by hand.”
18 rows and 55 cells from 5 reports, every cell linked to its page, 55 of 55 matching their source. Comparing the finished table with the written summary, the number checker finds a conflict in 2.6.6.3 and flags it for Lena.
“The text disagrees with this table. 2.6.6.3 says the dog AUC at the NOAEL is 22,400 ng·h/mL; this table took 24,200 from BA-512 page 11.”
Lena clicks the dog AUC at 15 mg/kg/day. The source panel opens BA-512 at Table 3, page 11, with 24,200 ng·h/mL highlighted. The extracted finding reads AUC0–24 at NOAEL, dog, Day 28, sexes combined, with a reader confidence of 99.
One click opens the 2.6.6 Toxicology Written Summary at the sentence that carries the dog NOAEL. A line runs from the flagged 22,400 to the value on the BA-512 page. 425 of 428 numbers in the section match.
“Disagrees with the source. Digits transposed. BA-512 Table 3 gives 24,200 ng·h/mL (sexes combined).”
The section's number check lists every number in the draft beside the value in the source: 364 match exactly, 61 are derived and recomputed — margins, human-equivalent dose, fold changes — and 3 disagree. Besides the dog exposure and the margin built on it, the female ALT change at 100 mg/kg/day reads 1.8-fold, the male value; TX-311 Table 11 gives 1.5.
The number checker cannot edit the draft. It proposes the source value, and sign-off stays blocked while any number disagrees.
Lena takes the source value. The margin in 2.6.6.9 was built on the mistyped AUC: 22,400 ÷ 610 gave 37-fold. The number checker recomputes it from BA-512 and the projected human AUC of 610 ng·h/mL in PK-230 — 24,200 ÷ 610, so 40-fold — and the change goes into the section history.
“Taken from the source. Exposure margin to the dog NOAEL: 40. Recorded in the history — recomputed by the number checker.”
With every number matching, Lena signs 2.6.6 with the meaning of her signature: “I drafted this section with the agents and checked every number against its source.” The AI use is recorded with the section — drafted by the summary writer, 428 numbers checked, 3 corrected from the sources — and the section goes to Dr. Omar Haddad for approval.
The summary writer drafts the 2.4 Nonclinical Overview from the 2.6 sections and the starting-dose rationale in PK-230: 142 statements, each cited. The number checker recomputes 79-fold to the rat NOAEL, 40-fold to the dog NOAEL and an MRSD of 29 mg; all 286 numbers match. It is ready for Lena's review, then for Dr. Maya Chen's approval.
PK-230, page 14: “Human-equivalent dose of the rat NOAEL: 30 ÷ 6.2 = 4.8 mg/kg; × 60 kg = 290 mg; with a 10-fold safety factor the maximum recommended starting dose is 29 mg.”
The same binder, seen by the writer and the four scientists who approve it — what the weeks before an IND looked like, and what they look like now.
Six agents read the reports, draft, build the tables and check every number; the writer and section owners sign.
Reads each nonclinical study report, appendices and tables included — design, GLP statement, doses, findings, incidences, toxicokinetics — and records every finding with study, species, sex, dose, day, value, unit, page and table.
Maps the findings to the ICH M4S headings of the house template, lists the studies cited in each section and flags any study that feeds no section.
Drafts 2.6.2, 2.6.4, 2.6.6 and the 2.4 overview in the house style, one cited sentence at a time, in the units of the template.
Builds the 2.6.3, 2.6.5 and 2.6.7 tabulated summaries from the same findings the text uses, each cell linked to its finding, report and page.
Compares every number in every section with its cited source and recomputes derived values — exposure margins, human-equivalent dose and the starting-dose maths. Reports each as match, rounding, derived or disagree, with the reason.
Compares the sections with each other, with the Investigator's Brochure and with the protocol, and reports conflicts with both citations.
Resolve, sign and approve. The agents propose; a named person decides.
Ask in plain words, or say what to change. Every answer cites the report page it came from.
What are the NOAELs, and what margin do they give the starting dose?
Two pivotal 28-day GLP studies set the NOAELs: rat 30 mg/kg/day, AUC0–24 48,200 ng·h/mL, a 79-fold margin at 10 mg; dog 15 mg/kg/day, AUC0–24 24,200 ng·h/mL, a 40-fold margin. Both margins use the projected human AUC of 610 ng·h/mL at the 10 mg starting dose. Note: 2.6.6 still says 22,400 and 37-fold for the dog — the number checker flagged both.
Which numbers still disagree with the study reports?
4 numbers disagree: 2.6.6 ALT fold change in females at 100 mg/kg/day — draft 1.8, source 1.5; 2.6.6 AUC0–24 at the dog NOAEL — draft 22,400, source 24,200; 2.6.6 exposure margin to the dog NOAEL — draft 37, recomputed 40; 2.6.4 human plasma protein binding — draft 97.8, source 97.2.
How was the starting dose chosen?
From the most sensitive species. Rat NOAEL 30 mg/kg ÷ 6.2 gives an HED of 4.8 mg/kg; × 60 kg = 290 mg; with a 10-fold safety factor the MRSD is 29 mg. The proposed 10 mg is about a third of that, with exposure margins of 79-fold (rat) and 40-fold (dog).
Add the in vivo micronucleus result to the 2.6.6 brief summary
Added to 2.6.6.1 Brief summary: “NWB-4417 was not genotoxic in a standard battery, including an in vivo micronucleus assay integrated into the 28-day rat study.” It cites the Ames test and TX-311 Appendix 12. The number checker confirmed the dose (100 mg/kg/day). The edit is in the section history.
13 screens from the working solution, on its sample data. Pick one to see it large.
The house template, the number check and who approves each section are settings — chosen once for the programme.
| Setting | Default | Choose from |
|---|---|---|
| House template | Northwind nonclinical v6.1 | Northwind nonclinical v6.1 · Northwind nonclinical v6.0 · Partner template (ICH M4S plain) |
| Citation style | Report · page · table (inline) | Report · page · table (inline) · Report number only · Footnotes |
| Units for exposure | ng·h/mL | ng·h/mL · µg·h/mL |
| Every number compared with its source | On — sign-off is blocked while any number disagrees | Always on — sign-off depends on it |
| Recompute derived values | On — margins, human-equivalent dose and starting-dose maths | On · Off |
| Flag rounding | Off | On · Off (a note when 12.1% is written as 12%) |
| 2.6.6 Toxicology Written Summary approver | Dr. Omar Haddad | Dr. Omar Haddad · Priya Raman · Sam Patel · Dr. Maya Chen |
| Record AI use on every section | On — required by the GenAI policy | Always on — required by the Northwind GenAI policy |
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 IND · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Eser et al., Human-AI Collaboration Increases Efficiency in Regulatory Writing, arXiv 2509.09738 (2025) — about 100 hours to first-draft 2.6.2, 2.6.4 and 2.6.6 by hand, a writer estimate, against 2.6–3.7 hours measured with AI assistance; section headings per ICH M4S; starting-dose method per the FDA 2005 guidance on the maximum recommended starting dose). People, companies, compounds and studies named on this page are fictional — characters and sample data in the working solution.
They are Module 2.6 of the Common Technical Document under ICH M4S: factual written summaries of the pharmacology (2.6.2), pharmacokinetics (2.6.4) and toxicology (2.6.6) studies, with tabulated summaries (2.6.3, 2.6.5, 2.6.7). The 2.4 nonclinical overview interprets them. Nonclinical Written Summaries drafts all of them from the study reports, for the writer to review and sign.
The study report reader records every finding with its page and table, the section planner maps the findings to the ICH M4S headings of your house template, and the summary writer drafts each section one cited sentence at a time. The table builder fills the tabulated summaries from the same findings, so text and tables share one source.
The number checker compares every number in every section with its cited source and recomputes derived values — exposure margins, human-equivalent dose and the starting-dose maths. On the sample IND it found 4 of 2,318 numbers that disagreed, including a transposed dog AUC and the margin built on it. Sign-off stays blocked until each one matches its source.
No. Every sentence must cite a finding, the writer does not make adverse or non-adverse calls the report does not make, and it leaves emphasis and judgement to the author. The table builder never types a value with no finding behind it.
Yes. Once the 2.6 sections are in, the summary writer drafts 2.4 from them and the starting-dose rationale, the number checker recomputes every margin, and the consistency checker compares the overview with each section and the Investigator's Brochure. The head of nonclinical safety approves it.
Yes. The agents draft, build tables and check; they never sign or approve. The writer resolves each flag and signs as author, each section's approver signs by e-signature with its meaning, and the AI use is recorded with every section.
Yes. The house template, citation style, exposure units, number-check rules, the approver for each section and whether 2.4 needs a second reviewer are all settings.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your house template, prior IND summaries and a set of final study reports — and it goes live once every quality gate has passed. We will walk you through it on your own reports 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 Nonclinical Written Summaries on a set of your own final study reports and house template.