JHTV Public-Data Portfolio Scoring

1. Cohort and provenance

The cohort comprises approximately 602 inventions listed under the Therapeutic Modalities category on the JHU Technology Ventures public technology publisher at https://jhu.technologypublisher.com/. Data were extracted via the Algolia search API on 2026-05-06, capturing only publicly disclosed listing fields: title, description, technology category, development stage (where listed), and JHU reference number. No PII, no internal IP filings, and no non-public data were persisted. Inventions without a Therapeutic Modalities category tag were excluded from the scored cohort.

The extracted listing data represent the public-facing invention disclosure only — not the full patent text, internal research status, or licensing activity. The development stage field is often absent or listed at preclinical; where not specified, the scoring pipeline defaults to Preclinical as the assumed stage for rNPV envelope construction. This is a conservative default: actual stage may be more advanced, which would reduce time-to-market and raise the rNPV envelope.

2. Rubric weights

The composite score for each invention is computed as a weighted sum across three dimensions. Weights are locked at the values below for the methodology@2026-05-07 rubric. Future revisions will be documented in Section 10 with a version bump.

• Clinical relevance (40%) — estimated unmet-need severity and market opportunity in the indicated disease area, relative to current standard of care. Higher scores for conditions with limited existing treatment options, high mortality or morbidity burden, and large addressable patient populations.
• Modality fit (30%) — alignment between the technology modality (small molecule, biologic, gene therapy, cell therapy, etc.) and the indication's historical probability of clinical success. Anchored to cumulative launch-of-approval (LoA) rates from the PoS calibration table, normalized to [0,1] using the actual ceiling of the modality-baseline cumulative-LoA table (see Section 3).
• Competitive whitespace (30%) — inverse of competitive crowding in the indication-modality space, measured as the soft-capped count of competing CT.gov programs and FDA approvals. Fewer competing programs yield a higher whitespace score.

What is not in the rubric, and why. Earlier methodology versions (through methodology@2026-06) included an "IRA exposure" dimension at 20% — a modality-derived proxy for Medicare Drug Price Negotiation risk under the Inflation Reduction Act. As of methodology@2026-05-07 this dimension was removed because it double-counts: IRA cliff risk is already modeled inside the rNPV engine via the Year 9 / Year 13 MFP discount on terminal-value cash flows, so weighting it again in a scouting-stage rubric counts the same effect twice (once in the composite score, once in the dollar number). The dimension also correlated with the modality-fit dimension already in the rubric, and provided ~zero discrimination across the JHTV cohort (every top-10 asset clustered at 0.85–0.90). The 20% weight was redistributed across the three retained dimensions; see Section 10 for changelog and substrate-guarantee details.

3. Subscore methodology

Clinical relevance is assessed using LLM-based enrichment via Claude (Anthropic) with a structured prompt designed to produce a 0–10 numerical score. The full scoring prompt is published in full and covers: disease burden (mortality, morbidity, prevalence), current standard-of-care adequacy, regulatory and reimbursement pathway clarity, and patient population size. The LLM output is normalized to [0,1] for weighting.

Modality fit is computed from the cumulative launch-of-approval (LoA) for the (indication, modality) pair using PhaseFolio's PoS calibration table (BIO/QLS 2021 base rates with modality-specific adjustments per Tufts NEWDIGS 2023 and Parker 2021). As of methodology@2026-06 the raw LoA is normalized to [0, 1] using the actual ceiling of the modality-baseline cumulative-LoA table (CAR-T at ~0.0088) as the divisor — the highest-LoA modality maps to 1.0 and lower-LoA modalities scale linearly down. Pre-@2026-06 the divisor was anchored at 0.15 (the Phase-1-onward LoA, not the preclinical-through-NDA LoA), which compressed all modalities into a 0.03–0.06 band with no discriminative work. See the version notes below.

Competitive whitespace is scored as a soft-capped count of competing programs: the count of active or recently completed CT.gov studies and FDA-approved drugs in the same indication-modality space, pulled from our ingested CT.gov clinical-trial and FDA application data. The raw count is transformed via a soft cap function so that the marginal penalty for the 50th competitor is smaller than for the 5th. Fewer competing programs yield a score closer to 1.0.

Where IRA risk is modeled instead. Inflation Reduction Act terminal-value risk is applied inside the rNPV engine, not in the rubric. The engine discounts post-cliff cash flows starting at Year 9 (for small molecules) or Year 13 (for biologics, gene therapies, and cell therapies) per the modality-specific Maximum Fair Price schedule. This affects the dollar rNPV envelope shown on each deep-dive sub-page directly. See the rNPV engine and IRA-framework methodology pages for the schedule and the substrate-discipline rationale for keeping rubric and engine concerns separated.

4. Top-50 asset-specific deep-dive treatment

The 50 highest-scoring inventions receive asset-specific engine treatment. Each top-10 deep-dive includes (a) one to three named real-world comparator programs anchoring stage costs, durations, and PoS, (b) a per-asset stage profile with each cost/duration/PoS value cited to comparator evidence, (c) an evidence-register table linking every load-bearing assumption to a verifiable public source, (d) a deterministic rNPV envelope (low / base / high) plus a 1,000-trial Monte Carlo distribution, (e) tornado sensitivity ranking the top six drivers of rNPV variance, (f) revenue trajectories of comparable launched programs where applicable, and (g) a thesis narrative plus key-risks list specific to that asset. The deep-dive payload is exposed at /api/v1/case-studies/jhtv-2026/deep-dive.

Comparator selection criteria, in priority order:

• Same target / mechanism class — e.g., POLθ inhibition is anchored to Artios ART558/ART4215 and Repare RP-3467 even when those programs are still preclinical or Phase 1, because mechanism-class identity beats unrelated launched comparators for stage calibration.
• Same indication + modality combination with a launched program — e.g., AAV gene therapy for ultra-rare neurological disease is anchored to Zolgensma economics regardless of the specific target.
• Same regulatory pathway (orphan, fast-track, RMAT, accelerated approval) — used for stage cost / duration calibration when target and indication-modality comparators don't exist.
• Same modality with comparable patient population size — last-resort comparator class.

Each deep-dive sub-page lists the comparator selection criteria its specific picks satisfy. Comparators that are discontinued or have been acquired are retained as cautionary anchors; this is flagged in the comparator selection_criteria text. All comparator citations resolve via curl validation; failed-validation entries are flagged in the evidence register.

Engine input perturbation for the rNPV envelope: low case applies costs ×1.25 and peak revenue ×0.75; high case applies costs ×0.75 and peak revenue ×1.25; base case uses the comparator-anchored values directly. WACC is 12% base case (institutional biotech / VC benchmark) unless an asset-specific WACC is justified in the deep-dive narrative. IRA terminal-value adjustment is applied per the engine's modality-specific cliff schedule.

5. Funding-path archetype classification

Each deep-dive carries a funding-path archetype assigned in Day-1 comparator research and validated against the engine output:

• VC-fundable — base rNPV is positive at industry-median WACC and standalone-equity assumptions. Standard biotech VC structures apply.
• Partnership candidate — base rNPV is negative or near-zero under standalone-equity assumptions but the asset is a real strategic prospect under co-development or pharma-deal economics. The narrative explains what deal structure or platform context would clear the asset.
• Grant / non-commercial — non-traditional revenue model (Gates, GAVI, NIH, product-development partnership). rNPV doesn't apply meaningfully; valuation is a different exercise (cost-effectiveness, public-health impact).

The archetype is shown alongside the rNPV envelope on every deep-dive sub-page so that a negative envelope can be honestly classified as “needs partnership” or “needs grant funding” rather than “not investable.” This is intentional: a sophisticated reader should be able to distinguish a structurally non-VC-fundable asset (e.g., a malaria vaccine) from a VC-fundable one with execution risk.

6. Multiplier handling — non-double-counting

The PhaseFolio methodology defines eight evidence-based PoS multipliers (genetic validation 2.6×, companion diagnostic 2.0×, biomarker enrichment 1.5×, orphan designation 1.5×, first-in-class 0.85×, CAR-T 1.73×, gene therapy 1.41×, and biomarker quality 1.35× — see the PoS Calibration methodology, Table 2). For the JHTV deep-dive, the comparator research pass already calibrates per-stage PoS values against asset-specific comparator programs that themselves carry the relevant evidence (orphan-validated genetic medicines, biomarker-stratified oncology, etc.). Re-applying the formal multiplier stack on top of the comparator-calibrated PoS would double-count that evidence and produce implausibly high cumulative LoA for preclinical assets.

Decision: the engine consumes Day-1 comparator-calibrated PoS values as final. The list of eligible multipliers for each asset is recorded in the deep-dive payload and surfaced on the sub-page for audit, but the values are not re-applied via log-odds stacking. This is a deliberate methodology choice for the JHTV case-study build; the formal multiplier stack remains active in the standard PhaseFolio rNPV engine for scenarios where a user provides raw BIO/QLS baseline PoS as input.

Empirical sanity check: post-deep-dive cumulative LoA across the 50 assets ranges from 0.1% to 10.8% (genuine therapeutic assets cluster ~5–11%; the lowest values are enabling-technology, research-tool, and discovery entries whose conservative illustrative profiles are shown for cohort consistency, not as forecasts), consistent with published cumulative success rates for orphan-validated genetic medicines (~10–15%) and biomarker-stratified oncology (~5–10%). Pre-fix values (with multipliers stacked) ranged from 59% to 73%, which is implausible for any preclinical asset and would have been correctly challenged by any academic-rigor reviewer.

7. Rubric versus rNPV — what each measures

The composite score (locked rubric weights 40/30/30) and the rNPV envelope are correlated but measure different things:

• The composite score ranks each invention by scientific and strategic merit — clinical unmet need, modality-indication fit, and competitive whitespace. It does not depend on market size, cost structure, or commercial-policy risk (the latter is captured in the engine, not the rubric).
• The rNPV envelope measures expected financial value under standard pharma economics — comparator-anchored development costs, indication-specific peak revenue, cumulative PoS, WACC discounting, and IRA terminal-value adjustment per the modality cliff schedule.

An invention can rubric-rank high but rNPV-rank low (e.g., an ultra-rare-disease platform with strong unmet need but small market) and vice versa. Within the top 50 under methodology@2026-05-07, four show positive base rNPV; seven are tagged VC-fundable (the gap reflects assets where the high-input scenario is positive but the base case is not); 26 are honestly classified as partnership candidates and 17 as grant / non-commercial. Many of the ranks 11–50 are enabling technologies, research tools, or basic-science discoveries rather than standalone drug candidates — for these the rNPV envelope is shown for cohort consistency, not as the decision criterion, and the asset is labelled with its true class on its deep-dive page. The 552 lower-ranked inventions are rubric-scored only — full-portfolio engine treatment is a separate engagement, not part of the public artifact.

Cohort selection vs display order. Top-10 membership is locked to the highest composite-score inventions — the rubric, not rNPV, decides who enters the deep-dive cohort. Display order on the index page is then sorted by base rNPV (descending) for investor readability, so the asset with the strongest standalone economics appears first. Composite-score rank is preserved on every deep-dive sub-page header (a footer line reads "composite-score rank #X of 50 deep-dive inventions · page header uses rNPV rank #Y"), so a reader who wants to follow the rubric's strict ordering can do so. Under methodology@2026-05-07, Personalized Allogeneic Neoantigen CAR-T is rank #1 by composite at 0.770 but rank #11 by rNPV (−$16M) — its CAR-T modality-baseline PoS = 1.00 lifts the composite while the platform-breadth assumptions hold the standalone-equity rNPV down. The override condition for dropping an asset from the top 50 is narrow: high-case rNPV below −$50M and no fitting funding-path archetype. In the 2026-Q2 build this condition triggered zero times — every asset has a defensible archetype framing even when the rNPV envelope is fully negative.

8. Limitations

• Public-data-only scope. Each listing is described in approximately 350 characters of public disclosure text. No internal IP filing details, preclinical data packages, or licensing activity are available. Clinical relevance scores based on this text are coarser than scores derived from a full data room.
• Comparator-anchored cost calibration. Top-10 stage profiles are anchored to named real-world comparator programs, but those comparators may differ from the JHU asset on details not captured in public listings (specific patient population definitions, manufacturing scale-up risk, IP overlap). Treat each stage profile as the closest defensible reference, not an asset-specific forecast.
• Preclinical-stage default. When the development stage is not specified in the public listing, the pipeline defaults to Preclinical. If an invention is actually in Phase 1 or later, the envelope understates the rNPV by embedding additional preclinical cost and time that has already been incurred.
• LLM classification noise. Clinical relevance subscores are generated by an LLM and subject to classification noise — particularly for novel modalities or indications with limited training data representation. The structured prompt reduces but does not eliminate variability across runs.
• IRA risk lives in the engine, not the rubric. The composite score deliberately does not weight Inflation Reduction Act exposure (see Section 2). The IRA effect is captured in the rNPV envelope through a Year 9 / Year 13 MFP discount on terminal-value cash flows. The cliff schedule is binary by modality class and does not model partial-lifecycle exposure, indication-specific negotiation probability, or potential exemptions (orphan, pediatric). Treat the engine's IRA adjustment as a proxy, not a legal or regulatory determination.
• Top-50 only. Asset-specific deep-dive treatment is applied to the 50 highest-composite-score inventions only. The remaining 552 are rubric-scored without engine output. Full-portfolio engine treatment is a separate engagement and out of scope for this public artifact.

9. Bucketing

Inventions are sorted by composite score and assigned to one of three tiers:

• Top 50 (highest composite scores): asset-specific deep-dive treatment with rNPV envelopes, per-dimension score breakdown, and indication-specific commentary. These inventions represent the strongest signal from the public data rubric.
• Mid tier(inventions ranked 51–110, ~60): secondary reference table with composite score, top two dimensions driving the score, and indication summary. No rNPV envelopes.
• Tail (inventions ranked 111–602, approximately 492 inventions): summarized in aggregate only — count by modality category, score distribution quartiles. No individual-invention detail.

Total cohort: ~602 inventions. Tier counts: Top 50 (asset-specific deep-dive) / Mid ~60 (ranks 51–110, rubric-only reference) / Tail ~492. All counts are approximate — the Algolia extract may include or exclude listings on the boundary of the Therapeutic Modalities category depending on tagging at time of extraction.

10. Methodology version notes

2026-05-19 — deep-dive coverage expansion (no methodology version bump). The asset-specific deep-dive set was expanded from the top 10 to the top 50 of the same pinned ranking. No rubric weight, equation, subscore method, cited source, or dataset changed, and the ranking was not re-derived — ranks 11–50 were read straight from the existing jhtv-portfolio@2026-Q2 ordering and worked up with the identical comparator-anchored method under rubric@2026-Q2-v3 / methodology@2026-05-07. Because the method is unchanged this is a coverage/transparency change with zero effect on any published number, so the methodology version is deliberately not bumped (same discipline as the 2026-05-16 version-history projection). The rNPV engine provenance was honestly re-stamped 1.0.0 → 2.1.0: the artifact is now produced by the current engine, and a demo-invariance check confirmed every original top-10 rNPV / eNPV / cumulative-PoS / Monte-Carlo / tornado value reproduces bit-identically (Δ$0.00) under 2.1.0 — already-issued exports remain verifiable. Ranks 11–50 skew toward enabling technologies, research tools, and basic-science discoveries; each carries an explicit asset-class and funding-path label and a per-asset CMO defensibility audit (every hard finding resolved in the asset's own thesis / key-risks / archetype before publication).

methodology@2026-05-07 (2026-05-07): IRA exposure was removed from the rubric. Pre-bump, IRA contributed 20% of the composite via a modality-derived risk score; post-bump, it contributes 0% in the rubric. The rationale is anti-double-counting: IRA cliff risk is already modeled inside the rNPV engine via the Year 9 / Year 13 MFP discount on terminal-value cash flows (per the IRA-framework methodology page), so weighting it again in a scouting-stage rubric counts the same effect twice — once in the composite score, once in the dollar number. The dimension also correlated with modality fit (which has its own 30% weight) and provided ~zero discrimination in the JHTV cohort — every top-10 asset clustered at 0.85–0.90. The 20% weight was redistributed across the three retained dimensions: Clinical relevance 30% → 40%, Modality fit 25% → 30%, Competitive whitespace 25% → 30%. The rubric version bumps from rubric@2026-Q2-v2 to rubric@2026-Q2-v3; jhtv-portfolio@2026-Q2 dataset version is unchanged. The top-10 reranks under the new weights — three assets exit (POLθ, mRNA Intensifier, malaria VIP) and three enter (Velocity Receptor CAR-T, retinal Muller-glia reprogramming, polymer-NA CFTR base-editing) — but the new top 10 stays within the previously-researched fixture inventory, so every deep-dive page renders as before with its full thesis prose, comparators, evidence register, Monte Carlo, and tornado.

methodology@2026-06 (2026-05-06): The modality fit dimension's normalization denominator was corrected from 0.15 (Phase-1-onward LoA) to the actual ceiling of the modality-baseline cumulative-LoA table (~0.0088, set by CAR-T per Tufts NEWDIGS 2023 1.73× per-stage multipliers on Phases I/II). Under the prior denominator the dimension produced 0.03–0.06 across all modalities, providing no discriminative work; under the corrected denominator modalities span [0.27, 1.0].

Format note. This monthly-form label (@2026-06) was the version string in use on its release date of 2026-05-06 — the calendar month in the label reflects the labeling scheme of the day, not the date of the ship. The version label was later promoted from monthly granularity (@YYYY-MM) to daily granularity (@YYYY-MM-DD) per our changelog format policy ("promote to daily only if multiple ships in one month introduce confusion"). Three methodology ships in May 2026 — @2026-05 additive, @2026-06, and methodology@2026-05-07 — had reached that bar. Going forward, all bumps use the daily form. Old monthly-form labels (@2026-04, @2026-05, @2026-06) remain valid forever as the strings stamped on already-issued signed exports; the public /verify endpoint accepts all forms indefinitely.

Composite scores under methodology@2026-05 remain valid as historical artifacts and continue to verify via the public /verify endpoint — substrate guarantee preserved per our versioning policy. The jhtv-portfolio@2026-Q2 dataset version is unchanged. The rubric version bumps from rubric@2026-Q2-v1 to rubric@2026-Q2-v2.

methodology@2026-05 (initial release): Public-data scoring of all ~602 JHU Tech Ventures Therapeutic Modalities inventions, four-dimensional rubric with locked weights 30/25/25/20, top-10 deep-dive treatment with comparator-anchored stage profiles. Deep-dive sub-pages additionally received a display-layer fix (no engine-math change) to the Monte Carlo "success tail".