BD search and evaluation, explained.

BD search and evaluation in plain words: finding a drug at another company, then testing the seller's story

From a registry entry to a signed license: where evaluation sits

1. SearchA standing sweep of registries, patents, abstracts and news, measured against written criteria. Most assets stay here for good.
2. First reading of the pitchNon-confidential material is read and answered. This front door has its own guide on partnering opportunity triage.
3. Evaluation in the data roomThe confidentiality agreement is signed, named people get access, six functions read the files against a bid date and write cited memos.
4. Committee and offerA scored recommendation with conditions and a price reaches the committee, which decides. A non-binding offer becomes a term sheet, a short summary of the main deal terms.
5. Contract and handoverLawyers turn conditions into covenants, promises written into the contract, and the findings pass to the alliance team that will manage the partnership after signing.

Four facts about buying drugs that surprise newcomers

Four words every licensing memo assumes: asset, thesis zone, data cut-off, wall-crossed

Fourteen terms are defined in the one-page summary.

From thousands of drugs watched to a handful in diligence: the funnel, the map and the thesis zone

In shortSearch is a long narrowing with few public numbers. Teams usually draw it as a map of fit against differentiation, with a shaded zone where they want to buy.

Large drug companies publish no funnel of their own. The clearest public one comes from a financier that buys drug royalties, and it is a fair analog because it reads the same kind of confidential data. In 2023 Royalty Pharma reviewed about 400 opportunities, signed confidentiality agreements on about 120, ran about 90 in-depth reviews, and closed nine deals.5

The source of those drugs matters too. A 2021 Boston Consulting Group paper said close to 60% of new therapeutic drugs in the previous five years were externally sourced, and that figure is now five years old.1 GlobalData estimated in 2025 that the US market is set to lose over $230 billion to patent expiry between 2025 and 2030, which is much of the reason buyers keep looking.6 Teams hold their favorites to two questions, usually drawn as two axes.

The weakness of the map is its age. A map built on public data and shown on a slide stays frozen while the memos pile up in folders. An asset that looks best in March can go on looking best until the manufacturing lead speaks at committee, because nothing connects the two documents.

What a public registry search can and cannot tell you about an asset

A registry search is the cheapest evidence, and the only kind you can show a colleague who is not on the deal. It tells you a study exists, its phase, size, endpoints and status. US sponsors of applicable trials owe results within a year of primary completion, with exceptions.7

In practice 40.9% of due trials met that deadline in one Lancet study.3 Failed programs leave less: of 74 late-stage trials of drugs that stopped for commercial reasons, only six were published.8 So a clean registry page about a rival means "we do not know" and should be written that way.

The law around a licensing diligence in November 2026: insider trading, antitrust, national security and loose words

In shortNo regulator inspects a licensing diligence, so no rule sets its method. The law bites at the edges: who may trade, what rivals may share, what a term sheet commits you to, and where the drug came from.

A data room's trading list, from the first phone call

Careful teams run the wall-cross in a set order. Someone asks whether the person is willing to be restricted, explains what that means, gets agreement, and only then gives the code name and the seller's name. Each person's name goes on a list with a date. The list should reach consultants, translators and outside counsel, because the duty follows the agreement they signed. Software that reads the files cannot trade, but the people who see its output can, so it belongs behind the same list.

What is in a seller's data room: numbered folders, several versions, and three kinds of file that mislead

In shortRooms look alike because the readers do. Version and data cut-off matter more than the folder a file sits in, because most contradictions come from date differences.

A virtual data room lets a seller show the same files to several bidders without losing control of them. It keeps one copy of each file, limits who can see which folder, prints the viewer's name across pages, and logs every view. One vendor's guide says 5,000 to 15,000 documents for a pharma licensing deal, with no source given.22 Our fictional case holds 214 files, smaller and closer to an early-stage asset. Count your own past rooms instead.

Questions to the seller are a record, and a signal

The room's question channel is the only written exchange most bidders get. An answer given there, dated and by name, can be cited in a memo; an answer on a phone call cannot. Every question also tells the seller what worries you, so careful teams review questions before they go, batch them by function, and keep them free of the buyer's own views on price.

The six opinions a licensing team owes its committee, and the document that settles each one

In shortA useful checklist is written backward from the sentence each expert must be able to sign. Each memo moves particular parts of the valuation, which is why they need to meet.

That grid shows why siloed reading is costly. A manufacturing finding can move the launch date and the cost of the next study, while the commercial memo prices that same launch year on a forecast written last month. If the two are not joined, the committee sees two versions of the same deal.

Three kinds of finding: deal-breaker, price item and covenant

A worry becomes a finding when it can be written as a sentence with a document, a consequence and an owner. Deal teams sort findings by what can be done about them. These words are common usage with no standard definition, so agree yours early.

Reading a small study and an investor deck: denominators, confirmed responses and the range around 61%

In shortEarly clinical data are real and thin. Before you price them, check the denominator (the patients counted), whether responses were confirmed, and the cut-off, and work out the range around the headline.

Take the fictional patient group (cohort) from the example. Of 31 patients on 80 mg once daily, 19 reached the study's primary response at week 12, which is 61.3%. On placebo, 1 of 10 did. A committee hears "61%". Here is what a careful reader adds, using ordinary arithmetic on those four numbers.

So the drug very probably does something, and whether it matches a rival near 70% is still open. That is the right note for the clinical memo: an honest signal from a small group, in the range of the class, with the size of the effect still to be pinned down.

Why two honest documents can give 45% and 38.1%

In cancer trials a tumor response usually has to be confirmed on a later scan. FDA describes objective response rate as the proportion of patients with tumor shrinkage "of a predefined amount and for a minimum time period".28 Investor decks often count everyone who shrank at any scan, while the study report counts those whose shrinkage held. In a second fictional case the June deck says 45% (18 of 40, confirmed and unconfirmed), and the interim report says 38.1% (16 of 42, confirmed).

Nobody needs to have lied for that gap to exist, and a buyer should still model the lower figure and ask why the two denominators differ. Keep both in the file, side by side, with the sources. A team that quietly picks one has made a judgment the committee never saw.

Five checks on any headline number in a room

1. Which denominator? Randomized, dosed, evaluable at the visit, or completers. The same responders give different rates.
2. Which cut-off? Compare only figures with the same one, or say why you cannot.
3. Which definition? Confirmed or not, and written before the study or chosen after.
4. Which dose and population? Pooled doses or the dose that would go forward; the study's patients or a subset.
5. What is missing? Ask for the clinical data tables behind any summary, and record what the seller holds back.

The clinical data also shape the plan after signing. For cancer drugs, FDA's August 2024 guidance on dosage says a recommended way to compare doses is a randomized, parallel dose-response trial, which a seller's plan may not include.29 In the second fictional case the regulatory lead prices that step at about nine months, and the launch date in the model moves with it.

When the lots in the clinic are not the lots in the animals: process changes, comparability and a new impurity

In shortA change of route or site after the animal studies can split one program into two materials. The buyer needs evidence linking them, or a study in the price, and the rule that applies depends on what kind of drug it is.

Drug makers change how they make things all the time, usually for good reasons: higher yield, bigger batches, a site that can reach commercial scale. The trouble starts when the change comes after the animal studies that justified human dosing.

A lot is one batch, and a route is the set of chemical steps that makes it. In the example, toxicology used two pilot-plant lots, and everything from lot HZ-DS-2104 on came from a different route at a contract manufacturer. The process report raises yield from 18% to 33% and says a comparison is planned for the first quarter of 2027.

Which comparability text applies to which kind of drug

People often reach for ICH Q5E, the international guideline on comparing a product before and after a manufacturing change. Its scope is proteins and polypeptides made in cell-culture systems.30 A chemically made peptide falls outside it, and outside ICH Q3A on impurities as well, because Q3A excludes peptides from its scope.31

EMA's guideline on synthetic peptides, in effect since June 1, 2026, was written for that gap. It asks for comparability results when two processes run in parallel, such as solid phase and hybrid, and says the tendency to epimerize, where one amino acid flips to its mirror-image form, should be investigated.32

ICH Q11 covers how drug substances are developed and made, and EMA calls its starting-material rules relevant to synthetic peptides, while ICH M3(R2) sets out the animal and lab (nonclinical) studies behind human trials.25,32,24 A memo should cite the text that covers the asset.

So the 0.31% in the example is below one threshold for identification and above another. Its importance comes from the comparison: it is in the new lots and absent from the lots that were tested. To qualify an impurity is to gather data showing it is safe at that level.31 A bridging toxicology study, an animal study on the new material, of about six months is the fictional team's proposed way through, and it is a cost a buyer can deduct.

That count covers applications at filing, which is later than most licensing deals. It still explains why experienced buyers read manufacturing early: a problem found in a data room costs a study, and the same problem found at filing can cost a review cycle. The publication made the count itself from the letters FDA released, and a letter can cite several reasons.34

Change-of-control clauses and royalty stacks: the contract pages that can stop a licensing deal

In shortA drug is often built on someone else's technology. The buyer inherits those licenses, their consent rights and their royalties, so the contracts folder deserves the same care as the science.

Take the second fictional asset, a two-target cancer antibody linked to a toxic payload, an antibody-drug conjugate. Its linker, the chemistry joining the two, is licensed in from another company. Section 14 of that license says a change of control, a sale of the company, needs the licensor's consent, or the license can be terminated. The patents can be sound and the clinical data strong, and a third party could still end the program after the deal closes.

Wording matters here more than instinct. In Meso Scale v. Roche, Delaware's Court of Chancery held that a reverse triangular merger, where the target survives as a subsidiary, "was not an assignment by operation of law or otherwise".27 A plain ban on assignment may therefore not catch that structure, while a clause that names change of control does. Which law governs matters too. A patent lawyer reads the clause; the first reader's job is to find every one.

A royalty stack is the sum of what you owe everyone else: university licenses, platform licenses, the seller. A stack of 13% on a drug with thin margins changes the deal even if each single royalty looks modest. Teams set a limit, 10% in our example, and treat assets above it as flagged. Read the anti-stacking clause too: it may let you offset some payments against others, or it may not.

How long the protection lasts: patent dates and the Medicare clock

The date on the composition-of-matter patent, the one covering the molecule, is the starting point. A granted US patent and a pending application are very different things to pay for. One fictional asset carries a non-final rejection on its US application, meaning the patent office has refused it so far, which leads the team to propose an option rather than a license until the patent is granted.

For a small-molecule drug the US Medicare negotiation law matters as well: selection can begin at 7 years after approval, against 11 for a biologic.35 A chemically made oral peptide approved as a new drug falls on the 7-year side, and the commercial memo should model that.

China-origin assets and foreign data: what FDA's rules accept, and what a buyer must still pay to bridge

In shortNo rule singles out any country. Foreign data can support a US application under stated conditions, and data from one region alone face a stricter test, so the cost of US patients belongs in the bid.

Chinese biopharma took 65% of the upfront money in those larger deals in 2026 to date, about $5.7 billion, up from 36% in 2025.2 Arnold & Porter, citing Reuters, put China cross-border licensing at a record $137 billion in total deal value in 2025, on a separate measure.36

For the search team, that shift means registries, patents and abstracts in Chinese, Korean and Japanese, with the same drug under several codes. Three tests help: same sponsor (the company running the trial), same mechanism and overlapping dates usually mean one asset; a shared protocol number settles it; an English abstract giving both codes is an alias worth keeping.

For the evaluation team the sharper question is what FDA will do with trials run abroad. Two rules and one public vote frame the answer.

A buyer reading a similar asset should ask three things and cost the answers: who were the patients, how did the comparison treatment differ from US care, and what bridging trial, linking those results to US patients, would FDA want? Add US sites to the plan, and the launch year moves.

Turning findings into a price: weighted scores, probability of success and a worked risk-adjusted NPV

In shortA score orders the conversation, and a risk-adjusted valuation prices it. Each finding should move one named input, so a committee can argue with one line instead of the whole number.

Base rates are where a probability starts. BIO's 2011–2020 study gives 7.9% from Phase 1 to approval across all diseases, 10.7% for autoimmune disease and 5.3% for oncology.26 A different method, the one Wong, Siah and Lo used on a larger set of trial records, gives 13.8% overall and 3.4% for oncology.41 Pick one source, say which, and use it across the shortlist, or your ranking partly ranks citations.

PTRS, probability of technical and regulatory success, is the base rate adjusted for what is known about this asset. Multiplying BIO's autoimmune rates from Phase 2 gives 19.3%.26

A weighted score, worked through

In the fictional case each function scores out of 5. Weights agreed in advance turn six opinions into one number. Clinical 3.5 at 30%, toxicology 4.0 at 20%, manufacturing 2.5 at 15%, patents 4.0 at 15%, regulatory 3.0 at 10% and commercial 3.5 at 10% come to 3.475. Rounded, that prints as 3.5, right on the line between "proceed" and "proceed with conditions". Keep the third decimal in the file: the weakest function, manufacturing, is what keeps it under.

Risk-adjusted NPV in four lines, using the example's numbers

Net present value (NPV) discounts future money to today. Risk-adjusted NPV (rNPV) also multiplies each future cash flow by the chance the drug survives to produce it. A common anchor for the discount rate is the 10.5% real (after-inflation) cost of capital used in published development-cost work.42 Put technical risk in the probability and market risk in the rate, and keep each risk in one place, because charging it twice is how good assets get rejected.

Our fictional asset's model, owned by BD finance, says $0.74 billion on public data. Here is a rough check on what the two findings did to it.

That simple walk lands near the model's figure. The habit is worth keeping: if a single line moves value by a quarter, you should be able to explain why with a pencil. A manufacturing finding already priced as a bridging study should reduce the probability only if the team believes the study could fail, and the memo should say which.

Published work also reminds the buyer what failure costs across a portfolio: a median of $985.3 million capitalized (counting the cost of money over time) to bring one new drug to market, in a study limited to 63 of 355 products with public data.42

Those medians are a check on a bid, and a thin one at later stages. Licensing value also moves in lumps: the annual totals below show 2025's record arriving after five flat years.

How licensing diligence goes wrong: the draft quoted, the slide never updated, the access nobody got

In shortNo inspector counts these errors, so no public error rate exists. The evidence comes from court records and from reading how rooms behave. The costliest failures are the ones found after signing.

Akorn: what a buyer learned after signing

Fresenius agreed to buy Akorn in 2017 and walked away in 2018. A Delaware court upheld the termination, in a case about regulatory and data problems at the target rather than a license. The judge's "rough estimate" of the valuation hit from the regulatory issues was about $900 million, 21% of the $4.3 billion implied equity value, wider than the cost of fixing them; the Delaware Supreme Court affirmed.43,44

Access to audit trails, the computer logs that show who changed a record, came after signing, during a whistleblower investigation. The lesson for a licensing reader: list what you were unable to see, and treat each gap as a question with a price.

AI in asset scouting and data-room reading: what it catches, which rules reach it and where a person takes over

In shortReading at volume suits AI when every statement keeps its page. Judgment on price and risk stays with named people. As of November 2026 the best-known AI rules were written for other work, and a careful buyer should say so.

Split the work by how it fails. Scouting fails by omission, the asset you never saw. Room reading fails by volume, the sentence nobody reached. Memos fail by inconsistency, two functions quoting different numbers. Those are coverage problems, and coverage is what software does well. Deciding that an untested impurity is worth a six-month study fails differently: it is a judgment about risk, and the person accountable for it should make it.

Even the best tool in that test missed about a quarter of the assets it should have found (recall 73.0%), on a benchmark that favors non-English regional assets, the hard case for general tools.45

A pilot in systematic reviews, where researchers pull data from published studies, gives the closest evidence for what a second reader does: single data extraction made 21.7% more errors than double extraction, and took 36.1% less time.46 That was one pilot on a different task. Its shape still fits an automated first pass that hands an expert the differences to judge.

Three limits that need a person, whatever the tool

Who reads a data room today: the S&E head, the deal lead, six function leads and a committee chair

In shortExperts with other day jobs read in parallel against a bid date, and the join between memos falls to whoever has time. No public figure says how the hours divide, so the split below is an example.

From head of S&E to committee chair: the people around one room

Spreadsheet and slides, outside advisers, a scouting firm, or software: four ways teams run it

From a 7:40 am upload to a signed recommendation: HZ-2301 moves into the zone, then dims

In shortEvery watched asset sits on one map. Reading a room moves its dot to where the evidence puts it, then marks the dot with what the files found, so a promising asset carries its open problems in plain sight.

1

Thirteen assets on one map, with the red flags drawn on the dots

Northwind Bio watches 18,240 assets and places 13 on its landscape: strategic fit across, differentiation up, dot size for deal value, color for therapy area. The shaded thesis zone is fit 82 or more and differentiation 65 or more.

TSV-208 from Tessivo Pharma sits at top right, at 92 and 84. It looks like the best asset on the map, but its room raised eight flags, five still open, so the dot is faded, ringed in red and outlined as a deal-breaker.

2

On public data alone, HZ-2301 misses the zone

Hazenmere Pharma's banker opened the room at 7:40 am: 214 files and eight people wall-crossed. HZ-2301 is an oral peptide for a chronic autoimmune disease, with rights open outside Greater China. Public sources stop at Phase 1 in healthy volunteers. Fit 87, differentiation 52, a hollow dot below the line.

3

The files are read in view, each step with its time

The Data Room Reader classifies all 214 files by function, type, version and cut-off, and pulls out 1,370 facts with their pages. Nine superseded drafts are set aside, among them a cohort summary at v0.6. The clinical analyst finds 19 of 31 responding at week 12 against 1 of 10, taken with or without food (blood levels with food came to 0.94 of those without, the fed-to-fasted AUC ratio).

In the sample, two oral rivals in its class must be taken on an empty stomach, per 19 public sources, so differentiation rises from 52 to 81 and the dot crosses the line.

4

Then two flags appear, and the dot dims

Toxicology shows a monkey no-observed-adverse-effect level, the highest dose with no harmful effect seen, of 200 mg/kg/day, about 30 times the patient exposure, on pilot-plant lots only. Manufacturing finds the route changed from lot HZ-DS-2104 with no comparison data (flag RF-41), and the new lots carry a D-arginine epimer at 0.31% that the tested lots did not have (flag RF-42).

The panel keeps both views: public data at 87 and 52, the room at 89 and 81, with a dashed ghost where the dot used to sit. The steps in this sample run add up to 64 minutes.

5

Each memo sentence opens at its page

Six tiles show each memo's score and flags. Lena Ortiz, the CMC lead, opens her draft: every sentence has a numbered citation, and the process report sits beside it at section 6, page 44, with the route change and the planned comparison highlighted. She has two questions to approve for the seller, and only she can sign.

6

The recommendation carries conditions, a lower probability and a later launch

Scored on the weights agreed in advance, the deal comes to 3.5 of 5, rounded from 3.475, just under the proceed line. The draft says proceed with conditions: a comparability study before signing and a bridging toxicology study that qualifies the epimer.

PTRS falls from 16% to 13%, rNPV from $0.74 billion to $0.55 billion, and first launch from 2031 to the first half of 2032. Bid guidance is a license outside Greater China at $180 million to $220 million up front. The offer is due October 31.

The other screens: batch tables, the deal-breaker on TSV-208, seller questions, the committee and your own settings

In shortEverything sits one click from a dot: the proof behind a flag, the register, the seller's answers, the activity trail, and the rules your team sets.

The same sample also holds a third room, BMO-77, where the agents find no granted US composition-of-matter patent and propose an option until one is granted, and a clean deal, QLR-310, which goes to committee at 4.0.

Seven specialist agents that read and draft, and the named people who sign each deal

In shortEach agent has one job and one limit. Function leads sign memos with the meaning recorded, two senior people sign the recommendation, and the committee decides. Agents never sign, send, price or decide.

Who signs what on a Northwind Bio deal

HZ-2301's inputs and outputs: seller files and your rules in, cited memos out

How data rooms, registries, valuation models and identity systems plug in

Design targets for the solution, to test on your own past deal

The bill for a finding made after signing: a study, a delayed launch and sometimes the whole upfront

In shortNo public benchmark prices a missed diligence finding. What you can measure is the money at risk at signing, the cost of a day's delay, and the hours your experts spent on sorting.

Pulling one closed deal apart to find your own numbers

Pick one closed licensing deal and one you passed on. A week of work by one analyst, with the memos, the emails and the room index, will fill this table.

Lost evenings, stale slides and a deal team's credibility

The return on faster diligence: expert hours on each room, and a price that reflects what the files say

In shortHours saved are easy to count and small beside a deal. Price moved by an earlier finding can be large and hard to prove, so test it on a closed deal before you claim it.

Expert hours, the offer, the calendar and scouting reach: how to count each

Work out your return from rooms a year, expert hours and the size of an upfront

With the defaults, hours are worth $48,000 and the price lever about $5 million, a hundred times more. Treat that gap with suspicion. The hours you can show this quarter. The price lever depends on one question: did the earlier flag change an offer that would otherwise have stood? A replay on a closed deal can only hint at the answer, so report it separately, and never add the two into one headline.

Scoring your BD team's readiness: rooms a year, written criteria, a closed deal to replay and a wall-crossed list

In shortEight questions score 0 to 2 each, and a ninth is for thinking. Your answers stay on this page. A low score tells you what to do first, and most of those steps help with or without software.

From 0 to 16: what to do before the next data room

Before a seller's files go in: questions for a diligence software vendor, a data-room provider or a scouting firm

In shortAsk for a demonstration on your files. These questions apply to any vendor, including us, and the strongest answers come with something you can open and check.

A seven-week replay of one closed data room before the next live bid

In shortFreeze what your team concluded, run the first pass on the same files, and count four things: coverage, citation accuracy, flags found and time. No seller or live deal is involved.

1. Clear the reuse. Ask counsel whether the confidentiality agreements allow a restricted replay, and use a room where they do.
2. Pick hard rooms. One with contradictions and a late finding that changed the price, plus one you passed on.
3. Freeze your answers first. The memos, flags, questions, hours and price, in writing, before any automated pass.
4. Use the same readers. Their disagreement with a draft is a finding in itself.
5. Report misses and additions separately. A flag your team missed is the most interesting result, and the easiest to dismiss.

Replay scorecard: what to count on a closed licensing deal

Answers on search and evaluation, licensing due diligence, data rooms and rNPV

In shortThe questions buyers and founders most often search, each answered on its own.