Indication prioritisation in under 20 days: the scoring framework

For a single-asset biotech, the choice of which disease to point the molecule at predicts approval better than the choice of molecule. That is not a rhetorical flourish. When BIO, Informa Pharma Intelligence and QLS Advisors ran a machine-learning feature-importance analysis across 12,728 phase transitions from 9,704 development programmes between 2011 and 2020, the indication was consistently ranked the top variable across all clinical development phases in predicting whether a drug would be approved. Ahead of the biological target, ahead of the modality, ahead of whether the programme was even running in its lead indication. [1]

So the highest-leverage decision on your board deck, the indication-prioritisation call, is the one most teams make on scientific intuition and a market-size slide.

You know the position. One or two assets, a molecule with three or four plausible indications, and a board asking which one you run first, with the answer wanted in weeks of runway rather than months of open-ended strategy work. There is a right way to make that call and two popular wrong ones. The first is the market-size reflex: rank the indications by the size of the prize and chase the biggest. The second is quieter and more respectable. Convene the experts, let the room debate, write down whatever the loudest voice talked everyone into. Call it committee laundering: groupthink with a spreadsheet bolted on afterwards for decency.

Both feel rigorous, and neither survives contact with the numbers. The alternative is to score the indication rather than argue it, on a small set of named factors, combined into one explicit weighted number you can defend and re-run. This piece lays out that model, and a realistic shape for running it in about twenty working days.

Therapeutic area is the wrong unit of analysis

Start with the factor that does most of the work, because it also explains why the market-size reflex fails.

Most teams reach for a therapeutic-area success rate. "Oncology's tough, roughly one in twenty." That average is close to useless for choosing between two oncology indications, because the variance inside a therapeutic area dwarfs the variance between them. In the BIO/Informa/QLS dataset, likelihood of approval from Phase I is 7.9% overall, but by disease area it spans 23.9% in haematology down to 3.6% in urology: a roughly seven-fold spread across fourteen areas. [2]

Now go inside a single therapeutic area. Within oncology alone, likelihood of approval ranges from 1.1% for pancreatic cancer (n=275) to 15.2% for alimentary-tract cancers (n=33), a fourteen-fold spread inside one disease grouping in the same BIO/Informa/QLS dataset. [3] Think of it as two doors in the same building, one opening onto a one-in-ninety chance of approval, the door beside it onto better than one in seven. "We're an oncology company" tells you nothing about which door you are standing in front of.

The indication-level literature says the same thing from the bottom up. Three independent studies benchmark specific indications against a roughly 10–11% convention: non-small-cell lung cancer at 11%, where biomarker-based patient selection is associated with about a six-fold increase in success (Falconi 2014); gastric cancer at 7%, where biologics succeeded 17% of the time against 1% for small molecules (Dhillon 2023); and multiple sclerosis at 27%, nearly three times the benchmark, sitting at the far end of the spread (De Gasperis-Brigante 2016). [4][5][6]

A word of honesty about these numbers. There is no single "true" success rate for any indication. Wong, Siah and Lo (2019) put oncology's overall rate at 3.4% in their sample against 5.1% in prior studies; Pronker (2011) found published cumulative estimates ranging from 7% to 78% depending on method and vintage. [7][8] That spread is precisely why you anchor on one internally consistent dataset's disease-area table and adjust from there, rather than splicing headline figures from different studies and different decades. Compute the number; don't inherit one.

Monday's version: never let "our therapeutic area's success rate is about X%" into the model, and score each candidate indication on its own odds.

The five factors that actually move indication-prioritisation odds

• Indication-specific probability of technical success. The load-bearing column, for all the reasons above. Derive an honest number: start from the disease-area or indication figure in the BIO/Informa/QLS data, then adjust for what that same data supports. Biomarker-preselected programmes showed roughly two-fold higher likelihood of approval, at 15.9%, driven by Phase II success near one in two (46.3%). [9] Modality matters too. Rare-disease therapies came in at 17.0% against 5.9% for chronic, high-prevalence disease, with oncology excluded from both. [10] These are modifiers you can defend, not guesses.

• Unmet need. Source this to an agency's own language. The EMA's orphan-designation criteria set a usable bar: a life-threatening or chronically debilitating condition, with no satisfactory authorised method of diagnosis, prevention or treatment, or, where one exists, significant benefit over it. [11] Its PRIME scheme defines unmet need as where "no treatment option exists, or where they can offer a major therapeutic advantage over existing treatments," requiring "meaningful improvement of clinical outcomes." [12] This is anything but a soft factor: in a validated MCDA rollout in Catalonia, "the most important evaluation criteria identified for orphan drugs were: 'disease severity', 'unmet needs' and 'comparative effectiveness', while the 'size of the population' had the lowest relevance." [13] Hold that finding for the market-size question below.

• Regulatory-pathway feasibility. Which expedited routes does this indication plausibly open? Ground it in the agencies' own thresholds rather than academic MCDA, which does not name this criterion. The EMA orphan route turns on a prevalence of no more than 5 in 10,000 in the EU; [11] PRIME is the EU's expedited analogue to breakthrough designation. [12] And the odds follow the pathway: rare-disease likelihood of approval of 17.0% against 5.9% for high-prevalence disease is, in part, this factor showing up in the outcomes. [10]

• Competitive density. The most numbers-light factor, so handle it as a structured judgement rather than a fake metric with invented decimals. A same-day scan of ClinicalTrials.gov for the candidate indication tells you how crowded the field is and where the white space sits. As illustration only: the RAINBOW trial (NCT06371417) tests six indications in parallel rather than ranking them, which is the "test everything at once" alternative to scoring, affordable only if you are not the one paying for all six; [14] and FUZE (NCT03834220), terminated, is the standing reminder that a clever basket design is not a shortcut around indication-level risk. [15]

• Realistic time-to-first-readout. Speed to a value-inflecting signal, not speed to approval. The BIO/Informa/QLS data put average time from Phase I to approval at about 10.5 years (2.3, 3.6, 3.3 and 1.3 years by phase), and the higher-odds disease areas tended to run shorter timelines too. [16] A better indication raises your odds and shortens the path to the readout that refinances the company. For a team burning runway, that second effect is often the whole point, not a tie-breaker.

"Where's market size?" It sits at the door, not in the grid, as a threshold gate you apply before scoring, where a market too small to sustain the company fails the gate and drops out. Everything past the gate is judged on the odds of getting there and the speed of getting there, not the size of the prize, because over-weighting the prize is precisely the reflex this whole framework exists to break. Put it back as a sixth weighted column and you have quietly rebuilt the thing you were trying to escape. (Recall Catalonia: population size ranked least relevant of all.)

Monday's version: put these five columns on one page for every candidate indication this week, and keep market size off the scoring grid on purpose. If you already run an impact/feasibility filter for evidence generation, this is the same discipline aimed one level up, at the indication itself.

Turning five columns into one number without faking the decimals

Four rules keep the arithmetic honest:

• Score in ordinal bands, say 1 to 5, not spurious decimals. A weighted score is a smoke detector, not a fire marshal. It tells you which room to look in. It does not put anything out, and dressing it in three decimal places only pretends otherwise.

• Set the weights before you see the data. Agree what your strategy values, in the abstract, and lock it. Weight after the numbers are in and you will, unconsciously, weight your way to the answer you already wanted.

• Treat any published weighting as one illustration, never a benchmark. You will find numeric splits online, for instance a five-axis 25/20/20/15/20 matrix. [19] Fine as a worked example of expressing weights as numbers. It is unvalidated content aimed at search engines, not a survey, and I would not let it near a board slide dressed as one.

• Run a sensitivity check. Nudge the weights and see whether the ranking holds. 

Monday's version: build the grid in ordinal bands with weights locked in advance, then torture-test the ranking with a sensitivity pass before it goes anywhere near the board.

"A weighted score is just groupthink with a spreadsheet"

You are taking a success figure the literature itself spreads from 7% to 78%, a competitive-density call that is pure judgement, and a set of weights expert panels demonstrably cannot agree on. You multiply them together, hand a board a number with a decimal point on it, and the decimal point does the persuading. That is false precision. Parmar's own paper flags the "reliance on expert opinions" and "instances of initial discordant results." [18] So is the scored model just committee laundering with extra steps and a nicer output?

I have watched that failure mode up close. Years ago, on a small oncology programme, I sat in a prioritisation meeting that ran two hours, was settled in the last ten minutes by the most senior person in the room stating a preference, and was written up the following week as a weighted scoring exercise. The number was real, but the decision it supposedly produced had been made before anyone opened the model.

Here is why the answer is still to score. The honest alternative to a soft-input score is not a hard-input one, because no hard-input version exists. The honest alternative is the unweighted room, which has every one of those same soft inputs plus whoever argues hardest and stays latest. Scoring does not remove the judgement; it drags the judgement into the open, one factor at a time, where each assumption can be named and contested, and the sensitivity pass shows exactly which factor is carrying the ranking. That is Rosati's (2002) old case for probability-weighted decision analysis over simple net-present-value ranking. The point was never to make the model objective. Its worth is that it makes the odds explicit, instead of burying them inside a market forecast where nobody can see them move. [20] The score is not the decision. It is the structured argument that produces one, and unlike the meeting, it is still legible in six months when the board reopens the question.

Monday's version: report the ranking's sensitivity, not a false point estimate, and treat any indication whose lead evaporates under a small weight change as unranked.

Twenty days, and why it isn't twelve weeks

Here is the shape, four phases across roughly twenty working days.

1. Days 1–3, frame and gate: list the candidate indications, apply the market-size and strategic-fit threshold gate to cut the non-viable ones, and agree the five factors and provisional weights before pulling a single data point.

2. Days 4–12, pull the numbers: this is the bulk of the work and the genuinely parallel part, covering indication-specific success odds from the indication literature and your biomarker and modality adjustments, unmet-need and regulatory-pathway evidence from the EMA criteria and the epidemiology, a ClinicalTrials.gov competitive scan, and a rough time-to-readout model. The speed comes from running four workstreams at once instead of in series, not from cutting corners.

3. Days 13–16, score, weight, rank: standardise to ordinal bands, apply the pre-set weights, compute the weighted sums, rank the indications, and run the sensitivity pass.

4. Days 17–20, pressure-test and document: hold one structured challenge session rather than a month of standing committees, then write the rationale down as a living document, so the decision is defensible now and re-runnable when the competitive or regulatory landscape shifts under it.

Monday's version: you can lift that shape onto a calendar today. The honest limiter is data readiness and a decision-maker who will actually decide, not the maths.

What the score buys, and what it doesn't

Run this indication-prioritisation score properly and you get a ranked, weighted decision you can audit, with the assumptions argued in the open, that a board or an investor can interrogate but cannot easily unpick, plus the discipline of having priced the odds instead of the prize. What you do not get is certainty. The model is a smoke detector: it tells you which room to look in, but it does not put out the fire, and no honest framework will claim otherwise.

References

[1] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Machine-learning feature-importance finding: the indication ranked the top variable across all phases; 12,728 phase transitions, 9,704 programmes.)

[2] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Likelihood of approval from Phase I: 7.9% overall; haematology 23.9%, urology 3.6%; 14 disease areas.)

[3] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Intra-oncology likelihood of approval: pancreatic 1.1%, n=275; alimentary cancers 15.2%, n=33.)

[4] Falconi et al. (2014). J Thorac Oncol. PMID: 24419412. https://pubmed.ncbi.nlm.nih.gov/24419412/ (NSCLC probability of technical success ~11%; biomarker use associated with ~6-fold increase.)

[5] Dhillon et al. (2023). Am J Clin Oncol. PMID: 36662871. https://pubmed.ncbi.nlm.nih.gov/36662871/ (Gastric cancer PTS 7% vs ~11% average; biologics 17% vs small molecules 1%.)

[6] De Gasperis-Brigante et al. (2016). Mult Scler Relat Disord. PMID: 26856949. https://pubmed.ncbi.nlm.nih.gov/26856949/ (Multiple sclerosis PTS 27%, ~3x the ~10% benchmark.)

[7] Wong, Siah & Lo (2019). Biostatistics. PMID: 29394327. https://pubmed.ncbi.nlm.nih.gov/29394327/ (Oncology 3.4% success in sample vs 5.1% in prior studies.)

[8] Pronker et al. (2011). Vaccine. PMID: 21722688. https://pubmed.ncbi.nlm.nih.gov/21722688/ (Published cumulative PTS estimates range 7%–78% depending on source and method.)

[9] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Biomarker-preselected programmes: two-fold higher likelihood of approval, 15.9%; Phase II success 46.3%.)

[10] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Rare-disease LOA 17.0% vs chronic/high-prevalence 5.9%, oncology excluded.)

[11] European Medicines Agency. "Orphan designation: overview." https://www.ema.europa.eu (Criteria: life-threatening or chronically debilitating; prevalence ≤5 in 10,000 in the EU, or insufficient return; no satisfactory authorised method, or significant benefit over an existing one.)

[12] European Medicines Agency. "PRIME: priority medicines." https://www.ema.europa.eu (Unmet need where "no treatment option exists, or where they can offer a major therapeutic advantage over existing treatments," requiring "meaningful improvement of clinical outcomes.")

[13] Gilabert-Perramon et al. (2017). Int J Technol Assess Health Care. PMID: 28434413. https://pubmed.ncbi.nlm.nih.gov/28434413/ (Validated MCDA rollout, Catalonia: disease severity, unmet needs and comparative effectiveness ranked most important; population size least relevant.)

[14] Sponsor of RAINBOW. "RAINBOW." ClinicalTrials.gov: NCT06371417. https://clinicaltrials.gov/study/NCT06371417 (Illustrative only: six indications tested in parallel rather than ranked.)

[15] Sponsor of FUZE. "FUZE." ClinicalTrials.gov: NCT03834220. https://clinicaltrials.gov/study/NCT03834220 (Illustrative only: terminated basket trial.)

[16] BIO, Informa Pharma Intelligence & QLS Advisors. "Clinical Development Success Rates and Contributing Factors 2011–2020." February 2021. (Average time Phase I to approval ~10.5 years: 2.3/3.6/3.3/1.3 years by phase; higher-LOA areas tend to shorter timelines.)

[17] Wagner et al. (2018). Adv Ther. PMID: 29270780. https://pubmed.ncbi.nlm.nih.gov/29270780/ (GEP-NET MCDA: linear-additive weighted scoring; expert weight variance, SD ~±0.12 for overall survival and ~±0.15 for progression-free survival.)

[18] Parmar et al. (2023). Curr Oncol. PMID: 37185399. https://pubmed.ncbi.nlm.nih.gov/37185399/ (CanREValue: equal-start weights, no "correct" set of weights; documented reliance on expert opinions and initial discordant results.)

[19] IntuitionLabs. Illustrative five-axis indication-prioritisation weighting matrix (25/20/20/15/20). Unvalidated web content, used here only as an example of expressing weights numerically.

[20] Rosati (2002). Expert Rev Pharmacoecon Outcomes Res. PMID: 19807328. https://pubmed.ncbi.nlm.nih.gov/19807328/ (Argument that probability-weighted decision analysis outperforms simple NPV/market-size ranking; argument cited, no figures.)

[21] Umbrex. "Biotechnology: Portfolio Strategy and Indication Prioritization." (Illustrative 12-week / 60-working-day full-rigour engagement, caveated as varying with "portfolio complexity, data readiness, and stakeholder availability.")