AI Drug Discovery — Evidence-led Decision Brief

01 / Executive decision

A narrow entry point, with hard gates.

The evidence supports tools that improve specified scientific decisions. It does not support a general claim that AI increases end-to-end clinical success.

Recommendation · FINAL-005

Build a governed evidence-and-diagnostics business through paid, reversible experiments. Do not begin with a general AI discovery platform, unrestricted genomic-data licensing, a self-funded companion diagnostic, or a broad therapeutic pipeline.

Inference · Moderate

AI’s strongest gains are upstream.

Search, measurement, structure prediction, and design improved more convincingly than causal biology or clinical outcomes. An auditable evidence layer is more defensible than another general model. S0003 S0006

Fact / inference · High

Demand exists—and so do incumbents.

Biopharma already buys diagnostics, data, analytics, CDx, and trial services. The sponsor must prove a measurable advantage against scaled providers, not assume one. S0068 S0121

Inference · High

Clinical uplift is not identified.

Small, unmatched, selectively disclosed cohorts cannot isolate an AI effect on cycle time or approval probability. Phase I signals have not translated into a demonstrated Phase II advantage. S0127

Decision	Posture now	Release condition
Gate 0: sponsor reality	Resolve now	Domicile, segment, permitted use, payer mix, liquidity
Biomarker / target-evidence pilots	Paid test	Named decision, independent cohort, locked analysis plan
Governed cohort analytics	Selective test	Scarce asset, controlled query, reproducible output
Trial enrollment	Partner	Contract against verified funnel outcomes
Companion diagnostic	Preserve option	Segment fit, reimbursement map, partner-funded plan
General platform / broad therapeutic JV	Do not lead	Separate thesis and board approval only

02 / Evidence & validation

Progress is real. Attribution remains hard.

Open each technical panel to inspect what is demonstrated, where the denominator breaks, and what management may safely infer.

Historical developments2020–2026: infrastructure before outcomes

2020–21Fact / inference

Structure at scale

AlphaFold 2 made high-quality structural hypotheses broadly reusable; confidence still does not establish binding, dynamics, function, or druggability. S0003 S0006

2022–24Fact / inference

Design crosses wet-lab thresholds

Generative chemistry and protein design produced prospectively synthesized molecules and experimental binders, but selected campaigns do not disclose the full denominator or a clinical advantage. S0011 S0015

2025–26Fact / inference

Reality moves to prospective decisions

Complex perturbation models can fail to beat simple baselines, while a 20,707-patient randomized notification study did not improve therapeutic enrollment. S0010 S0022

Decision rule

Name the decision, comparator, denominator, elapsed time, direct cost, downstream outcome, and failure rule before claiming productivity.

Clinical validationRentosertib, clinical cohorts, and genetic support

Fact / inference · Moderate

Rentosertib: a traceable chain, not proof of efficacy

The 71-person, 12-week Phase IIa trial at 21 China sites was primarily a safety study. There was no formal between-group efficacy comparison or p-value; 22 week-12 FVC values were imputed, and response was non-monotonic. S0013

30 mg QD−27.0 mL

30 mg BID+19.7 mL

60 mg QD+98.4 mL

Placebo−20.3 mL

Inference · High

Phase I ≠ platform validation

Reported 80–90% Phase I success among AI-associated molecules falls to about 40% in Phase II from a sample of roughly ten programs. The cohort is small, immature, unmatched, and exposed to selection and censoring. S0032

Illustrative orientation

Genetic support improves the prior

2.6 × 7.9% ≈ 20.5%

Even this cross-dataset orientation implies roughly four in five programs do not reach approval. It is not a causal sponsor forecast. S0033 S0112

Commercial evidenceReal demand, crowded categories

Tempus

Diagnostics, multimodal data, analytics, and trial matching; filings describe multi-year data contracts and a pharma customer base. S0068

Guardant

Oncology testing, biopharma services, CDx, and real-world data; reported $210.1m FY2025 biopharma-and-data revenue. S0121

Foundation / Roche

Research assays, CDx, data, enrollment, and monitoring backed by a large regulated installed base. S0122

IQVIA & CROs

Established pharma procurement routes for real-world data, cohort design, recruitment, and operations. S0123

A moat is not a data asset. It is a buyer-verified advantage in rights, cohort scarcity, assay quality, recontact, longitudinal outcomes, ancestry, cost, speed, or regulated delivery.

03 / U.S. & China

Domicile and segment are Gate 0.

Regional data planes reduce exposure; they do not create a legal exemption. Move the approved question or permitted output—not an unrestricted genomic lake.

United States

Fact / inference · jurisdiction-specific

Commercial entry requires six separate proofs.

Intended useResearch service, LDT/IVD, CDx, or another device route
CodeCPT/HCPCS/PLA, miscellaneous, or a new-code path
CoverageNational, MolDX, Medicare Advantage, and commercial policies
PaymentCLFS/contractor rate, gapfill, crosswalk, ADLT, and billing flow
EvidenceAnalytical validity, clinical validity, utility, and comparator
EconomicsPayer mix, denials, collection lag, cost, and subsidy

MolDX coverage review and CLFS payment mechanics are separate; neither guarantees viable margin. S0125 S0126

China / cross-border

Fact / recommendation · legal review required

Ownership can change the addressable market.

Non-PRC domiciled

China is an expansion and counterparty decision; regional custody, purpose limitation, and output review still apply. S0085

PRC domiciled or controlled

China-local HGR and data compliance remain necessary. U.S. procurement, federal funding, diligence, and independence requirements may narrow the buyer set. S0088 S0096 S0099

Management input required

Resolve legal domicile, ultimate ownership, controlling shareholders, data location, contracting entity, and intended customer jurisdiction before customer targeting or architecture design.

How testing segment changes the strategySomatic, germline, reproductive, and consumer paths diverge

01

Somatic / liquid biopsy / CGP

Biomarkers, CDx, therapy selection, enrollment, and longitudinal response rise.

02

Germline / rare disease

Target validation, natural history, family genetics, and recontact rise.

03

Reproductive / prenatal

Most near-term drug offers fall; clinical-purpose and consent boundaries can be substantial.

04

Consumer genomics

Recruitment may be possible with strong consent and trust; regulated CDx and raw-data licensing fall.

04 / Scenarios 2027–2035

Demand can rise while the core weakens.

Scenario weights are intentionally omitted until sponsor data exist. No sector headline unlocks capital; each option advances on its own gates.

Scenario · unweighted, conditional

High evidence demandWeak evidence demand

A

Selective productivity

Moderate/rising demand · durable core

Scale paid evidence; preserve one partner-funded CDx option.

B

Prospective proof

High demand · durable core

Add networks and regulated programs only after return gates.

D

Core squeeze

High demand · impaired core

Protect liquidity; partner or license rather than build.

C / E

Translational reset

Weak demand · durable or impaired core

Fee-covered work only; freeze speculative capital.

Durable core diagnosticsImpaired core diagnostics

05 / Opportunity ranking

Value and ease are different decisions.

Scores are transparent analyst judgments, not valuations. Strategic value and execution ease each use four 1–5 components.

Lead paid pilot

17Value /20

15Ease /20

Biomarker / stratification service

Strong adjacency and a visible buyer path. Crowding makes win/loss evidence mandatory.

Partner-funded option

18Value /20

9Ease /20

Companion diagnostic

Highest durable value for the right somatic asset base; heavy reimbursement, regulatory, and timing burden.

Selective paid pilot

16Value /20

14Ease /20

Governed cohort analytics

Demand exists; defensibility depends on scarce outcomes and rights, not access alone.

Opportunity	Value	Ease	Posture
Genetic target / indication validation	16/20	15/20	Paid pilot
Multi-omic / longitudinal evidence	15/20	11/20	Partner-funded build
Biobank expansion	15/20	8/20	Option only
Independent model validation	14/20	19/20	Bounded demand test
Molecular prescreening / enrollment	14/20	15/20	Partner against baseline
Structured licensing	14/20	15/20	After evidence exists
Broad raw-data licensing	8/20	12/20	Do not lead
Broad therapeutic JV	10/20	5/20	Do not lead

06 / Capital allocation

Capital follows proof—not possibility.

All figures are estimates in 2026 USD. They exclude core lab overhead, tax, financing, acquisitions, owned-drug IND work, manufacturing, and clinical trials.

90 days$0.5–1.2m

One rights-cleared cohort and credible buyer path

12 months$3–7m

Two paying customers, repeat demand, prospective pass

Year 3$12–30m

Six programs, three customers, two replications

Year 5$25–70m

Repeatable revenue, diversified customers, ring-fenced downside

The seven-gate capital sequenceFrom sponsor reality to asset-level risk

0

Sponsor reality

Domicile, segment, rights, cohort power, payer mix, liquidity

1

Paid design

Buyer, protocol, cash, IP/data, stop terms

2

Replication

Simple comparator, subgroup analysis, denominator

3

Utility

External validation tied to a decision

4

Repeatability

Repeat buyer, margin, reuse, revenue plan

5

Regulated scale

Agency, reimbursement, quality, security, custody

6

Asset risk

Biology, FTO, CMC, partner funding, capped downside

07 / Risks & controls

Stop rules are part of the strategy.

Stop expansion if any two operating failures persist for two quarterly reviews—or once for a critical rights or security event.

Rights

Consent, provenance, recontact, or derived use fails

Maintain a dataset-level rights registry and purpose-specific review.

Stop when lead use is absent, ambiguous, or breaks participant promises.

Core

Adjacent growth impairs diagnostics liquidity

Separate the liquidity model and ring-fence capital.

Stop when core cash falls below the board floor.

Science

Performance does not transfer

Use an external cohort, simple baseline, locked version, and subgroup calibration.

Stop when pre-specified replication or utility fails.

Market

Incumbents commoditize the offer

Track win/loss evidence against named alternatives.

Stop when two pilots reveal no measurable reason to repurchase.

Trust

Privacy, cyber, HGR, or national-security exposure

Regional custody, least privilege, output review, and a current incident plan.

Stop on a material breach, prohibited transaction, or unmanageable counterparty.

08 / First 90 days

Buy information, not platform code.

The first tranche should eliminate the uncertainties that can change the strategy’s shape.

01
Resolve domicile, ownership, testing segment, customer jurisdictions, and board liquidity floor.
02
Map dataset consent, provenance, recontact, withdrawal, derivative, publication, retention, and regional access.
03
Quantify cohort power, ancestry, specimen quality, outcomes, missingness, and external validation.
04
Map code, coverage, payment, denial, collection, and margin for core and proposed offers.
05
Run 8–12 buyer interviews, secure 3–5 written scopes, and obtain at least two paid commitments.
06
Build bottom-up offer economics; lock protocols, baselines, negative-result policy, and stop criteria.

09 / Sources

Evidence you can inspect.

Source IDs match the finalized Codex source ledger. Links open the original publisher, regulator, filing, or paper.

S0003Highly accurate protein structure prediction with AlphaFoldNature S0006Experimental comparison of AlphaFold predictionsNature Methods S0010Perturbation-model benchmark evidenceNature Methods S0011Generative design with prospective synthesisNature Machine Intelligence S0013Rentosertib Phase IIa studyNature Medicine S0015RFdiffusion protein designNature S0022Randomized trial-notification studyJAMA Network Open S0023FDA authorized companion diagnostics listU.S. FDA S0024AI supporting drug regulatory decisionsU.S. FDA S0032AI-discovered drugs in clinical trialsDrug Discovery Today S0033Clinical development success rates, 2011–2020BIO S0068Tempus annual filingSEC S0085Data Security ProgramU.S. Department of Justice S0088FY2026 NDAA, Section 851U.S. Congress S0096Human Genetic Resources implementing rulesChina MOST S0099Cross-border data-flow provisionsState Council Gazette S0112Genetic support and drug-development outcomesNature S0118UK Biobank Pharma Proteomics ProjectUK Biobank S0120About FinnGenFinnGen S0121Guardant Health annual filingSEC S0122Foundation Medicine biopharma solutionsFoundation Medicine S0123Patient recruitment solutionsIQVIA S0124Myriad Genetics annual filingSEC S0125MolDX molecular diagnostic coverage policyCMS S0126Clinical Laboratory Fee Schedule 2026 updateCMS S0127Clinical-success cohort sensitivityNature Communications S0128AI-enabled medical device inventoryU.S. FDA

10 / Methodology

Facts, inferences, estimates, and recommendations stay separate.

The analysis prefers primary and authoritative sources, attributes company claims, retains negative evidence and denominators, and does not treat transaction headline values as realized revenue.

Fact / inference Inference Estimate Scenario Recommendation

Material limitations: public disclosure is selective; private prices and failed programs are underreported; the clinical set is not a census; living regulatory lists can change; China legal conclusions require qualified PRC counsel; and sponsor-specific rights, payer mix, cash generation, and internal evidence were unavailable.

Corporate strategy research—not legal, regulatory, valuation, tax, investment, or medical advice.