AI Systems Could Compress Brain Drug Discovery from Decades to Years, Researchers Say

The pitch is seductive: artificial intelligence can compress the search for new brain drugs from decades to years, rescuing patients from diseases that have defeated hundreds of well-funded pharmaceutical programs. More than $18 billion has flowed into over 200 biotech companies making some version of this claim [1]. Eli Lilly signed a $1.75 billion deal with Isomorphic Labs. Bayer committed $1.5 billion to Recursion Pharmaceuticals [2]. Academic publications on "AI drug discovery" surged from roughly 2,200 in 2011 to over 40,000 in 2025 [3].

But underneath the capital and the conference keynotes lies a set of harder questions. CNS (central nervous system) drug development has a 45% higher failure rate than drugs targeting other organ systems [4]. The average timeline from compound identification to FDA approval runs 13 to 15 years and costs $2 to $3 billion [5]. Disease-modifying treatments for major neurodegenerative conditions have historically had a 100% clinical failure rate [4]. Against that record, what exactly does AI change — and what doesn't it?

Source: OpenAlex

Data as of Jan 1, 2026CSV

The Barren Landscape: Three Decades of CNS Drug Approvals

Between 2008 and 2023, the FDA approved 118 medications across psychiatric categories — but only 37 of those were genuinely new drug applications rather than reformulations or repurposed compounds [6]. Of the new psychiatric drugs approved between 2012 and 2024, just seven (31.8%) were first-in-class, meaning they worked through a previously unexploited mechanism. Thirteen (59.1%) were classified as "addition-to-class" — incremental additions to existing drug families [7].

Source: Psychiatrist.com / ResearchGate Review

Data as of Jan 1, 2024CSV

The distribution tells its own story. Schizophrenia received the most approvals (26), followed by ADHD (18), depression and substance use disorders (16 each), and sleep-wake disorders (13). Neurocognitive disorders — the category that includes Alzheimer's disease, the condition attracting the most AI investment — received just six [6]. Between 2013 and 2024, the FDA approved only 16 truly novel drugs to treat psychiatric diseases [8].

For neurological conditions more broadly, the picture is similarly sparse. The recent accelerated approvals of lecanemab (Leqembi) for Alzheimer's and tofersen (Qalsody) for ALS represented genuine milestones, but both remain controversial — lecanemab's clinical benefit is modest and its side effects are significant [9]. Average development timelines broken down by disease category remain punishing: Alzheimer's programs routinely span 15 or more years from target identification to market, with most ending in failure. Depression timelines are somewhat shorter (8-12 years) partly because existing drug classes provide validated mechanisms to build on, while schizophrenia and epilepsy programs fall somewhere in between [5][7].

The financial toll is staggering. CNS disorders generate upwards of $80 billion per year in pharmaceutical revenue, yet the failure rate has "sent billions of dollars down the drain," as one analysis from Drug Discovery World put it [10]. Each failed late-stage CNS trial can cost a sponsor $500 million or more. The total global annual R&D spend on CNS drug discovery is difficult to isolate precisely, but estimates from the Tufts Center for the Study of Drug Development and industry analyses suggest it falls in the range of $15-20 billion annually, with the majority of that investment historically written off as failure [5][10].

What AI Actually Does — and Where It Stops

The AI drug discovery industry makes its case primarily in two domains: molecular design and target identification. Platforms like Insilico Medicine's Chemistry42 can generate a lead compound in 21 days — a process that conventionally takes months to years [2]. Recursion Pharmaceuticals uses high-throughput biological imaging combined with machine learning to identify drug-disease connections across massive datasets. Exscientia (now merged with Recursion as of July 2025) designed EXS-21546, an adenosine receptor antagonist, using only 163 synthesized compounds to reach a Phase I-ready candidate — a fraction of the thousands typically required [2][11].

For blood-brain barrier (BBB) permeability — one of the central obstacles in CNS drug development — AI models predict which molecules can cross into brain tissue far more rapidly than traditional physicochemical screening. A 2024 study in Nature Scientific Reports demonstrated that large language models and machine learning classifiers can predict BBB permeability with increasing accuracy [12]. One group is making a $70 million bet that AI can solve the BBB problem outright [13].

The clinical trial data supports part of the thesis. AI-discovered molecules show an 80-90% success rate in Phase I trials, substantially exceeding the historic industry average of roughly 52% [14]. This makes sense: AI excels at designing molecules that are safe, drug-like, and well-tolerated. The algorithms are good at what computational chemistry has always been good at, only faster and at greater scale.

Source: ScienceDirect / Drug Discovery Today

Data as of Jun 1, 2024CSV

But at Phase II — where efficacy in actual patients is tested — the AI advantage evaporates. Success rates drop to approximately 40%, comparable to the industry average of 29% [14]. This convergence is the most important number in the entire AI drug discovery debate, because it suggests that the technology's power lies in speeding up the early, computational stages of discovery while leaving the fundamental biological uncertainties of human disease intact.

No AI-identified CNS drug candidate has yet completed a Phase III clinical trial. Recursion's REC-994 for cerebral cavernous malformation and REC-2282 for NF2-related schwannomatosis are in Phase 2 — the most advanced CNS-specific AI programs in clinical development [11]. The most consequential data is expected to arrive from Phase III readouts in late 2026 and 2027 [15].

The Bottleneck Downstream

Independent neuropharmacologists and clinical trialists are increasingly vocal about what they see as a misdiagnosis of the problem. The claim that AI can compress CNS drug timelines rests on the assumption that the bottleneck is in identifying candidate molecules. But as an analysis in Applied Clinical Trials argued, "discovery innovation is driven by cheap-to-deploy computational methods, while development requires touching actual trial infrastructure: sites, IRBs, patient populations, regulatory submissions, comparator datasets, endpoint definitions" [16].

Clinical development — the Phase I through Phase III process — takes six to ten years from first-in-human dosing to regulatory approval, and AI cannot meaningfully compress this timeline. Patients must be recruited. Safety must be monitored over months or years. Regulatory agencies require specific data packages that cannot be generated computationally. As DeepCeutix, an AI drug formulation company, put it bluntly: "Your AI Can Design a Molecule. It Can't Formulate a Drug" [17].

The formulation gap is real but often overlooked in public discussions of AI-driven timelines. A computationally designed molecule that works in a test tube must still be manufactured at scale, packaged in a deliverable form, and shown to reach its target in the human body at therapeutic concentrations. For CNS drugs specifically, the BBB adds an additional layer: even if AI can predict which molecules cross the barrier in silico, confirming that prediction in living patients requires years of clinical work.

Derek Lowe, a medicinal chemist and prominent industry commentator, described the current moment as "between hype and hope" at the 2025 Bio-IT World conference. He noted that the foundational ideas behind computational drug design "have been around for decades" — a 1981 Fortune magazine cover story touted similar promises [18]. The question is whether today's tools represent a qualitative break from that history or merely the latest iteration.

Echoes of Previous Revolutions

The AI drug discovery wave is the fourth major technological promise to sweep CNS medicine in 35 years. In the 1990s, the Human Genome Project was expected to unlock the molecular basis of brain disease and usher in targeted therapies within a decade. High-throughput screening in the 2000s promised to brute-force drug discovery by testing millions of compounds simultaneously. Stem cell therapies in the 2010s offered the prospect of replacing damaged neurons entirely [18][19].

Each wave produced real scientific advances. Genomics yielded GWAS (genome-wide association studies) that identified hundreds of risk loci for schizophrenia, Alzheimer's, and depression. High-throughput screening contributed to several marketed drugs. Stem cell research advanced understanding of neuronal differentiation. But none delivered on the timeline promises made to investors and patients. The gap between scientific capability and clinical translation remained stubbornly wide [18].

AI occupies a position that industry analysts describe as "between the Peak of Inflated Expectations and the Trough of Disillusionment" on the Gartner hype cycle [19]. What arguably distinguishes this moment is the scale of biological data now available (millions of protein structures via AlphaFold, petabytes of clinical imaging, electronic health records) and the computational power to process it. Whether that distinction is sufficient to overcome the biological obstacles that defeated previous waves remains an open question.

A key critique, voiced by Lowe and others: "AI has not solved — and likely cannot solve — the fundamental challenges of clinical validation, regulatory approval and commercial success." The technology is "a powerful tool, not a panacea" [18][19].

Following the Money

The financial incentives shaping AI drug discovery claims deserve scrutiny. Of the more than $18 billion invested, the strongest funding rounds increasingly combine AI with experimental execution — not just computational predictions [1][20]. But investor confidence has been wavering: deal flow has slowed, with fewer than 20 qualifying deals in recent periods worth roughly half of the 2021 peak [20].

The venture capital concentration is notable. Investment is heavily clustered in California and Massachusetts, with fewer than a handful of firms (Menlo Ventures and Sequoia Capital among them) appearing across multiple deals [1]. The licensing deals that generate the largest headlines — Eli Lilly's $1.75 billion agreement with Isomorphic Labs, Bayer's $1.5 billion commitment to Recursion — are structured with heavy milestone payments, meaning the upfront cash is a fraction of the announced figure [2].

NIH grant competition adds another layer of pressure. Academic researchers who frame their work in terms of AI-driven acceleration are competing for a finite pool of funding against traditional approaches. This creates structural incentives to emphasize speed and disruption in grant applications, even when the underlying science supports more modest claims [20].

As Fortune reported in September 2025, "the AI drug breakthrough is taking a long time to arrive, for reasons that may have little to do with the technology's limits" — pointing to the gap between computational prediction and the messy realities of drug development [21].

Who Benefits First — and Who Waits

The equity implications of AI-driven CNS drug discovery are significant and underexamined. A 2024 analysis in ACS Journal of Chemical Information and Modeling found that only a fraction of neurological diseases has been the focus of machine learning research, generally those with larger patient populations and more published data [22].

Alzheimer's disease — affecting an aging, predominantly Western demographic with purchasing power and insurance coverage — dominates the AI drug discovery pipeline. Parkinson's disease is a distant second. Rare pediatric neurological disorders, despite being theoretically well-suited to AI approaches (many are monogenic, meaning a single gene is responsible), have received comparatively little attention [22][23].

This is a paradox worth pausing on. Monogenic diseases are, in computational terms, among the most tractable problems in drug discovery. A single gene target, a clear mechanism, a defined patient population. Yet the economics of drug development — the need to recoup billions in R&D costs — mean that conditions affecting tens of thousands rather than tens of millions are consistently deprioritized [22].

Conditions prevalent in low-income countries face a double disadvantage: smaller datasets for training AI models and weaker commercial incentives for investment. Drug repurposing — using AI to identify new uses for existing approved drugs — offers a partial solution for rare and neglected diseases, but this remains a modest part of the literature compared to cancer and common chronic conditions [23].

The technology is not neutral in whose suffering it addresses first. It follows the money, and the money follows the markets.

What the Skeptics Get Right

The most important critique of AI drug discovery timelines does not come from Luddites or competitors. It comes from computational chemists and clinical trialists who understand both the power and the limits of the technology.

Their argument can be summarized in three points. First, AI has genuinely compressed the design-make-test cycle in early discovery. A computational chemist who once analyzed 20-50 compounds per week can now evaluate thousands [16]. That is a real advance.

Second, the bottleneck has shifted, not disappeared. It now sits in clinical development: trial design, patient recruitment, safety monitoring, and regulatory review. These processes involve human biology, institutional infrastructure, and regulatory frameworks that cannot be accelerated computationally [16][17].

Third, data quality — not algorithmic sophistication — is the binding constraint. AI models are only as good as the data they are trained on, and much of the relevant biological data for CNS diseases is noisy, incomplete, or locked in proprietary silos. The BBB permeability models that perform well on benchmark datasets may fail when applied to structurally novel compounds [12][13].

The 85% Phase I success rate for AI-discovered drugs is real, but it answers the wrong question. The question was never "can we design molecules that are safe enough to give to humans?" The question is "can we find molecules that actually treat brain disease?" And there, the 40% Phase II success rate — statistically indistinguishable from the industry average — suggests that AI has not yet changed the fundamental calculus [14].

What Comes Next

The next 18 to 24 months will provide the first genuine test of AI's impact on CNS drug discovery. Recursion's Phase 2 CNS programs will read out. The first AI-designed drugs will enter Phase III trials. The merged Recursion-Exscientia entity will attempt to prove that integrating high-throughput biology with automated chemistry produces better outcomes than either approach alone [11][15].

If those trials succeed, the implications would be substantial — not because AI "solved" brain disease, but because it demonstrated that computational tools can identify viable drug candidates that human researchers missed. If they fail, the industry will face a reckoning similar to the genomics correction of the early 2000s: real science, real tools, but timelines measured in decades rather than years.

The 75 drugs or vaccines from AI companies currently in clinical trials represent the largest-ever test of computationally driven drug discovery [1]. For the millions of patients with neurological and psychiatric conditions who have exhausted existing treatment options, the stakes extend well beyond stock prices and licensing deals. The technology's potential is real. The question is whether the promises being made to investors are the same as the promises being kept to patients.