Drug target assessment, explained.

Drug target assessment in plain words: reading before spending

From a protein someone noticed to a program with a budget

1. A candidate appearsFrom a genetic study, a lab screen, a paper, a clinician's observation or a partner's pitch. It is a name and a hunch.
2. Scientists assess itRead the published evidence and internal data, answer six questions, grade each one, and challenge the weak points.
3. The lead signs a recommendationAdvance, advance with conditions, hold for more evidence, or stop, with the reasons and the open questions on the record.
4. The committee decidesIt weighs this target against the others on a common scale, and money moves, or does not.
5. Experiments test it, and the review comes backLab work checks the weakest grades, new studies appear elsewhere, and the assessment is read again.

Step 2 is the subject of this guide, with steps 3 and 5 where they touch it. How any medicine is designed or made is outside this guide.

Few published papers cover safety, and three more surprises about target reviews

Target, modality, loss of function, tractability: terms for your first target review

Fourteen terms are defined in the one-page summary.

The six questions every drug target review has to answer

In shortIdentification names a candidate, validation tests whether it drives the disease, and assessment adds what a company must know before paying: expression (where the target is present), tractability, what happens when the target is removed, and who else is there. Six questions cover that ground, and published research rarely addresses more than one.

Three words get used as if they meant the same thing. Identification produces a candidate. Validation tests a causal claim. The published GOT-IT recommendations say the term "can be understood in numerous ways depending on the context" and "usually describes the technical evaluation of whether a target has a key role in a disease process".2 Assessment is the wider judgment: validation plus the questions a scientific paper rarely asks.

Two public frameworks shape what teams ask. GOT-IT (Guidelines On Target Assessment for Innovative Therapeutics), funded by Germany's Federal Ministry of Education and Research and written for academic scientists, groups the work into five blocks: target-disease linkage, safety aspects, microbial targets, strategic issues, and technical feasibility.2 AstraZeneca's review of its own pipeline produced the "five Rs": right target, right patient, right tissue, right safety and right commercial potential.5 The six questions below are one common way of cutting that ground, and the one this guide follows.

Read that chart as a warning about method. A paper makes one argument, and the link to disease is the argument journals publish. So a review written by summarizing the literature will speak confidently about disease and go quiet on everything else, and a section left empty can read like a question nobody needed to ask. Treating each question as required, with a low grade when evidence is missing, makes the silence visible.

We found no regulator that reviews a target choice: the rules that begin later, and the AI dates that matter

In shortWe found nothing in FDA or EU rules that approves or inspects a company's choice of target. FDA's draft on AI excludes drug discovery. EMA says discovery can be low impact unless its results feed a filing. The EU AI Act excludes systems built solely for research, and the review itself sits outside good laboratory practice. As of November 2026, your governance is your own.

This is the rare part of medicine development with no regulatory filing of its own. No regulator says so in those words; it is our reading of the rules below, and we found no rule that asks an inspector to read a target review. FDA's good laboratory practice rule defines the studies it covers and leaves out "basic exploratory studies carried out to determine whether a test article has any potential utility".6 Rules on trial applications, safety studies and manufacturing all arrive later.

That freedom has limits. A discovery finding can end up inside a filing as part of the reason for a program, and then its origins start to matter. EMA says so: AI in discovery "can be low regulatory impact if non-optimal performance only affects the developer", but if results "contribute to the total body of evidence presented for regulatory review", the principles for non-clinical development should be followed.7

Where drug target evidence lives: papers, genetics portals, biobanks and your own lab

In shortThe evidence sits in at least six separate places, and each answers some questions and not others. Public resources are huge and free. Internal data is the part nobody else can read. Stitching them together, with each claim tied to its source, is where the reading weeks go.

Three terms recur here. A genome-wide study scans genetic variants across many people for one trait. A biobank holds genetic and health data on many people. Omics means large measurements of RNA, proteins and the like in a sample.

Public target portals bring much of this together. The Open Targets Platform, for example, assists "drug target identification and prioritisation using publicly available data".22 It changes often: its June 24, 2026 release added cell-level expression data and new genetic studies, and used a language model to mine clinical records.23

Its own documentation says how to read its target view: green marks "potentially positive attributes" and red "potentially negative attributes", NA means no information is available, and the view describes a target "in a disease agnostic way".24 A portal gives you a starting list and a set of pointers, and grading a claim for your decision stays with your team.

Each source also ages at its own speed. A genome-wide study can change a genetics grade the week it appears. A competitor's program can start with no paper at all. Preprints arrive before peer review, so a team has to decide whether it reads them and how it labels them. Licensed journals add a practical limit: what you may store, quote and pass to a tool is set by your contract, so check before any full text moves.

Why reading target evidence takes weeks, compared with a formal literature review

We found no public figure for how long a target review takes. The nearest published measure comes from a different task with a far higher bar, the formal systematic review of medical treatments. Across 195 such reviews, the mean was 67.3 weeks from registration to publication, with a mean of 5 authors, and only 2.94% of the studies found were included.25 A target review is smaller and quicker than that, and has the same shape: most of what you retrieve does not count, and someone has to read it to find that out.

Target prioritization starts with a grade: how evidence is scored, and how an average can hide a problem

In shortWe found no public scale for target reviews, so each team has to write its own; this guide's example uses A to D per question. A grade is useful when the bar was written before anyone read the evidence, when "independent" is defined, and when the person who sets the bar is different from the one who grades. An overall grade summarizes and can mislead.

Medicine has graded evidence for decades. The best-known scheme, GRADE, rates clinical evidence as high, moderate, low or very low, and lowers the rating for study limitations, inconsistent results, indirectness, imprecision and reporting bias.17 Genetics has its own: the ClinGen framework classes a gene's link to a disease as definitive, strong, moderate, limited, no reported evidence or conflicting evidence.18 Neither was built to rank targets. Both teach the same lesson: state in advance what earns each level, and allow "conflicting" as an honest answer.

A target team can use something simpler, such as A for strong, B for good, C for weak and D for a gap or a red flag. The value lies in the definitions. Here is what a written bar, or rubric, can look like. It is an example to adapt, drawn from the fictional case in this guide.

A worked example: turning six grades for NWT-31 into one overall grade

Northwind's fictional team scores A as 4, B as 3, C as 2 and D as 1, averages the six questions, and reads the result on a scale: 3.15 or above is B+, 2.85 or above is B, 2.6 or above is B−. NWT-31 scores A, B, A, B, C and B. That is 4 + 3 + 4 + 3 + 2 + 3 = 19, and 19 ÷ 6 = 3.17, which is a B+.

So a target with a real open question on safety still reads as B+ overall. That is why careful rubrics add overrides. A D on safety can cap the overall grade at C, whatever the rest says: in the same fictional portfolio, a target scoring C, B, C, A, D, B averages 2.5 and is held at C. A C or D on safety can also call for a second signature. An overall grade helps a committee sort a list. The decision belongs to the six grades, read one at a time.

Three habits that keep target grades honest

What human genetic evidence proves about a drug target, with the numbers worked through

In shortMedicines whose targets have human genetic support are estimated to succeed 2.6 times as often. The evidence counts only when it clears the strict statistical bar for genome scans, points to the right gene, and runs in the right direction. It raises the odds across many programs and promises nothing about one.

Genetics gets special weight because it is close to an experiment on people. Variants are handed out at conception, before any disease or treatment can affect them. That is the logic of Mendelian randomization: genotypes "are independent of non-genetic confounding, and are unmodified by disease processes".26 If people born with less of a protein get a disease less often, that protein probably sits on the causal path.

The size of that advantage has been estimated three times. In 2015, researchers saw the share of mechanisms with genetic support rise from 2.0% at the preclinical stage to 8.2% among approved drugs, and estimated it "could double" success in clinical development.27 A 2019 study confirmed the effect in Phases 2 and 3, and found more than a two-fold gain where the causal gene was clear.28

The 2024 update puts the figure at 2.6 times.3 In the same analysis, relative success was above 2 in 11 of 17 therapy areas and above 3 in hematology, metabolic, respiratory and endocrine disease.3

Four numbers on a genetic claim, read for fictional targets NWT-31 and NWT-30

Direction of effect: the question the 2.6 times figure leaves open

A genetic link tells you a protein matters. Direction tells you what a medicine should do to it. If carriers with less of the protein are protected, lowering it makes sense. If those with less of it get sick, lowering it could cause the disease. The 2024 authors state that "all of our analyses are naive to direction of genetic effect",3 so their figure cannot settle this for you. Your reviewer must state it for every genetic claim.

What does a 2.6 times advantage change for a portfolio? An illustration

The 2.6 figure compares groups of programs. Suppose, purely to show the arithmetic, that 8% of programs without genetic support reach approval. Those with support would reach 8% × 2.6 = 20.8%. To expect one approval you would need 1 ÷ 0.08 = 12.5 programs in the first group and 1 ÷ 0.208 = 4.8 in the second.

The 8% is our assumption. Overall approval sits around 10% of clinical programs.3 The authors add that the measure "does not address the probability" that a gene linked to a disease will yield a successful drug if targeted.3 So it can shape a portfolio, and it cannot promise you this target.

When genetics corrects a story, and when it starts one: HDL and PCSK9

Two public cases show the method at work. For years, high HDL cholesterol looked protective against heart attacks: in observational studies, one standard deviation higher HDL went with an odds ratio of 0.62. A 2012 Mendelian randomization study found that the same rise caused by genes gave 0.93. Its confidence interval, the range likely to hold the true value, ran from 0.68 to 1.26, and that range crosses 1, the point of no effect.26 Genetics said the link was probably not causal.

In the other direction, people with nonsense variants in the PCSK9 gene, which cut a protein short, had 28% lower LDL cholesterol and 88% lower risk of coronary heart disease in a 2006 study of 3,363 Black participants. The authors concluded that "moderate lifelong reduction" in LDL went with a substantial fall in coronary events.32

A reviewer should check Mendelian randomization papers against the STROBE-MR checklist, which asks authors to justify the method, describe how genetic variants were chosen, and report tests of its assumptions.16 One useful property: proteins sit closer to the gene than most risk factors do, which strengthens one of the method's key assumptions.33

Why a published target result may not hold: replication, reagents and animal models

In shortPublished lab results are claims to weigh. In the largest public test, repeated cancer biology effects came out 85% smaller at the median. Many commercial antibodies fail, negative studies stay unpublished, and scientists still disagree about mouse models. A grade has to reflect who found a result, how often, and with what tools.

The Reproducibility Project: Cancer Biology repeated 50 experiments from 23 high-impact papers. It tested 158 effects. For positive effects, the median replication effect was 85% smaller than the original, and 92% of replications came out smaller.4

The authors are careful about what this means: a failed replication "does not disconfirm a finding", though it "does suggest that additional investigation is needed".4 An older analysis estimated that more than half of preclinical research is not reproducible, at about $28 billion a year in the US.34 It is one analysis of past studies, so read it for scale.

What undermines a lab result, and what a target reviewer should check

MELK: a target that survived years of papers and failed one knockout test

A public case shows what the table protects against. Many reports described a protein called MELK as something several cancers depend on, with results from lab tools called RNA interference and small-molecule inhibitors, and a drug described as a MELK inhibitor entered trials.

In 2017, researchers deleted the gene with CRISPR, a precise gene-editing tool: cancer lines grew normally without it, and still died when given the drug.41 That suggests the drug was killing the cells some other way. A 2019 follow-up found the same pattern in other cancer drugs and targets.42 A review that asked "has anyone removed this gene cleanly?" would have graded that biology lower.

A modeling study suggests how much is at stake. It estimated that about one in 200 protein-disease pairings is truly causal, giving a false discovery rate of 92.6% in preclinical research on target selection.43 That number rests on assumptions, and the authors vary them. It does explain why a positive paper alone moves a careful reviewer so little.

What happens when a target is taken away: reading human carriers, knockouts and tissues

In shortThe safety question at this stage concerns biology, before any medicine exists. People who naturally carry less of the protein are the best evidence available, and they are rare. Conflicts between evidence types are the normal case, so a grade has to say which evidence it followed and why.

In this guide, safety means the target's own biology: what the body does without this protein. Later questions about a specific molecule and how it behaves belong to other teams and other documents. At the review stage you are asking what removing or dampening the protein might do, wherever it is found.

Natural human "knockouts" are the most direct evidence. The public gnomAD database identified 443,769 high-confidence predicted loss-of-function variants across 141,456 people.44

Three findings from the same period should shape how you read them. Even essential genes, where such variants are not tolerated, "can be highly successful as targets of inhibitory drugs". For most genes, finding people with both copies affected will take samples about 1,000 times larger than those available, unless recruitment focuses on consanguineous individuals, people whose parents are related. And automated filtering is powerful, though "manual curation remains crucial" to remove errors.45

How the fictional NWT-31 safety grade came out a C

Return to the example at the top. One copy affected, 212 people, 1,400 traits, nothing found. Both copies partly affected, two siblings, chronic fungal infection since childhood. Mice with reduced activity clear a fungal infection more slowly. Low expression in heart, liver and kidney.

Which line wins? The carriers argue for a B. The siblings and the mice point to a real effect that may depend on dose, and a strong medicine might reach it. A cautious reader grades C and names what would raise it. Someone else can reasonably argue for B, with a follow-up study as a condition. Both views should stay on the record, and a scientist decides.

Can the planned modality reach it, and who is already there? Tractability and competition

In shortTractability is a property of a target and a planned kind of medicine together, so it is graded against one stated modality. Competition ages faster than any other evidence. Patent questions go to counsel. Both are business judgments as much as scientific ones.

At the review stage, a tractability grade summarizes whether there is reason to believe the planned kind of medicine could reach the target. Portals sort the question into plain categories: is the protein on the cell surface, has a binder been reported, are good research tools available, has any drug on this target reached the clinic?24

The reviewer's job is to record what the evidence says for the modality the team intends, and to keep measured results apart from predictions. A predicted structure or binding site is a hypothesis. How to design a medicine is separate work for specialists, and has no place in the review.

A grade that names no modality says little, because the same protein can be a poor fit for one kind of medicine and a good fit for another. A sensible template refuses to plan the review until the modality is stated, and regrades if the plan changes.

Why the competition section ages first, and how a committee uses it

Competition has a different job from the other questions. Nothing in it says whether the biology is right. It says whether being right would still be worth the money by the time you arrive. A review might record each rival program by company, modality and phase, with the date of the check. The fictional template in this guide treats a list older than 30 days as out of date.

A rival's success can also help: it is evidence that the target works in people, called clinical precedent. So a crowded field can raise biology confidence while lowering commercial appeal.

Eight ways a target review misleads a portfolio committee

In shortMost failures are ordinary: a claim with no passage behind it, one study counted twice, a grade nobody can explain, a document nobody re-reads. We found no inspection record for this work, so we list errors from the published literature and from the logic of the method, and mark our judgment as ours.

A pattern also shows up in what happened to decisions. In Pfizer's review of 44 Phase 2 programs, most failures came from lack of efficacy, and in 43% of cases it "was not possible to conclude whether the mechanism had been tested adequately".49 So the company had paid for a trial and still could not say whether its target idea was wrong. A review that records what would count as a fair test helps with that later question.

What choosing the wrong target costs a drug program, and when you find out

In shortAbout one in ten clinical programs reaches approval. Lack of efficacy is the major cause,43 and that points back to the target. A review costs weeks of a scientist's time, by our estimate; a Phase 2 trial averaged $13.35 million in a 2014 federal study. We found no study that measured how many failures a better review would have prevented.

Estimates differ by what they count. Wong and colleagues, using 406,038 trial records, put the path from Phase 1 to approval at 13.8%.50 Minikel and colleagues say "about 10%" of clinical programs.3 Both are defensible, because they count different groups.

The loss after Phase 2 is the one most tied to target choice, since that is where efficacy is first seriously tested. One trade report of AstraZeneca's review of 142 projects from 2005 to 2010 says that 65% of projects "failed during phase II due to lack of efficacy", in many cases because of a poor understanding of the target. It does not say 65% of which group. We read that report and did not read the paper itself.51

That spread matters inside a portfolio. A B-graded target in oncology and a B-graded target in vaccines carry different odds, so a committee should keep therapy-area base rates beside the grades and avoid applying one number across the portfolio.

The cost of a new medicine, and why estimates differ by a factor of two or more

Before their own estimate, Wouters and colleagues note earlier figures from $314 million to $2.8 billion.52 What each estimate counts explains much of the gap. A capitalized figure also charges for the return an investor requires while the money is tied up.

The AstraZeneca record: what the five Rs did, in the company's own account

AstraZeneca is the best-known public test of whether more discipline pays. After its 2011 revision around the five Rs, success from candidate drug nomination to Phase 3 completion rose from 4% in 2005-2010 to 19% in 2012-2016.55 Read this carefully. It is one company's own account, many things changed at once, and the authors say only that the framework is "beginning to have an impact". It is the most encouraging number available, and it does not isolate the effect of a target review.

What AI changes in target review, and where a scientist still has to decide

In shortLanguage models help find, read, extract and re-check literature at a speed people cannot match. They can also invent references, miss what a table says, and sound confident about thin evidence. Keep every claim tied to its passage, separate writing from challenging, and leave the decision to a named scientist.

Target review suits this technology on the surface: lots of text, a fixed set of questions, and a need to look again when something new appears.

One published screening tool, tested by its developers on 10 Cochrane review libraries with over 58,000 references, reached a recall of 1.00 with a specificity of 0.61, meaning that it missed no relevant paper in those tests and still passed many irrelevant ones along for people to read.56 That is a sensible division of labor for early filtering, and an evaluation you would want repeated on your own subject.

Citations are where a general chatbot fails most visibly, so test that directly. Asked to write reviews in 2023, GPT-3.5 fabricated 55% of its citations and GPT-4 18%.57 A separate 2023 study of GPT-3.5 medical papers found 47% fabricated.58 The problem has shrunk and has not gone: in a test of ChatGPT-5 on orthopedic guidelines, run in early 2026, 7.13% of references were fabricated and only 49.34% were bibliographically accurate.59 The authors conclude that independent verification "remains necessary".

Three limits on AI that better models will not remove from target review

How discovery teams run a target review today: people, hand-offs and the reading weeks

In shortBy our account, one scientist does most of the reading, pulling in specialists for chemistry, safety and competitors. The result is a document or slide deck, signed by a therapeutic-area head and brought to a committee. There is no public benchmark for time. In our estimate most days go to reading and writing, and little goes to arguing about what the evidence means.

Who shapes a target decision, from the requesting scientist to the committee

In house, outsourced or software: three ways of producing a target review

From a graded matrix to a conditional yes: how Dr. Maya Chen decides on NWT-31

In shortIt is Wednesday, October 7, 2026 at Northwind Bio. Dr. Maya Chen, Head of Immunology target discovery, has NWT-31 in ulcerative colitis to sign before the October 22 committee. AI agents researched it, graded it against her team's bar and challenged the result. She reads the evidence, decides each challenge, and signs with conditions.

1

Twelve targets graded on six questions, in one view

Targets run down the side, grouped by therapeutic area, with the six questions across. Each cell is a grade from A to D, with the overall grade and stage at the right. There are 53 grades and 851 cited claims behind them. A dark bar above the matrix shows the evidence watch and what it last found.

The first thing the eye finds is the pale C in the NWT-31 row, then the red D on a target already stopped.

2

One click on the safety C shows why it is a C

The panel opens the claims behind the grade: 212 people with one faulty copy and no immune deficiency across 1,400 traits; two siblings with chronic fungal infection since childhood; mice that clear a fungal infection more slowly; low expression in heart, liver and kidney.

Each claim carries its evidence type, so a human case report never reads like a mouse model.

3

Every claim opens the passage it came from

In the full assessment, section 5 of 6 asks the same question in words: what happens when the target is taken away? The grade, the agent's stated confidence of 81%, the summary and the key claims sit in the middle. On the right, each source shows the passage with the supporting words highlighted, and a link opens it in place.

The rail on the left shows the six grades and how far each is from settled.

4

The critic argues the other side, and Dr. Chen decides

A separate agent reads each section as a skeptical committee member would. On safety it argues that two siblings are a small record, that 212 carriers are healthy, and that the risk looks dose-dependent. It proposes moving from C to B, with a fungal-infection study as a condition.

She can accept, keep the grade as it is, or ask the agent to look again. The critic never changes a grade itself. A resolved challenge sits beneath it, with its reason on the record.

5

Three targets compared on one rubric

For the committee she opens NWT-31 beside NWT-33 and NWT-35. Each has its shape on one radar and its grades in one table. NWT-31 is strongest on genetics and pathway biology, with safety to watch. NWT-33 leads on expression and tractability for its planned modality. NWT-35 leads on genetics.

Because all three used one bar, the argument in the room can be about the targets.

6

She signs with conditions, and the safety grade takes a second signature

She chooses Advance with conditions over the other three options. The open questions become the conditions: whether the closest family member matters in human heart cells, a freedom-to-operate opinion from counsel, a serum marker of pathway activity, and a rival's Phase 1 results. Two open challenges are recorded with her signature.

The meaning of the signature is printed on the screen: scientific approval of this assessment. Because safety is a C, Lena Ortiz co-signs after her.

A held target gets new genetics, and a new target is researched while you watch

In shortThe same morning, a weekly re-check finds a new genome-wide study on a target held for weak genetics. Its grade moves from C to B, in view, and its owner is asked to confirm before the committee can use it. A new request goes from one line to a graded, challenged assessment for a scientist to review.

1

NWT-30 was held because nobody had replicated its genetics

On September 21, Dr. Aiko Tanaka signed NWT-30 in ALS as Hold for more evidence. Its genetics rested on one small study at p = 4 × 10⁻⁵, 800 times weaker than the genome-wide bar. Overall: B−.

2

New study, new citation, a grade moved in plain view

The watch reads 214 new papers and preprints. One fictional meta-analysis of 31,480 ALS cases and 118,930 controls replicates the region at p = 2.8 × 10⁻⁹, and the risk allele raises the target's expression in motor neurons. The genetics cell flips from C to B with a "was C" tag. The overall goes from B− to B, since 17 points became 18, and 18 ÷ 6 = 3.0.

It stays a B until fine-mapping is published. Dr. Tanaka confirms before any committee pack uses it, and a new open question is added.

3

Sam Patel types one line: NWT-44 in atopic dermatitis

A computational biologist asks for an assessment for the November 19 committee. The planner splits the template into 6 sections and 38 research questions, sets the bar for each grade, and never grades anything itself.

4

Six agents research in parallel, and the scorecard fills as grades land

Each agent has its own lane, showing what it reads and how many sources and claims it has. Genetics, expression, biology, tractability, safety and competition run at once. A scientist can watch or go to lunch.

5

Built: 128 claims, 54 sources, a B+, and a problem named

Expression and competition come back A. Tractability comes back C, since the evidence for the planned modality is thin. The critic raises 6 challenges, resolves 5 with the agents and leaves 1 for the owner. The assembler routes the result to Dr. Chen. That B+ is 19 points over six questions.

A person still has to read it. Agent time ran to minutes; the days that follow belong to scientists.

The target list, the full assessment, the review record, the dashboard and your rules

In shortBeyond the matrix, each screen answers one question. Which targets need me? What does the whole document say? Who did what, and when? Where do the conditions stand? Where does the evidence come from? And what is your bar? Each figure here comes from the sample portfolio.

What the sample dashboard shows about grades and evidence mix

Design targets for a Botminds target assessment deployment

Nine agents, three human signatures: what goes into a target assessment and what comes out

In shortOne agent plans, six research, one challenges and one assembles. None of them decides. The therapeutic-area lead signs, a translational safety scientist co-signs a weak safety grade, and the owner confirms any grade that new evidence moved. The assessment says nothing on how a medicine should be designed or made.

Who signs a target assessment, and what each signature means

Target assessments are research records, marked as such. These controls support your compliance if a finding later feeds regulated work; they make no one compliant by themselves. On credentials, Botminds is SOC 2 Type II audited, with information security managed to ISO 27001.

From literature and your own screens to a signed target assessment

Connecting literature licenses, omics stores and competitor databases

Roles, audit trail and scale for discovery organizations

Roles separate who may request, edit, sign, co-sign and read. Every change to the template, the bar or the sources is recorded. Each therapeutic area can run its own targets under a shared rubric, and committee secretaries build packs from signed versions. Your own data stays under your access rules, and agents read only the sources you switch on.

What a slow or weak target review costs your discovery team

In shortThe time is modest and visible. The expensive errors are later and hard to count: a trial on a target that was never tested fairly, a hold nobody reopened, a committee hour spent re-arguing a paper. We found no study that isolates these costs, so measure your own in a week before you believe anyone's figure.

That ratio shows scale: review time is cheap against what depends on it, so a better review needs to change few decisions to pay for itself. It is also easy to misuse. We found no test of whether a review would have prevented any given failure.

Auditing your last ten target reviews to find your own numbers

Costs of a weak target review that no calculator captures

Five levers on target review value, and a calculator for your own portfolio

In shortScientist hours are the part you can count, and they may pay for a tool alone at enough volume. Avoided mistakes could be worth far more and cannot be proven in advance, so keep them in a separate line and be skeptical of anyone who adds them up for you.

The 160 hours is the solution's own estimate of 20 working days by hand, at 8 hours, and no public figure backs it. The 41 hours treats the solution's 5.1-day design target as full days of people's time, which is cautious. The trial cost is ASPE's 2014 average for one Phase 2 trial, unadjusted.54 A default of 0.1 means one trial every ten years, chosen so it does not swamp the sum. Look at "Time alone" first, and treat the trial line as a question to test with your own closure reports.

Look for the shape in any vendor's numbers, including that one. If challenge and review time also fall to nothing, someone has probably skipped them.

Is your discovery team ready to automate target reviews? Eight questions and a score

In shortScore each question 0, 1 or 2. Write down your rubric, your owner and your baseline first. Without them, a tool only makes your current habits faster.

Answer honestly; the last question is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like. The figures in those answers are examples.

What your target review readiness score means

What to ask a target assessment vendor about citations, grades and who decides

In shortAsk every vendor and research firm the same questions, us included. Good answers can be shown on your own target in a meeting. Be wary of accuracy claims with no test set behind them, and of any tool whose grade can change with no human on the record.

Rebuild past target reviews first: an eight-week pilot your scientists can mark

In shortStart with targets you have already decided, so nothing live is at risk. Keep your own review as the answer key, plant a few errors, and have scientists mark each claim. Then run one live target through a real committee cycle, and decide on evidence by week eight.

1. Write the bar and clear the sources with library and legal.
2. Pick eight past reviews, including two your scientists now believe were wrong, and record how long each took.
3. Plant checks: a retracted paper, a single-laboratory claim, two papers on one group of people. Keep the key outside the system.
4. Rebuild blind, then have two scientists mark every claim for accuracy and every grade for agreement.
5. Run one live target through your next committee cycle, with your people signing, and decide.

What to measure when you test a target assessment tool

What people ask about drug target assessment and validation

In shortThe questions people search about target assessment, each answered on its own.