NextFin News - Artificial intelligence is compressing early-stage drug discovery from years into months, but the bottleneck that destroys nine out of ten drug candidates has not moved an inch. That is the sobering message from Ruth McKernan, the veteran biotech executive and venture investor, who argues the technology is real and already useful — but far narrower than the hype suggests. Her point reframes the entire AI-drug-discovery debate: the question is no longer whether machines can design molecules, but who actually captures the value when the hard work of clinical development remains stubbornly human, slow, and expensive.
McKernan, a venture partner at SV Health Investors, former chief executive of the UK government's Innovate agency and executive chair of neuropsychiatry biotech Draig Therapeutics, has spent decades inside the industry's R&D engine — from Pfizer's research labs to the boardrooms of companies racing to turn biology into medicine. In a recent interview, she offered a verdict that cuts against both the cheerleaders and the doubters: AI in biotech is "real, but narrow," and its most concrete wins so far sit not in molecule generation but in understanding patients well enough to run smarter trials.
The Productivity Miracle Is Real — and It Stops at the Lab Door
The evidence for acceleration in early discovery is no longer anecdotal. Industry analysis compiled in early 2026 shows AI-enabled workflows compressing early discovery timelines by 30 to 40 percent and cutting preclinical candidate development to 13 to 18 months, versus the traditional three to four years. In antibody design, reported hit rates have reached 16 to 20 percent against computational benchmarks near 0.1 percent — a two-orders-of-magnitude improvement in target-to-candidate efficiency. Protein structure prediction models now deliver 50 percent-plus accuracy gains over traditional methods, and generative design cycles that once consumed supercomputing resources can now produce tens of thousands of compliant candidates in hours.
Yet this productivity miracle has a hard boundary. Clinical trial duration, regulatory review timelines, and manufacturing scale-up remain unchanged. Biology, patient enrollment, and regulatory requirements impose non-negotiable constraints that no algorithm can bypass. Claims of "10x faster drug development" conflate preclinical acceleration with total development time — a category error that has cost the sector credibility. The industry's persistent failure rate, roughly 90 percent of candidates never reaching approval, is a clinical-phase phenomenon, and that is precisely where AI has so far delivered the least.
The clinical data supports a measured read. A widely cited 2024 analysis in Drug Discovery Today found AI-discovered molecules achieving an 80 to 90 percent success rate in Phase I trials, compared with historic industry averages of 40 to 65 percent — an encouraging start. But Phase II success for the same molecules was approximately 40 percent, in line with historical norms. The sample remains tiny: of 6,147 drugs in clinical development, only 67 — about 1 percent — were AI-discovered molecules, and only 24 involved AI-discovered targets, narrowing to between three and nine for genuinely novel AI-originated targets. A technology that touches barely one percent of the pipeline cannot yet be said to have transformed medicine.
There's a lot of hype about AI, but I'll give you a few specific examples of where we've used it and it's it's been fantastic. So one of the areas that's really emerging is speech analytics and being able to break down speech into different features. So for example, people who suffer from depression are very slow to respond to a question. If you use speech analytics while doing that, you can now be giving begin to get an understanding of is this person truly depressed? How depressed are they? What sort of depression do they have. This is that sort of analytics is really making a difference in terms of understanding patients and their needs.
That quote — McKernan describing how her teams actually use the technology — is the clearest statement of where AI is earning its keep today: not as an autonomous inventor of drugs, but as a tool for patient stratification, the single biggest determinant of whether a trial succeeds or fails. In neuropsychiatry, where heterogeneous patient populations and noisy endpoints have sunk countless programs, the ability to measure depression severity objectively through speech patterns is not incremental. It is the difference between a trial that reads a signal and one that drowns in noise.
The Investment Cycle Is Cyclical; the Technology Is Structural
Separating the cyclical from the structural is the only way to think clearly about this sector, and the two forces are currently pointing in opposite directions. The investment cycle is unmistakably cyclical and mean-reverting: the 2021-22 boom in which AI drug-discovery platforms raised billions on "biobucks" — milestone payments that may never materialize — has given way to a brutal correction. The ratio between announced biobucks and actual upfront cash sits near 50 to 1, a valuation discipline that has already forced multiple companies to shut down, cut workforces by more than 20 percent, or seek delisting. Venture capital has consolidated around the best-funded players, and the easy-money era for pure-platform stories is over. That leg will revert; capital will return when clinical proof arrives, and it will flee again when it does not.
The technology, by contrast, is structural. A regime change has occurred in how early discovery is performed, and it will not reverse. The capability shift is permanent — foundation models for biology, physics-enabled generative design, and multi-omics datasets do not un-invent themselves. The economics are durably altered: a 30 to 40 percent compression of the most expensive preclinical phase compounds over every program a company runs. And the history that governed the old model no longer applies — brute-force screening of hundreds of thousands of compounds is no longer the binding constraint. Generative AI could deliver between $60 billion and $110 billion in annual value for the pharmaceutical industry overall, according to one widely cited estimate, and that value accrues regardless of where the investment cycle happens to be.
But the structural shift in discovery does not imply a structural shift in total development time, and confusing the two is the most common error in this debate. AI changes the front end of the funnel. It does not change the fact that a Phase III trial in a chronic disease runs for years, that regulators require evidence no simulation can yet substitute, and that a molecule must still be manufactured at scale under conditions no model fully controls. The productivity gain is real, durable, and structurally important — and it is also bounded.
Who Captures the Value: Platforms Versus Portfolios
Here is the second-order question the market is only beginning to price correctly: if AI compresses discovery, who gets rich? The intuitive answer — the AI platform companies — is precisely the one most likely to be wrong. McKernan's framing is blunt: no large pharmaceutical buyer is paying billions for an AI platform alone. They buy portfolios of assets. The implication is that value migrates away from the toolmakers and toward the entities that can execute the parts of the process AI cannot touch — running trials, navigating regulators, and commercializing approved drugs.
This is already visible in deal structure. The 50-to-1 biobucks-to-upfront ratio is not merely a sign of caution; it is a repricing of bargaining power. When discovery becomes cheaper and faster, the scarce asset is no longer a molecule — it is clinical execution capacity, regulatory credibility, and patient access. Big pharma, sitting on all three, pays less upfront and defers value into milestones. The AI startups, meanwhile, face a capital-intensive valley between a promising preclinical candidate and the Phase II data that would prove their technology actually improves outcomes. Several have not survived the crossing.
The pattern is visible in the largest deals of the past few years. When AbbVie paid $8.7 billion for Cerevel Therapeutics and Bristol Myers Squibb paid $14 billion for Karuna Therapeutics, they were not buying AI platforms. They were buying neuroscience portfolios of approved or near-approved assets — precisely the kind of de-risked, clinical-stage medicine that big pharma values and that pure discovery tools cannot replicate on their own.
The counter-move for AI-native companies is to become asset owners rather than service providers — to advance their own molecules into the clinic and capture the full value chain. Insilico Medicine's lung-drug candidate Rentosertib, discovered and designed through its Pharma.AI platform and advanced from target identification to Phase I in about 30 months, is the model. Its Phase IIa results in idiopathic pulmonary fibrosis were published in June 2025, and the company has since initiated a Phase III trial. But that path demands hundreds of millions of dollars and a tolerance for risk that many venture funds lost after the 2025 correction. The sector is converging on an uncomfortable truth: being good at AI and being good at drug development are different capabilities, and the market pays for the latter.
The Counter-Case: Skepticism and the Phase III Reckoning
The strongest argument against the bullish thesis does not come from Luddites. It comes from Jennifer Doudna, the Crispr pioneer and Nobel laureate, who has publicly challenged the idea that generative models might soon get credit for drug discovery, questioning whether AI can replace the human effort at the core of scientific discovery. Her skepticism is not anti-technology; it is a demand for evidence at the only place that matters. Scientific commentators have gone further, noting that AI-discovered compounds have so far shown progression rates similar to traditionally discovered molecules, and warning that the Phase III readouts expected through 2026 and 2027 may demonstrate accelerated timelines without improved efficacy — commercially valuable, but scientifically underwhelming.
This counter-thesis attacks the bull case at its foundation. If AI only makes discovery faster without making it better — if the molecules it produces fail in Phase II and III at the same rate as the old ones — then the technology is a cost cutter for pharma R&D, not a revolution in medicine. The value would accrue as margin improvement for large drugmakers, not as a wave of breakthrough therapies or platform-company valuations. That outcome is entirely consistent with the data we have: Phase I success is elevated, Phase II is flat, and the sample is too small to conclude anything about Phase III.
The falsifying signal is specific and observable. If, across the Phase III readouts expected through 2026 and 2027, AI-originated candidates do not show a materially higher approval rate than the historic industry baseline — roughly one in ten candidates entering clinical development — then the structural-revolution thesis is wrong, and the sector should be valued as a process-optimization tool, not a paradigm shift. Conversely, if multiple AI-designed drugs reach approval with success rates demonstrably above baseline, the platform skeptics are wrong and the valuation reset of 2025-26 will look like a buying opportunity in hindsight. Watch the Phase III data, not the press releases about partnership announcements.
What Comes Next: Scenarios by Time Horizon
Short term (6-18 months): volatility around clinical readouts. The first wave of pivotal AI-drug data will hit in 2026 and 2027, and the market will react violently to each result. A single high-profile failure would compress valuations for the whole category. A single clear approval would do the opposite. Regulatory clarity will also arrive: the US Food and Drug Administration's AI guidance is expected to be finalized in 2026, requiring credibility-assessment plans for high-risk AI applications, while the EU AI Act's high-risk provisions take effect in August 2026 and may classify some drug-development AI as high-risk. Compliance costs will rise, favoring larger players.
Medium term (2-5 years): consolidation and the asset test. The sector will consolidate. Companies that can pair AI discovery with credible clinical execution will be acquired or will survive as independents; pure-platform vendors without proprietary data or assets will be forced into service-model economics or shut down. The winners will be those that have manufactured their own proprietary experimental data — the only moat that matters when models themselves are commoditizing. Expect the 50-to-1 biobucks ratio to compress toward reality as upfront cash becomes the currency of trust.
Long term (5+ years): a structurally different industry, but not a magic one. If the Phase III evidence validates the technology, drug discovery becomes faster, cheaper, and more targeted as a permanent feature of the industry — a structural shift comparable to the advent of high-throughput screening or structure-based design, not a discontinuity. The beneficiaries will be patients (more candidates tested against more targets), large pharma (higher R&D productivity), and the asset-owning biotechs that survive the consolidation. The losers will be the platform companies that mistook a tool for a business model.
The base case is the bounded-revolution scenario: AI delivers durable, structural gains in early discovery and patient stratification, lifts Phase I success rates, but leaves the clinical attrition problem largely intact until better biomarkers and trial designs mature. Approvals will arrive — more than a handful by the end of the decade — but not enough to justify the 2021-22 platform valuations. The investment cycle will remain cyclical, swinging on individual readouts, while the technology underneath keeps improving regardless.
McKernan's verdict — real, but narrow — is the right lens. AI is accelerating drug discovery in the ways that are measurable today: faster target identification, better hit rates, smarter patient selection. What it has not done, and may not do for years, is solve the part of drug development that actually determines whether a patient gets a medicine. The companies that understand that distinction will be the ones that capture value from the revolution. The ones that don't will be the consolidation statistics.
Explore more exclusive insights at nextfin.ai.
