How a Top-Twenty Pharma Team Cut Stage-Gate Review From Weeks to Days
This case study describes a real workflow running inside a large pharma company. Per brand policy, the company is not named.
A drug discovery team at a top-twenty pharma runs a molecular screening workflow four thousand times a day.
The workflow is a reviewed promotion decision. A computational chemist generates candidate molecules. Each candidate is scored by the team's in-house models. The highest-scoring candidates go to a stage-gate review, where a senior chemist decides whether the candidate promotes to the next stage of the pipeline. The decision has to be defensible to a regulator. The decision has to be traceable to the reviewer who made it. The decision has to be reproducible six months later when somebody asks why a candidate was rejected.
Before the Domain Lab engagement, the stage-gate review took five weeks per candidate batch.
After six months on the Drug Discovery Domain Lab, the same review takes three days.
The workflow did not change. The record under the workflow did. This is the story of what changed.
Where the five weeks went
Walk the old workflow slowly.
Week one. Computational chemistry batch-generates candidates in an internal notebook environment. Scores print in a spreadsheet. The spreadsheet is forwarded to the stage-gate review chair.
Week two. The chair assigns reviewers. Reviewers pull the candidate files from a shared drive. Each reviewer reads the generation notebook, the scoring output, the supporting literature references, and the synthesis feasibility notes. Each reviewer writes a disposition in a separate document. The documents aggregate in a folder.
Week three. The chair reads the dispositions. Disagreements between reviewers surface. The chair schedules an adjudication meeting. The meeting happens two weeks out on most calendars.
Week four. Adjudication meeting. Decisions made. Meeting minutes written. Minutes attached to the candidate files in the shared drive.
Week five. Regulatory affairs reviews the minutes. The decision chain is reconstructed from the dispositions, the adjudication notes, and the reviewer identities. A stage-gate promotion packet is assembled. Packet is signed. Candidate moves forward.
Five weeks. Five handoffs. Five places where context falls out. Five systems where the record lives — a notebook environment, a spreadsheet, a shared drive, a meeting minutes tool, and a regulatory affairs archive.
The team knew the workflow was slow. The team also knew that every one of the five steps was there for a real compliance reason. You cannot skip any of them.
What the team could not do, on their own, was run all five on one record.
What the Domain Lab changed
One screen. One record. Every review step, every adjudication, every sign-off, tied to the candidate data and the reviewer identity and the ChemBERTa-class model output that started the chain.
Generation. The computational chemistry team's notebook still runs. Its output now writes directly to the Domain Lab record instead of to a spreadsheet. The candidate metadata, the scoring output, the lineage to the training data the model used — all structured, all tied to the candidate.
Review. The chair and the reviewers open the same queue. Each candidate has its full generation record attached. Each reviewer's disposition is captured with a timestamp, a credential, and a rationale. Disagreements are surfaced by the system, not by a calendar.
Adjudication. Reviewers with conflicting dispositions are routed to an adjudication discussion that happens inside the record. The discussion is captured. The decision is attached to the candidate.
Sign-off. Regulatory affairs pulls the full chain with a single query. The stage-gate promotion packet is assembled automatically — generation record, review dispositions, adjudication discussion, sign-off signature. The packet is in the ALCOA+ format the team's auditor expects.
Same workflow. Same compliance rigor. Same reviewer roster. The difference is that no one is re-keying data between systems, no one is reconstructing the decision chain from fragments, and no one is waiting two weeks for a meeting that could have been resolved inline.
What the numbers moved
Track three metrics over the six months of the engagement.
Cycle time. Stage-gate review dropped from a median of thirty-six days to a median of three days. Not because any one step got faster. Because the handoffs between steps disappeared.
Reviewer agreement. Per-reviewer agreement on calibration cases, which the team had never tracked before the engagement, is now published to the chair every week. When a reviewer drifts, the chair sees it and re-calibrates inside a session. Before the engagement, drift was caught in quarterly reviews. Sometimes.
Audit readiness. Regulatory affairs used to spend forty to sixty hours reconstructing the decision chain for every candidate that reached an audit request. The reconstruction is now a ten-minute export. The team calculated that the freed regulatory-affairs capacity is equivalent to a full-time hire they now do not have to make.
What the team kept
The engagement ended in month six. The team retained five things.
The tuned ChemBERTa checkpoint. Fine-tuned on the team's six months of reviewed work. Performance on the team's internal validation set materially better than the starter ChemBERTa. Portable. Customer-hosted.
The regression set. Every candidate that was rejected, with the reason for the rejection and the reviewer who rejected it. A replay suite that blocks a future model from promoting a candidate that looks like one the team already rejected.
The decision history. Every reviewed candidate from the six-month period. The full chain. Immutable. Queryable. Auditor-ready.
The training signal. Labels, adjudications, reviewer notes, in a format the team's training pipeline reads natively. Ready for the next fine-tune run the team wants to do.
The measurable curve. A time-series of model quality against the team's own internal validation set. Not against a public benchmark. Against the work.
These five portable artifacts are what the "you keep the weights" promise means concretely for this team.
What did not change
It is worth writing this down, because it matters for a buyer reading this.
The team's regulatory posture did not change. The compliance rigor of the stage-gate review did not relax. The reviewer credential requirements did not change. The ALCOA+ format requirement did not change. The team's IP posture on generated candidates did not change.
What changed was the substrate under the work. The work is the same. The record under the work is one record now, instead of five.
What this says about Domain Labs
The pattern holds across every vertical.
A team already runs a hard workflow. The workflow is there for a real reason. The workflow is also the slowest, most handoff-heavy, most audit-expensive part of the team's operation. A Domain Lab does not replace the workflow. It carries the workflow on a record that every step of the team already writes to, so that the handoffs stop costing the team four weeks per cycle.
The reviewed work becomes training signal. The starter OSS model (ChemBERTa in drug discovery, MONAI + MedSAM in medical imaging, CatBoost in financial, YOLO + anomalib in manufacturing, Nucleotide Transformer in genomics) gets fine-tuned on that signal. The tuned model is the customer's.
Run the workflow. Own the model.
That is what a Domain Lab is.
The team in this case study ran their workflow. They own their model. The next audit request is a ten-minute export.
That is the promise.
---
Ready to run your workflow on one record?