The Next Fifteen Years

A forecast built from first principles
Section future / 03-domains / physical / medicine / drug-discovery.md

Drug discovery - preclinical speed, clinical wall#


Contents

Real acceleration in target identification, structure prediction, and molecular design. That sentence is true and often followed by a fantasy about timelines collapsing to years. The fantasy skips where drugs actually fail.

Where the failures are#

~60% of clinical failures are efficacy or toxicity in humans, which AI cannot predict without better biology - biology that still has to be measured in organisms, not only in models of organisms.

In silico progress moves the early funnel: more candidates, cheaper prioritization, better structures. It does not move the clock on Phase II/III, manufacturing scale-up, or regulators reading multi-year safety data.

~20–30% cost reduction in preclinical by the early 2030s, and minimal change to the ~10-year, ~$2B clinical gauntlet before ~2032.

After 2032 the clinical side can move if automated experimentation and better translational models land - that is a science bottleneck, not a pharma-marketing bottleneck.

The rate limiter is wet lab and recruitment#

StageWhat AI helpsWhat still binds
Target ID / designA lotBiological validity of the target
Preclinical in vitro / in vivoThroughput, design of experimentsAnimal models, assay development
Phase I–IIIProtocol design, site selection, document generationPatient recruitment, endpoints, safety clocks
Approval / manufacturingDossier draftingProcess validation, inspection

Whoever industrializes automated experimentation captures the largest available prize - same conclusion as Science, for the same reason. Hypothesis generation is no longer scarce; experimental cycle time is.

Why cheaper candidates may not mean more approvals#

There is a counterintuitive consequence of making the early funnel cheap that the optimistic case usually skips. If candidate generation costs collapse while clinical capacity stays fixed, the constraint moves from having a good molecule to choosing which of many to spend a decade and a billion dollars testing. The industry's selection process is already the weakest link in the chain, as evidenced by the failure rate quoted above, and nothing about generating ten times more candidates improves it. Abundance upstream of an unchanged filter raises throughput at the filter's input, not its output.

The optimistic reply is that better selection is exactly what models should be good at, and in principle it is. The catch is that selection quality is validated only by the clinical outcome, so improving the filter requires the very feedback loop that runs on a ten-year clock. The one part of the pipeline where better prediction would pay most is the part where ground truth is most expensive, which is the master asymmetry stating the case against its own most attractive application.

A second-order effect worth watching: if candidate abundance is real, the scarce asset shifts to clinical trial infrastructure - patient access, site networks, regulatory relationships, and the capital to run parallel programs. Those are the inelastic complements in pharma, they are held by incumbents, and the framework predicts value accrues there rather than to the design tools. That is consistent with what the deal structures have looked like: discovery platforms license to large pharma rather than displacing it.

What "AI drug" headlines measure#

They measure entry into the pipeline or preclinical milestones, not reduction in expected time-to-approval for a new molecular entity. Those are different random variables. A flood of AI-designed candidates can coexist with an unchanged median approval timeline if the clinical wall does not move.

Watch for:

Connection to the rest of medicine#

Faster preclinical discovery does not fix delivery. A cheaper candidate still meets reimbursement, licensure, and care-capacity constraints on the way to patients. The consumer-side diagnostic boom and the drug-discovery boom can both be real while population health moves slowly - because health is produced by the delivery system, not by the molecule inventory.

Failure modes#

Dual-use is not a sidebar#

The same stack that shortens design cycles for legitimate candidates shortens them for biosecurity threat agents. The binding public-interest question is not whether AI "helps drug discovery" - it does, where verification is cheap - but whether the marginal capability is more valuable in the clinic or more dangerous in the wrong hands, given that screening and synthesis barriers sit elsewhere in the chain. A page that celebrates preclinical throughput without naming that joint is incomplete; a page that freezes discovery over dual-use without pricing the disease burden is incomplete the other way. The corpus's answer is layered barriers and state capacity, not a discovery freeze - see the biosecurity chain table.

Preclinical is not health#

A decade of cheaper candidates with flat Phase II success and flat population outcomes is the base case, not a paradox. Score this page on validated experiments and automated-lab utilization, not press releases of "AI-designed" molecules that still die in humans. Health is produced in delivery; this page only changes the molecule inventory rate.


Related: Science · Data · Biosecurity · 2032–2040

View markdown source

select · Enter open · Esc close