DeepMind CEO Demis Hassabis Pursues AI-Automated Drug Design

At the World Economic Forum in Davos in January 2026, Demis Hassabis, the Nobel Prize–winning CEO of Google DeepMind and Isomorphic Labs, declared that the first AI-designed cancer drug would enter Phase I clinical trials by early 2026 [1]. A month later, Isomorphic Labs released IsoDDE — a drug design engine it called the most advanced publicly disclosed AI system for computational drug design [2]. The company now manages 17 active drug programs across oncology, immunology, and cardiovascular disease [1], backed by pharmaceutical partnerships worth nearly $3 billion [3].

The promise is extraordinary: compress drug discovery from a decade-long, $2.6-billion ordeal into something measured in months and millions. The question is whether the biology will cooperate.

The $3 Billion Bet — and What It Actually Buys

Isomorphic Labs, spun out of DeepMind in 2021, secured two major partnerships in January 2024. Eli Lilly committed $45 million upfront with up to $1.7 billion in milestone payments. Novartis put up $37.5 million upfront with $1.2 billion in milestones [3][4]. The combined headline figure of nearly $3 billion is almost entirely contingent on Isomorphic hitting specific drug development targets — meaning the guaranteed cash is under $83 million.

Source: FierceBiotech

Data as of Feb 1, 2025CSV

In February 2025, Novartis doubled down, expanding the partnership to include up to three additional research programs targeting biological mechanisms previously considered "undruggable" [4]. In January 2026, Isomorphic announced a new research collaboration with Johnson & Johnson [2]. The deals cover small molecules and biologics across oncology and immune-mediated disorders [3][4].

Several lead candidates from these partnerships are now in IND-enabling studies — the final preclinical stage before human trials — with Phase I entry projected for late 2026 [4]. Hassabis had originally targeted 2025 for clinical entry; the timeline has slipped by roughly a year [2].

What IsoDDE Can Do — and What It Cannot

AlphaFold solved one problem: predicting how a protein folds into its three-dimensional shape from its amino acid sequence. Drug design requires solving several others. A candidate molecule must bind to the right protein target with enough affinity to be therapeutically useful, avoid binding to the wrong targets, survive metabolism in the human body, and not poison the patient. These properties — collectively called ADMET (absorption, distribution, metabolism, excretion, and toxicity) — have historically been where computational models fail and human chemists take over [5][6].

IsoDDE, released in February 2026, represents Isomorphic's attempt to close these gaps. The system has four principal capabilities: protein-ligand structure prediction, antibody-antigen interface modeling, binding affinity estimation, and ligandable pocket identification [2].

On benchmarks, it is a substantial advance. IsoDDE more than doubles AlphaFold 3's accuracy on the hardest protein-ligand generalization targets. For antibody-antigen complexes, it achieves 39% accuracy in high-fidelity predictions, compared with 17% for AlphaFold 3 and 2% for Boltz-2 [2]. Its binding affinity predictions match or exceed physics-based methods like free energy perturbation (FEP), a computationally expensive gold standard, while running orders of magnitude faster [2].

But IsoDDE does not model ADMET properties. It does not predict off-target binding across the human proteome. It does not simulate how a molecule will behave in a living organism over weeks of dosing [5][6]. These steps still require separate computational tools, wet-lab experiments, and the judgment of medicinal chemists. AlphaFold and its successors are also inherently limited to static protein conformations — they do not capture the dynamic shape-shifting that proteins undergo when they interact with drugs or other molecules, including allosteric transitions and ligand-induced rearrangements [5].

In short, IsoDDE addresses the front end of drug design — finding the right target and designing a molecule that fits — while much of the back end remains a hybrid of AI, traditional computational chemistry, and human expertise.

The Pipeline: Where Things Stand

Across the entire AI drug discovery industry — not just Isomorphic — the clinical pipeline has grown rapidly. As of early 2026, roughly 173 programs are in preclinical development, 94 are in Phase I, 56 in Phase II, and 15 have reached Phase III [7][8].

Source: Axis Intelligence / Drug Target Review

Data as of Apr 1, 2026CSV

The early-stage numbers are encouraging. AI-designed compounds show Phase I success rates of 80–90%, significantly above the historical average of 52% [8]. But Phase I primarily tests safety, not efficacy. In Phase II, where drugs must demonstrate they actually work against a disease, the success rate for AI-discovered molecules drops to approximately 40% — statistically indistinguishable from the industry baseline [7][8].

No AI-discovered drug has received FDA approval as of April 2026 [7][8]. Industry analysts assign roughly a 60% probability that the first approval will come in 2026 or 2027 [8].

The Failures That Define the Field

The most instructive data points are the failures.

Recursion Pharmaceuticals' REC-994, targeting cerebral cavernous malformation (a rare brain vascular condition), was the company's highest-profile AI-discovered compound. It met its Phase II primary endpoint of safety and tolerability, and early MRI data showed a trend toward reduced lesion volume at the higher dose [9][10]. But long-term extension data told a different story: patients who crossed from placebo to the active drug showed no improvement, and the group that stayed on the drug did not sustain its initial results. Recursion discontinued the program in May 2025 [9][10].

Exscientia's DSP-1181, developed for obsessive-compulsive disorder, was celebrated for reaching Phase I in just 12 months — roughly a quarter of the typical timeline. It was later discontinued after Phase I [8][11].

Insilico Medicine's rentosertib — the first drug where both the target and the molecule were identified entirely by generative AI — has had a more complex trajectory. Phase IIa results published in Nature Medicine in June 2025 showed patients receiving the highest dose had a mean improvement in lung function of 98.4 milliliters, compared to a 62.3-milliliter decline in the placebo group [8][12]. Insilico is pursuing Phase III discussions with regulators, but the data fell short of conventional statistical significance thresholds, and the sample size was small [7][12].

These results suggest a consistent pattern: AI can accelerate early discovery and improve the odds of reaching human trials, but it has not yet demonstrated an ability to improve the odds of a drug actually working in patients.

How Much Can AI Actually Save?

The pharmaceutical industry spends an average of $2.6 billion and 10–15 years to bring a single drug from initial research to market approval [13]. AI's clearest impact is on the discovery phase — the first 3–5 years of identifying targets and designing candidate molecules.

A BCG and Wellcome Trust analysis estimated that AI could yield 25–50% cost and time savings through the preclinical stage [13][14]. More aggressive projections from industry advocates suggest 40–60% total cost reductions if AI is integrated across the full development pipeline, including clinical trial design and patient selection [13].

The concrete evidence so far is more modest. Insilico Medicine compressed target identification to Phase I from a typical 6–8 years to 30 months [8]. Exscientia reduced compound synthesis requirements by 85% in its OCD program [13]. These are real efficiencies, but they apply to the cheapest phase of development. Clinical trials — Phase I through Phase III — account for roughly 60–70% of total drug development costs, and AI's impact on trial design and execution remains largely theoretical [13][14].

Source: OpenAlex

Data as of Jan 1, 2026CSV

Academic interest in the field has exploded: over 40,000 papers on AI drug discovery were published in 2025 alone, a nearly 25% increase over 2024 [15]. But publication volume is a measure of enthusiasm, not clinical results.

Who Captures the Savings?

Even if AI cuts discovery costs substantially, the economic benefits may not reach patients. Drug pricing in the United States is not primarily a function of R&D cost — it is determined by what the market will bear, shaped by patent exclusivity, insurance reimbursement structures, and the absence of direct government price negotiation for most drugs [16][17].

AI-accelerated development could actually increase the effective patent life of a drug. If a company discovers a molecule two years faster, it enters the market two years earlier while the 20-year patent clock started at filing. The result is more years of exclusivity pricing, not lower prices [16]. Pharmaceutical companies have historically priced drugs based on the value they deliver to patients and what competing treatments cost — not on what it cost to develop them [17].

The structural argument is straightforward: unless regulatory or legislative changes force the savings downstream, AI-driven cost reductions will flow to shareholders and reinvestment, not to patients and payers. The Inflation Reduction Act's provisions for Medicare drug price negotiation cover only a limited number of drugs, and do not apply to commercial insurance [17].

The China Factor

While Hassabis builds his pipeline from London and Google's Mountain View headquarters, China has emerged as a serious competitor. In 2025, five of the ten largest pharmaceutical R&D licensing deals originated from Chinese companies, accounting for 38% of all big-pharma in-licenses with $50 million or more in upfront payments [18].

XtalPi, one of China's leading AI drug discovery firms, reported revenue growth of over 400% in the first half of 2025, driven by partnerships with Pfizer and Eli Lilly [18][19]. The company struck a deal with U.S.-based DoveTree potentially worth over $10 billion — among the largest AI drug discovery agreements ever [19]. Insilico Medicine listed on the Hong Kong Stock Exchange in late 2025, raising nearly $300 million and becoming the first AI-driven biotech to IPO on the Main Board [18].

China now accounts for nearly 30% of global pharmaceutical innovation pipeline assets, up from 10% in 2019 [18]. The country's "Four Little Dragons" of AI drug discovery — XtalPi, Insilico Medicine, DeuteRx, and Deep Intelligent Pharma — are both designing molecules and licensing them to Western pharmaceutical companies at scale [20].

The competitive dynamic is not simply East versus West. Western pharma is actively buying Chinese AI-discovered molecules, creating an interdependent ecosystem. But the speed and scale of China's investment raises questions about whether companies like Isomorphic Labs, which have yet to produce a clinical-stage drug from their own pipeline, risk being overtaken by competitors with faster paths to clinical data.

The Skeptic's Case

The strongest argument against the Hassabis narrative is not that AI is useless — it is that AI solves the wrong bottleneck.

Most drugs fail not because chemists cannot design molecules that bind to targets. They fail because the targets themselves turn out to be wrong, because the disease biology is more complex than the model assumed, or because the drug causes unacceptable side effects in humans that no computational model predicted [7][8]. Phase II and Phase III failures are dominated by lack of clinical efficacy (roughly 50% of failures) and safety concerns (roughly 30%), according to long-standing FDA analyses [7].

Structure prediction — the problem AlphaFold solved — addresses the earliest and cheapest step in drug development. A skeptical pharmacologist would note that knowing a protein's structure tells you little about how a disease actually works in a patient, whether interfering with that protein will produce a therapeutic benefit, or whether the benefit will outweigh the harm. These are questions that require clinical trials in human beings, and no AI system has shown it can reduce the failure rate at that stage [7][8].

The track record reinforces this view. AI-discovered molecules progress through Phase I at higher rates than traditionally discovered ones — but Phase I is a safety screen, not an efficacy test. Once drugs reach Phase II, where they must demonstrate clinical benefit, AI-discovered compounds perform no better than the historical average [7][8]. If this pattern holds through Phase III, the technology's contribution will be real but narrower than advertised: faster and cheaper discovery of drug candidates, but no improvement in the probability that those candidates become approved medicines.

What 2026 Will Reveal

The next 12–18 months represent the most consequential period for AI-driven drug design since AlphaFold's debut. Fifteen programs are in Phase III trials, with results expected through 2026 and into 2027 [8]. Isomorphic Labs' own candidates are approaching clinical entry [1][2]. The FDA and European Medicines Agency released joint guiding principles on AI in drug development in January 2026, signaling regulatory readiness [8].

If Phase III data shows that AI-discovered drugs succeed at rates meaningfully above the industry average — currently around 50–60% for drugs that reach that stage — the case for automated drug design becomes difficult to argue against. If the success rates match the baseline, the technology will still have value as a speed and cost optimizer, but the grander vision Hassabis has articulated — AI that "solves disease" — will require significant revision.

The $3 billion in partnership commitments, the 17 active programs, the benchmark-beating IsoDDE engine: these are substantial achievements. Whether they translate into medicines that reach patients depends on whether a problem that starts in silicon can be solved without the messy, expensive, irreducibly biological reality of human clinical trials. That question has no computational shortcut.