21-aducanumab-contested-retrospective

The first three pre-staged assets (/05 adagrasib, /19 lifileucel, /18 divarasib, /20 tulisokibart) all share one feature: the framework's output and the actual outcome are reconcilable in principle. The math composes; the call is defensible.

Aducanumab is different. It is the asset where the framework was honest and wrong — pre-approval bare-chain PoS ~54% <!-- parity-allow: superseded --> (now ~14% under the v1.7.7 hard-cap, see below), FDA approved Aducanumab in June 2021 over a 10-0 Advisory Committee vote against, and voluntary withdrawn by Biogen in January 2024. This writeup walks through what the framework would have said in March 2019 (after the ENGAGE/EMERGE futility analyses) and uses the framework's miss to delineate where the framework is silent and why that silence is acceptable.

Why this asset

The framework's design commitment is that it should never invent a PoS for things it doesn't know how to price. Aducanumab is the cleanest test of that commitment: at the March 2019 information cutoff, the science was contested (Phase 3 futility on the original analysis, positive post-hoc reanalysis of the high-dose EMERGE arm), and the actual approval ran through a political process the framework does not model (FDA Center Director Janet Woodcock's accelerated approval over the Advisory Committee's 10-0 reject vote, the surrogate amyloid endpoint reinterpretation, the post-approval CMS coverage restriction).

The asset is interesting because:

1. The framework gets the science approximately right. Anti-amyloid in AD has a deep cohort base rate of ~5-8% at Phase 3 (CNS · biologic cohort). With target_validated: false (the amyloid hypothesis was contested in 2019) and biomarker_enrichment: false for the original Phase 3 designs, the bare multiplier chain would produce PoS ~50-55% <!-- parity-allow: superseded --> conditional on a positive read — and the read was not clearly positive at the cutoff. The v1.7.7 hard-cap pulls the framework's output down to ~14%, which is the honest population-anchored read for an asset where neither validation lever is on.
2. The framework is silent on political process. No multiplier in the multiplier chain corresponds to "FDA Center Director will accelerate-approve over a unanimous AdCom reject." That outcome path is unmodeled, and should be unmodeled. A framework that tried to price political risk would be untrustworthy.
3. The framework's downside signal is the downside_failed_p3 line, not the success-conditional base case. This is a subtle read but it matters. With the inputs documented below, post-cap rNPV base case is $1,983M (what the asset is worth if Phase 3 succeeds — the framework can't model the political accelerated- approval-and-withdrawal arc, so a successful Phase 3 is the only path to commercialization it sees). The base case is large because the bullish peak ($4.5B) and well-capitalized launch economics compound through the PoS-gated cash flows. The framework's actual downside line is downside_failed_p3 = $106M — what's left if ENGAGE/EMERGE failure terminates the program. The realized outcome (peak revenue ~$5M before 2024 withdrawal) sat below even this downside line, because the framework prices a clean failure (program terminates) but reality was a contested approval that limped along consuming SG&A before being killed. The framework honestly under-priced the political-process tail risk on the downside — see "Where the framework is silent" below.

Information cutoff: March 2019, immediately post-ENGAGE/EMERGE futility analyses. Pre-Biogen-restart-announcement (October 2019), pre-Center-Director-approval (June 2021), pre-withdrawal (January 2024).

The asset

rNPV inputs

These mirror the workbench staging at web/app/diligence/[asset]/page.tsx (constants ADUCANUMAB_RNPV) so the methodology numbers below reconstruct exactly when you visit /diligence/aducanumab.

Framework outputs (what March 2019 inputs produce)

• PoS final LOA: ~14% (post-v1.7.7 hard-cap) — down from the pre-cap ~54% <!-- parity-allow: superseded --> that the bare multiplier chain produces. The CNS · biologic cohort base rate is ~5-8%; with target_validated: false the modality multiplier should drag PoS down hard, but the bare chain (Fast Track + well-capitalized reflexivity + mAb modality default) pushed back up to ~54% <!-- parity-allow: superseded -->. This was the framework's first honest signal: when target_validated is false AND biomarker_enrichment is false, no other lever should be able to push PoS above cohort-base-rate × 3. Shipped in v1.7.7: a post-multiplier-chain hard-cap implemented in api/app/modules/pos/engine.py. The cap anchors on the long-run P1→approval cohort LOA (modality-adjusted) rather than the phase-conditional base rate, because the phase-conditional rate already absorbs the survival bias of having reached the current phase — exactly what the cap is intended to push back against. For CNS · mAb that's 0.041 × 1.15 × 3 ≈ 0.14, which is where aducanumab's PoS now lands. The cap is recorded as an explicit row in the audit-trail waterfall so the final number remains reconstructable.
• rNPV base case: $1,983M (post-cap; success-conditional, not risk-adjusted further) <!-- parity-allow: model-output -->. Computed live by api/app/modules/rnpv/engine.py against the input table above. Pre-cap rNPV (bare multiplier chain at ~54% PoS) would have been ~$4,300M <!-- parity-allow: model-output --> — the v1.7.7 cap pulls rNPV down by ~$2.3B. Monte Carlo P25/P50/P75 = $1,109M / $1,829M / $2,850M <!-- parity-allow: model-output -->. These numbers are what they are: at a $4.5B bullish peak with a 14% LOA, the success branch is still worth ~$2B in expected present value.
• The honest downside signal is downside_failed_p3, not the base case. Reading rNPV as the "downside bound" mistakes the framework's outputs. The base case is success-conditional — what the asset is worth if Phase 3 succeeds and the asset reaches market. The framework's actual downside line is the downside_failed_p3 field, which prices the Phase 3 failure path separately. For aducanumab with the inputs above: downside_failed_p3 = $106M <!-- parity-allow: model-output -->. That's the framework's honest read on "what's this asset worth if the Phase 3 trials fail" — small, positive, reflecting residual IP value. The actual outcome (peak revenue ~$5M before withdrawal) was even worse than this downside line predicted, because the framework prices a clean failure (program terminates) rather than a contested approval-and-withdrawal (program limps along consuming SG&A before being killed). That's a framework limitation, not a framework failure — see "Where the framework is silent" below.
• Comparables cohort: CNS · biologic (currently falls back — insufficient deal density at the precise CNS · biologic cohort in the comparable database; this is a known limitation called out in the Limitations Panel)
• Implied value: bounded by the rNPV — the comparables fallback denies an EV/peak-sales upside multiplier, which is appropriate for an asset where the science is contested

What actually happened (the political path the framework cannot see)

• October 2019: Biogen announces post-hoc reanalysis of the EMERGE high-dose arm with a 22% reduction in clinical decline. <!-- parity-allow: external-stat --> The ENGAGE arm did not show the same effect. The reanalysis is controversial — the prespecified analysis was futile; the new analysis re-cut the data.
• November 2020: FDA AdCom votes 10-0 against approval (with 1 uncertain). The AdCom reads the post-hoc analysis as insufficient.
• June 2021: FDA Center Director Janet Woodcock invokes accelerated approval based on the surrogate amyloid PET endpoint, over the AdCom's reject vote. Three AdCom members resign in protest.
• April 2022: CMS issues a National Coverage Determination restricting Aducanumab to CMS-approved clinical trials, effectively blocking commercial coverage.
• January 2024: Biogen voluntarily withdraws aducanumab from the market. Peak revenue: ~$5M (vs. the $3-5B sell-side bull case).

What the framework correctly does not model

1. Accelerated approval on contested surrogate endpoints. The framework's modality and TA multipliers do not contain a "FDA will approve on surrogate endpoint despite AdCom reject" branch. This is intentional. A framework that priced political-process risk would be illegitimate for a tool meant to encode scientific PoS.
2. Post-hoc reanalysis as evidence. The framework does not reweight PoS based on subgroup analyses generated after primary endpoint failure. Doing so would invite reflexive reasoning.
3. Reimbursement battles as PoS proxies. The CMS coverage determination was the actual commercial death blow. The framework prices a clean Phase 3 failure (downside_failed_p3 = $106M <!-- parity-allow: model-output --> for these inputs) but does not model the CMS-approval-vs-FDA-approval gap — the path where the FDA approves and CMS effectively denies reimbursement, leaving the asset technically on-market but commercially dead. This is the contested-political-process tail risk on the downside, and it's a known framework limitation documented in the LimitationsPanel.

What the framework should do better

1. Tighter target_validated × cohort-base-rate coupling. SHIPPED in v1.7.7. When target_validated: false AND biomarker_enrichment: false, the framework now caps PoS at the long-run P1→approval cohort LOA (modality-adjusted) × 3. For CNS · mAb the cap is ~14%, replacing the previous bare-chain ~54% <!-- parity-allow: superseded -->. Implementation: a post-multiplier-chain ceiling applied at the end of _compute_for_asset in api/app/modules/pos/engine.py, emitted as an explicit unvalidated target + no biomarker enrichment cap row in the audit-trail waterfall so the final number remains reconstructable. The cap is anchored on the cohort cumulative LOA (not the phase-conditional base_rate) because the phase-conditional rate already absorbs the survival bias of having reached the current phase — exactly what the cap is intended to push back against.
2. CNS · biologic comparables cohort. Currently the framework falls back when this cohort is queried — there's not enough deal density. Adding aducanumab itself as a comparable (peak ~$5M, withdrawn) would be honest but tiny. The honest fix: keep the fallback but tag the implied-value cell with "no cohort — estimate not reliable."

Cross-references

• The framework-limits discussion in this writeup is the empirical anchor for the existing LimitationsPanel content. Edits to the panel should cross-reference this writeup so the limitation + example travel together.
• /05 adagrasib and /19 lifileucel both retrospectively match their outcomes within tolerance. Aducanumab is the counterpoint that demonstrates the framework's honest silence on a class of risks it cannot price.

Sources

• Biogen 10-K FY2018
• ENGAGE (NCT02484547)
• EMERGE (NCT02484547 paired arm)
• Sevigny J et al. The antibody aducanumab reduces Aβ plaques in Alzheimer's disease. Nature 537, 50-56 (2016)
• FDA accelerated approval of Aduhelm, June 7 2021
• CMS National Coverage Determination, April 2022
• Biogen withdrawal announcement, January 31 2024