R&D stage-gate governance, explained.

What a stage-gate review is: the meeting where a research program earns its next budget

One program's route from a guess to a first human dose

1. Target to hitIs this protein worth a screening campaign? Northwind Bio's NWB-6032 asks for $2.9 million over 8 months.
2. Hit to leadDoes one chemical series deserve an optimization team? NWB-3105 has only one series, and its committee is leaning toward hold.
3. Lead to candidateIs this exact molecule worth the IND-enabling money? NWB-4417 asks for $11.8 million. This is the decision the example follows.
4. Candidate to INDIs the package complete enough to ask regulators for a first human dose? NWB-8803 asks for $31.0 million, and this is where it stops.
5. The clinic, and more reviewsPhase 1, 2 and 3 each bring their own, larger decisions. The habits built here carry over.

Every review works the same way, and step 3 is the example's main thread. All programs and amounts here are fictional.

Why a score of 72 can mean two things, and three more portfolio surprises

Four terms that run through every portfolio review

Fourteen terms are defined in the one-page summary.

The four stage-gate reviews before a first human dose, and the question each one asks

In shortCompanies name these reviews differently, but the order rarely changes, because each one reads evidence that did not exist before the previous stage finished. Each ends in one of four outcomes, and each outcome needs its own design.

A stage is the work: lead optimization, say. The gate is the decision that opens the next stage. The pack is the evidence. Teams argue about the work and the pack, and spend less time designing the decision itself: the criteria, the thresholds, the band between pass and fail, and who may sign.

Read the four questions in order. The first review decides on arguments about biology and competition. The second decides on a family of molecules. The third is the first time one exact structure is in the room, with a number against every property. The fourth checks that the package meets what an IND must contain. In the chart, GLP studies are those run under Good Laboratory Practice, FDA's rule for certain nonclinical studies, and a GMP batch is one made under Good Manufacturing Practice. A company that runs all four on one template is scoring four different questions on one answer sheet.

Go, go with conditions, hold, stop: what each outcome has to specify

The right-hand column is our reading of common practice, drawn from the literature cited in sections 08 and 09. We found no standard that sets these terms.

Which rules touch a portfolio decision, which do not, and what moved in 2025 and 2026

In shortWe found no law that tells you how to run a stage-gate review; that is our reading of the texts below. Rules apply at its edges: what an IND must contain, how long study records stay, when an electronic signature rule applies, and how inside information is handled. Several of these edges moved recently.

The first thing to know is what FDA looks at. Its aim in reviewing an IND is, in all phases, "to assure the safety and rights of subjects", and the submission must carry chemistry, manufacturing and control information and pharmacology and toxicology information.2,5 FDA reads that package. We found no FDA rule about the committee that decided to assemble it. A criterion such as "first-in-human plan agreed" exists because the IND will ask for it, and the rule stops there.

This guide's examples use efficacy, exposure, manufacturing, market and cost. Real packages for FDA also carry nonclinical toxicology, as 21 CFR 312.23(a)(8) requires, and your own specialists judge those studies.

What goes in a lead-to-candidate evidence pack: ten documents from six functions

In shortA pack draws on six functions: biology and pharmacology, drug metabolism, chemistry, legal, competitive intelligence and finance. Most reports arrive as drafts first and change later. Knowing which version a value came from matters as much as the value.

Here is the pack behind Northwind Bio's NWB-4417 decision, by the function that owns each piece. Your documents will carry other names. The list tends to look familiar, because each item answers a question the committee will ask.

Draft, final, superseded: why the version sits beside the value

In the example, a draft solubility report said 38 µM on September 18, and the final said 42 µM on October 5. Finance's estimate went from $12.4 million to $11.8 million two days before that. Both changes came after a deck would normally be frozen. Neither moves the outcome here. On another day, a moved value could turn a met criterion into a partly met one, and the room would argue over a number that was already out of date. Every value should carry its document, version, page and date, and someone should see a list of what changed since the last build.

Evidence quality also varies in ways a version number cannot show. An analysis of past studies put the cumulative prevalence of irreproducible preclinical research above 50%, at about $28 billion a year in the United States alone.22 Those authors are estimating across studies, and the figure is debated. For a committee, the practical lesson is a question for each result the decision leans on: who has repeated it?

Who prepares, who votes and who signs on a portfolio review committee

In shortThree jobs get muddled: preparing the evidence, voting and signing. Separating them is the cheapest governance a company can buy. The person who wants the money presents the case and does not vote on it.

We found no regulation that sets this structure, so every company writes its own. A common pattern gives the chair the signature, gives heads of function a vote, gives finance a co-signature on larger budgets, and gives the portfolio office the calendar and the templates. In the example, the committee has five voters, and a quorum of two voting members before the chair may sign.

Four design choices for a committee that decides

1. Write down who signs at each review. Early reviews with small budgets often need only the chair's signature. Larger budgets may need a co-signature from finance, and a named deputy for absences.
2. Set a quorum and a pre-read rule. A vote from someone who has not opened the pack adds little. Track who has read what, by name, before the meeting.
3. Handle conflicts. A therapeutic-area head who owns the program and votes on it has a conflict everyone can see. Decide in advance whether that person abstains, or votes on the record with the conflict noted.
4. Protect dissent. Published accounts of one company's reform stress a culture of "truth seeking", backed by more quantitative decisions.23 A written dissent on the record is one small way to make that real.

Weighted scoring at a gate review: how NWB-4417 reaches 86 while one dimension sits at 60

In shortA weighted score multiplies each dimension's result by its weight and adds them up. It is useful as a way to find gaps. It is weak as a verdict: it hides a failing criterion behind strong ones, and two templates produce numbers that do not compare.

Here is the arithmetic for the example, on the sample company's lead-to-candidate template, version 4.2.

A reader of the number 86 sees a healthy program. A reader of the row sees that one dimension carrying a fifth of the weight scored 60, and that the 7.8 times margin behind it is a draft from an outside lab. Both views are true, and the second is the one a committee needs.

Same total, different weights: two programs that both score 72

Give two area teams their own weights, then look at two programs. These numbers are made up to show the effect.

Program A is strong biology carried by weights that favor biology. Program B is sound chemistry, with thin biology. Each area can defend its template, and a therapeutic area may have good reasons for its weights. The mistake is ranking the two totals against each other when the portfolio is tight. If leadership wants to compare across areas, it needs one shared set of dimensions, and it should read the pattern of shortfalls alongside the total.

Why the quality of the test behind a criterion outweighs the number of criteria

Scannell and Bosley modeled R&D as a series of decisions and found that an absolute change of 0.1 in the correlation between a model's output and clinical outcomes can offset a 10-fold, even 100-fold, change in brute-force efficiency.24 A later review with co-authors made the same case for predictive validity across discovery tools.25 For a committee, that means writing down, whenever a threshold is set, the evidence that it predicts anything.

Drug development attrition: the published figures, who counted them and what each one measures

In shortMost programs fail, and the exact percentage depends on who counted what and when. BIO's latest edition gives 7.9% from Phase I to approval for 2011 to 2020. Quote the source, the dates and the method, or leave the number out.

Those rates come from an industry association and two data firms, using a commercial database. The same series gave 9.6% in its 2006 to 2015 edition.30 A different method gives a different answer. Wong, Siah and Lo at MIT used 406,038 trial entries and more than 21,143 compounds, followed industry-sponsored programs along their whole paths, and reported that 13.8% eventually led to approval.31 All three are honest numbers answering slightly different questions.

The classic model: where research money sits before the clinic

A 2010 model from authors at Eli Lilly starts about 24 projects at target to hit for each launch, and puts the capitalized cost at about $1.8 billion over 13.5 years.35,36 We read the paper's abstract; the stage figures below come from later slides that reproduce its model.

This 2010 model comes from authors at one company, on assumptions they spell out, and it is 16 years old. Treat its dollars as a shape. Its point still holds: because of the cost of capital and the many programs that fail, the decisions before the clinic carry a large share of the cost per launch, and they are decided with thin evidence.

Does better governance move the numbers? The best public case

AstraZeneca reviewed its small-molecule projects from 2005 to 2010 and built decisions on five determinants: right target, tissue, safety, patient and commercial potential.21 In 2018 its authors reported success from candidate nomination to Phase III completion rising from 4% to 19%.23 They describe where the approach failed as well. It is one company's account of itself, the period includes other changes, and we found no outside replication. It remains the most detailed public example of fixed criteria at candidate nomination coinciding with better results.

Returns, pipeline size and approvals: why the next funding choice matters

Deloitte's analysis of the 20 largest companies, published May 7, 2026, puts the forecast return at 7% for 2025, and 2.9% without GLP-1 and GLP-1/GIP drugs.37 Scannell and colleagues showed that new drugs approved per billion R&D dollars fell by about half every nine years from 1950, around 80-fold in inflation-adjusted terms.38 Pipelines stay very large: Citeline counted 22,940 drugs in 2026, down 3.92% and the first fall since the mid-1990s, though it says data changes may affect that comparison.39 FDA's CDER approved 46 novel drugs in 2025.40 A tall pipeline and a short list of approvals are what make the choice of what to fund next matter.

No public benchmark exists for how long a pack takes to prepare, what share of conditions close on time, or how often a committee reverses itself. Section 18 shows how to measure your own.

What a decision record must keep so someone can still read the reason in five years

In shortMinutes say what happened in a meeting. A decision record says what was decided, on what evidence, under which template, by whom, and what has to stay true. Write it so a stranger can follow it after every person in the room has moved on.

The test comes later. A program fails in Phase II, and the board asks what the nomination rested on. Or a partner's diligence team wants the history of a molecule. Or the program that worked was nearly stopped, and someone wants to know what saved it. If the answer lives in a slide deck on a former employee's drive, the learning is gone. The AstraZeneca authors could only do their review because they could trace projects across years.21

Check your own last ten sets of minutes against the table. The two signature rows borrow Part 11's design, and section 03 sets out its scope.9 In our reading, most portfolio records sit outside that scope, so this is good practice, and your counsel decides any question of scope.

Corrections that leave the first version standing

A record will sometimes turn out wrong: a value was mistyped, a report was withdrawn. Overwriting it removes the history a later reader needs. The cleaner habit is a new record that points to the old one and says what changed and why. The same goes for criteria. When a template changes, packs already in committee keep the version their votes were cast on, and the template's owner approves the new one.

Conditions that close: how to write one, and how to chase it

1. Make it measurable. "Confirm a coverage margin of at least 10× before GLP studies start" can be checked. "Further evaluate exposure" cannot.
2. Give it one name and one date. The example's three: January 15, 2027, February 26, 2027, and the IND review in November 2027.
3. Link it to its evidence. Name the report and page that made it necessary.
4. Say what happens if it fails. Back to committee, automatic hold, or a stated exception. Pick one in advance.
5. Put overdue items in front of the chair. A late report that shows up only in last month's minutes will stay late.

What happens to a stopped program: the reason, the people, the records, the molecule and the money

In shortCompanies spend months deciding on a stop and an afternoon carrying it out. A stop needs its own close-out plan: the reason on the record, a team released with care, study records and samples kept, contracts wound down, and a real answer on what becomes of the molecule.

Stopping is hard for human reasons. "Quick kill" strategies aim to reduce late-stage attrition by bringing termination forward, and Peck and colleagues ask what barriers stand in the way.3 Lendrem and colleagues name a progression-seeking bias in R&D decisions.4 The program leader has years of work in it, the team fears for its jobs, and a plausible next experiment always exists. The defense is written in advance. A kill criterion is a threshold agreed before the data arrives: "if no series reaches 30× selectivity by month 14, this program stops". The test of a real one: could a stranger apply it?

Hold deserves the same care. A hold with no restart test becomes a program that is neither funded nor closed, drawing attention and a little money for a year. Write down the result that restarts it, who re-tests, and the date when a hold becomes a stop by default.

Where stopped molecules can go, with real programs that took them

A molecule that fails one purpose may still have value for another. In a pilot run by NIH's National Center for Advancing Translational Sciences, eight drug companies made 58 compounds available to academic researchers for new uses, and the first nine awards came in June 2013.43 In December 2014, seven firms said they would make 68 stalled compounds available for academic research.44 NCATS listed clinical trial information for each of its compounds, so in our reading a molecule stopped before any human study has fewer takers. Out-licensing or a spin-out is the commercial route, and each depends on a clean package and a clear reason for the stop. A partner reading your record wants to see that you stopped for portfolio reasons or for a specific, stated result.

Learning from a stop: which criterion would have caught it sooner

The stop is also data. One useful habit is a short review: which criterion would have caught this earlier? Would a stricter threshold have stopped three good programs as well? Over a few years a committee builds its own evidence on its own thresholds. Without that, each new template is an opinion. Run the review after the team has moved on, keep the discussion on the template and the evidence, and assign no blame to people.

Where portfolio reviews go wrong: eight failures and the control that catches each

In shortWe found no regulator that inspects these decisions, and so no public finding data. The failures below come from the published studies cited here and from practitioner accounts. Check your last ten decisions for each one.

Treat the placements as a starting point for a conversation. One thing the published work does support: process has a large effect. In McKinsey's 2010 study of 1,048 major decisions across industries, managers' reports of the decision process explained more of the outcome than the quality of their analysis, by a factor of six.28 That is a consultancy's study on self-reported practices, outside pharma, and we read quotations of it, not the full article. It fits what the AstraZeneca authors say about culture.23

AI at a portfolio review: what FDA, EMA and the EU AI Act say, and which choices stay with people

In shortCurrent regulator guidance largely leaves discovery-stage AI to the developer, while warning that results feeding a submission get more scrutiny. Software can read reports, pull values and draft text, and a named person must still own the choice to spend or to stop.

Three limits of software in a portfolio review

First, published claims about AI in drug discovery measure something else. Authors at Boston Consulting Group reported that AI-discovered molecules had 80% to 90% success in Phase I and about 40% in Phase II, the latter on a small sample and comparable to history.49 That concerns molecules, and it is a consultancy's analysis. No public study we found measures how accurately software reads an evidence pack, so test it on your own past packs.

Second, a confident summary can mislead a tired reader. Every value needs a link to the page it came from, and the checker needs to click. Third, weights and thresholds are the company's judgment. Software should apply them exactly as written, record the version, and never change them. If a recommendation appears on screen, label it advisory, and show the committee how it was built.

A portfolio cycle today: five roles, an estimated 19-day pack by hand and three ways to run it

In shortMost companies run this on slides, shared drives and a spreadsheet of decisions. A program leader assembles the pack, the portfolio office chases it, the committee reads between other meetings, and the secretary types the minutes. No public benchmark gives the hours, so the split below is an estimate you should replace.

The quotes are from the guide's fictional sample company.

Three ways companies run the cycle, each with its catch

From the portfolio map to a signed record: NWB-4417's Go with conditions

In shortThis is the example from the top of the page, inside the product. Dr. Omar Haddad, the chair, opens the map, takes the 86 apart down to a page, reads the members' votes, keeps the three drafted conditions, writes the rationale and signs. The agents recommended. He decided.

1

The whole portfolio on one funnel

Omar's first screen is the funnel: 12 programs at four reviews, $73.5 million asked at the October 14 meeting. Each bubble is sized by its cost to the next review and colored by its evidence score: green at 80 and above, amber from 65 to 79, red below 65, dashed while documents are still arriving. The pass rates sit on the lines between stages: 62%, 48% and 55%, sample figures.

Omar clicks NWB-4417, the Pathway Q inhibitor scoring 86 in stage three.

2

A score he can take apart

The rail shows how the 86 is built: seven dimensions, each with its weight, its score and the document behind it. Exposure carries 20% and scores 60. Below that sit three conditions, each tied to a study code. The first opens the rat study at page 11, with the values marked: 7.8 times coverage against a 10 times criterion, in a report still marked as a draft from the contract research organization (CRO), the outside lab that ran it.

This is the question the chair asked on the first page, answered in two clicks.

3

The meeting, its agenda and who has read what

Seven days out, 13 of 20 votes are in and 18 of 20 pre-reads are open. Dana Okafor, the portfolio governance lead, can see which two pre-reads are still unopened, and whose they are. The pack went to the five members on October 7 at 7:50 am, marked market-sensitive.

4

Fourteen criteria, each with a value and a source

Inside the dossier, every criterion shows what was measured against its threshold, the result, a confidence and a numbered source. Dr. Maya Chen, the program leader, checked it and released it at 8:05 am. The line above the tabs says what built it: six agents, ten documents, template v4.2, 3 minutes 12 seconds.

5

The votes, then the chair's own decision

Four of five votes are in: three for Go with conditions, one for Go. Dr. Ines Moreau, head of DMPK, ties her vote to the 28-day study: "Condition 1 is mine." Sam Patel, from finance, notes the $11.8 million is in the FY27 plan. Omar picks from the four outcomes and keeps all three drafted conditions. He could untick one or add his own.

6

A signature with a stated meaning

Omar chooses what the signature means, enters his password and signs. The dialog tells him the record seals at that moment, and that a later correction becomes a new record pointing to this one.

7

A sealed record, and a bubble that moves

PD-26-038 holds the rationale, the three conditions with owners and dates, the signer, the time and the meaning, sealed with a hash. Back on the map, NWB-4417 sits at candidate to IND, with its conditions now under "conditions signed".

How two changed values reach the program leader before the committee reads them

In shortWhen new evidence lands, the agents rebuild the dossier in view and flag what moved. Here, solubility goes from 38 to 42 µM and cost to IND from $12.4 million to $11.8 million. The score stays at 86, and the program leader checks the changes before release.

The same rail opens any partly met criterion. Solid form is a judgment against a written threshold, so the score shows "judgment · see source", and the passage holds the facts: 7 steps at 420 g scale, form B selected, and partial conversion to form A at 60 °C and 75% relative humidity after 4 weeks.

NWB-8803 stops: what fell short, the chair's reason and what the registry keeps

In shortNWB-8803, a fictional antibody-drug conjugate at candidate to IND, scores 58. Four criteria are met, three partly and three not. The agents recommend Stop, members agree, and the chair writes the reason in his own words. The record joins the other stopped programs, each with a reason someone can read later.

NWB-8803's four shortfalls against its criteria

• Coverage margin of 3× against a 10× criterion
• Conjugation yield of 61% against a 75% target
• 14 rival programs, 3 of them in Phase 3
• $31.0 million and 16 months, $6.4 million over budget

Lena Ortiz, head of immunology and deputy chair: "Fourteen rivals and a 3× coverage margin — we cannot win this one." Sam Patel: "$6.4 million over budget with a weak competitive case." Neither the score nor the recommendation decides it. Omar picks Stop and writes his reason.

The record holds the reason. Your close-out tasks, such as study records, contracts, disclosure review and the molecule's future (section 09), stay with the owners your company names. A Hold carries its restart tests as conditions: NWB-5098's record names a back-up series read-out and a predicted dose at or below 300 mg, each with an owner and January 30, 2027 as the date.

Six agents read and draft, members vote, and the chair signs each decision

In shortThree readers pull values with citations. A scorer applies your template exactly as written. A risk analyst and a writer draft the rest. The agents hold no vote and no signature, and their recommendation is advisory.

The people who check, vote and sign at Northwind Bio

Open conditions, shortfalls by area and criteria you own, for the whole research portfolio

In shortThe same record feeds the portfolio view: overdue conditions on the map, a dashboard of preparation time and shortfalls by therapeutic area, and criteria templates you control. Design targets are labeled as targets.

Evidence in, signed decision out, with the systems you already keep

Evidence comes from where you already keep it: document management, shared drives, electronic lab notebooks, finance planning tools and legal matter systems. Enterprise controls include roles and committee-only access, maker and checker steps, watermarked exports, a full audit trail, and signature with a stated meaning. These controls support your own compliance program. Botminds is SOC 2 Type II audited, with information security managed to ISO 27001.

What a late, weak or unrecorded research decision costs your portfolio

In shortPreparation hours are the small number. The large numbers are a wrong go that consumes a stage's budget, a late stop, and a delay to the programs that deserved the money. Few of these have public benchmarks, so a week of counting in your own records will tell you more than any industry average.

Six costs of your portfolio reviews, and how to count each one in a week

What no spreadsheet counts: the good program never funded, and the team that lost faith

Sizing your return from faster packs and one earlier stop a year

In shortThree levers: preparation time, re-checking, and stopping a weak program one review sooner. The first two are easy to measure and small. The third is hard to claim and large. Build the case on the first two, and report the third as upside.

Three places portfolio review software can pay back

Portfolio review calculator: enter your own figures

Formulas: time saved = reviews × (days today − days with software) × daily cost. Upside = programs stopped earlier × spend avoided. Every default is illustrative. Software cost is not included. A research program's spend avoided is real only if the money goes to something better.

A readiness score for your portfolio review: eight questions, from templates to the chair

In shortScore each question 0, 1 or 2. Low scores point to work on your process first, and software cannot substitute for that work. Nothing you choose leaves this page.

Reading your band: what to do next at your portfolio committee

What to ask any portfolio decision vendor, and a ten-week pilot on your own past packs

In shortAsk for proof on your documents. Ask how it handles a draft and a final that disagree. Ask who may change a threshold, and what a signature means. Then run a pilot on packs where you already know what happened.

A ten-week pilot on packs where you already know the outcome

Five pilot measures, from citation accuracy to committee use

Stage-gate questions people search for, each answered on its own

In shortEach answer stands on its own, with sources in the sections above.