Model card, methodology M5

• Not backtested. PhaseFolio has not tested this model's outputs against realised outcomes. Earlier internal backtests were retired in M4 and are kept only as historical records.
• Reference data, not a prediction. The default probabilities, costs, margins and rates are starting values taken from published sources, public filings and statute, plus assumptions PhaseFolio has labelled as its own. They describe groups of historical programs or companies, not your asset.
• Not advice. Nothing the model produces is investment, legal, tax, medical or regulatory advice, a fairness opinion or a fair-value measurement.
• Your values, your judgement. Every default can be changed in the asset modeler, and the model computes with the values you keep. You are responsible for the values you keep and for decisions made with the output.
• Use at your own risk. The Terms of Service govern use of PhaseFolio, including the sections on models and consequential decisions, warranties and limitation of liability.

PhaseFolio values clinical-stage drug assets with risk-adjusted net present value (rNPV). It is built for people who already make these judgements: business development and licensing teams, biotech investors, technology-transfer offices and analysts preparing an internal view of an asset, a deal or a portfolio. It gives them a transparent, editable starting model and records which inputs came from where.

It is not built for, and should not be used for:

• treatment, prescribing or any clinical decision about a patient;
• a prediction that a specific program will or will not succeed;
• IFRS 13 or US GAAP fair-value measurement, Section 409A valuations, litigation or damages work, or a fairness opinion;
• ranking the efficacy of drugs across studies;
• any decision made on the defaults alone, without a person reviewing and owning the inputs.

Development is treated as a sequence of gated stages. Each stage has a probability of success, a cost and a duration. The cumulative probability of approval is the product of the stage probabilities from the asset's current stage through approval.

Each stage cost is discounted at the weighted average cost of capital (WACC) from the stage midpoint and weighted by the probability of reaching that stage. On approval the model builds an annual revenue schedule (ramp, plateau and loss-of-exclusivity erosion inside the exclusivity window), converts gross revenue to net with net = gross × (1 − COGS% − SG&A%) × (1 − tax rate), applies the Inflation Reduction Act price cliff to US revenue where it applies, and discounts to present value.

The headline is rNPV = revenue PV × cumulative probability of approval − total risk-adjusted cost. The model also reports the unadjusted NPV (approval assumed), a Monte Carlo distribution whose mean is the expected NPV, and a one-at-a-time tornado sensitivity. Deal structures (royalties, milestones, profit splits) transform the same revenue and cost series into the partner's position.

The full equations are on the rNPV model page. The method follows the risk-adjusted valuation approach described by Stewart, Allison and Johnson (2001).

The defaults do not all have the same standing. The table states the basis of each one. "PhaseFolio assumption" means we chose the value ourselves and label it as ours.

Details and caveats for each default are on its own page: PoS calibration, COGS benchmark, SG&A benchmark, tax benchmark and IRA framework.

BIO 2021 reports success rates by disease area, by modality and by biomarker use, but never crossed. Each PhaseFolio cell starts from the area row and adds, in log odds and stage by stage, how far the modality row and the biomarker row sit from the all-indications row:

• Area rows come from BIO 2021 Figures 2, 7 and 8b. Hematologic oncology uses the Figure 7 hematologic row. Dermatology has no BIO row and uses the Other row as a proxy.
• Modality rows come from Figure 10b. Bispecifics use the monoclonal antibody row as a proxy. Cell therapy and other modalities take no modality shift. Sparse rows are smoothed as (k + 0.5) / (n + 1).
• Biomarker rows come from Figure 11. A companion diagnostic takes the preselection row; enrichment takes half the preselection shift, which is a PhaseFolio assumption.
• Every BIO row used reproduces its printed likelihood of approval, and a test holds the model table to this derivation.

Adding the two shifts assumes that modality and biomarker effects are independent of the disease area and of each other. That is PhaseFolio's modelling assumption; BIO 2021 does not report the crossed rates, so it has not been tested against them. The PoS reference data explorer shows every cell beside its BIO 2021 area row with the stage-by-stage derivation, so the difference can always be traced to the published rows it came from.

Multipliers are optional. Genetic validation, orphan designation and first-in-class apply only when a scenario is marked with them; biomarker quality applies only when a reviewed classification exists. Preclinical and NDA/BLA rates are never adjusted.

Where a source reports a relative success rate, PhaseFolio applies the same number as an odds ratio in log-odds space so a probability can never exceed 100%. A relative rate and an odds ratio are not the same quantity, and the gap grows as the base probability rises. Using one as the other is a PhaseFolio modelling choice, disclosed here and on the PoS calibration page.

Phase 1 objective response rate is extracted as a flag only and does not change any value.

• Historical transition rates from 2011 to 2020 are a reasonable starting point for programs today.
• Modality and biomarker effects add in log odds, independent of disease area (section 5).
• Relative success rates can be applied as odds ratios (section 6).
• Preclinical rates, the first-in-class multiplier, the enrichment half-shift and every stage cost and duration default are PhaseFolio's own values.
• Newer modalities route to the nearest baseline: RNA therapeutics and oncolytic viruses to gene therapy, protein degraders to small molecule, radioligands and cancer vaccines to other. This is a proxy, not a claim that the risks are the same.
• Revenue follows a stylised ramp, plateau and erosion curve, annually, with costs timed at stage midpoints.
• Operating margin is additive: COGS% and SG&A% are subtracted from the same gross revenue.

• No backtest. The outputs have not been compared with what happened to real programs (section 10).
• Population rates. Every default describes a group. It carries no information about a specific molecule, sponsor, trial design or dataset.
• Coverage. Rare disease has no literature cost profile. Dermatology and several modalities rely on proxy rows. Biomarker quality is scoped to oncology solid tumors.
• Age of sources. BIO 2021 ends in 2020 and Schwaederle et al. was published in 2016; practice has changed since.
• Uncertainty. The deterministic rNPV is a point estimate; read uncertainty from the Monte Carlo distribution, which is only as good as the input ranges it samples.
• Sensitivity is one at a time and does not capture interactions between inputs.
• Loss-of-exclusivity erosion uses stylised small-molecule and biologic curves, not forecasts of generic or biosimilar entry.

• M2 found that a hand-typed copy of the baseline table on the methodology page had drifted from the table the model computes with. The page now renders from the model table, and a test keeps them together. Exports for five areas made before M2 were computed on the model table, not on the drifted copy.
• M2 removed an attribution to Citeline that the sources did not support.
• M4 derived every Phase I to NDA/BLA rate from named BIO 2021 rows. That filled 125 cells that had fallen back to an area's other row and corrected hematologic oncology, whose earlier rows sat well above BIO's hematologic cancers group.
• M4 removed the CAR-T modifier, which moved the relevant cell from 7.6% to 7.5% and so did almost nothing, and replaced two stacked gene therapy modifiers with one calibrated value.
• M4 retired the published backtests. They were small, internal, run on earlier model versions and not independently reviewed, and they could be read as validation they were not.
• In May 2026 PhaseFolio ran two internal checks on a small set of oncology programs. They are recorded, as written at the time, in the internal checks record. They are not a validation of the model and were not peer reviewed.
• M5 rewrote the public methodology as reference data: every default is labelled with its source, the estimand the source reports and how the model uses it.

This edition makes no claim about predictive accuracy. A public evaluation of the model will be published only when it has been designed and signed by qualified statistics or bioinformatics staff, with its data, method and results open to review. Until then, the model card reports the model's sources, assumptions and limits, and nothing more.

Later editions will add sections as that work happens: out-of-sample evaluation, calibration of the stage probabilities, and any change to the defaults it supports.

A new edition of the model card is published with each methodology release. Earlier editions stay available at their own address and are never rewritten. Section-by-section changes are listed in the version history. Signed exports carry the methodology release and model version they were computed with.

Methodology and model card by Tamal Adebisi, founder and methodology author, PhaseFolio.

With thanks to Jeromie Ballreich, PhD, Associate Research Professor in the Department of Health Policy and Management at the Johns Hopkins Bloomberg School of Public Health and scientific and strategic advisor to PhaseFolio, for discussion of the methodology, help refining it, and sanity checks. This acknowledgement is not a review of every figure in this card and not an endorsement by Johns Hopkins University.

Questions and corrections: contact@phasefolio.com.