IPM Take
AI drug discovery is often sold on speed, but AstraZeneca’s new partnership with insitro is aimed at something potentially more important: deciding what biology is worth pursuing before years of development and hundreds of millions of dollars are committed to it.
The collaboration covers metabolic dysfunction-associated steatohepatitis, or MASH, and cardiomyopathies, two areas where superficially similar patients can have very different underlying disease biology. Insitro will use its Virtual Human platform to combine large-scale human genetics, clinical information and experimentally generated cellular data in an attempt to identify causal disease drivers and nominate new therapeutic targets.
AstraZeneca will then take selected programmes forward, holding exclusive worldwide rights and assuming responsibility for clinical development, regulatory work and commercialisation.
The headline financial figure needs perspective. This is not a $1.5 billion upfront investment in proven drugs. Insitro receives $40 million initially, while approximately $1.46 billion depends on future development and commercial milestones being achieved, with potential single-digit royalties on global sales.
That structure tells its own story.
Big Pharma is willing to pay heavily for AI-derived biology when it works, but much of the financial risk remains tied to whether algorithms can ultimately produce targets that survive experimental validation, clinical trials and regulatory scrutiny.
The real benchmark for AI in drug discovery is therefore shifting. The question is no longer whether a model can identify an interesting target.
It is whether that target becomes a medicine.
Executive Summary
AstraZeneca and insitro announced a strategic drug-discovery collaboration on 15 October focused on discovering new therapies for MASH and cardiomyopathies, two heterogeneous cardiometabolic disease areas in which identifying causal biological drivers remains difficult.
Under the agreement, AstraZeneca will pay insitro $40 million upfront. Insitro is eligible for approximately $1.46 billion in potential development and commercial milestone payments, plus potential single-digit royalties on global sales. The milestone figure is contingent and should not be interpreted as money already committed or guaranteed.
Insitro will apply its Virtual Human AI platform, which integrates large-scale human and cellular datasets with machine learning, statistical genetics and computational approaches designed to identify disease mechanisms and therapeutic targets.
The collaboration will cover AstraZeneca-nominated targets, with insitro working from target discovery through development-candidate generation. AstraZeneca will hold exclusive worldwide rights to selected programmes and will lead subsequent clinical development, regulatory activities and commercialisation.
The agreement reflects a broader shift in pharmaceutical AI from optimising individual molecules toward using multimodal datasets to improve target selection, one of the earliest and most consequential decisions in drug development.
However, no clinical efficacy has yet been demonstrated from programmes created under the collaboration. The partnership is fundamentally an early discovery initiative, and whether AI-derived targets produce clinically useful medicines will require experimental, preclinical and ultimately human validation.
Why it matters
- HTA bodies: The collaboration is far upstream of HTA, but it highlights a growing issue for future evidence assessment. AI may increasingly influence which disease subtypes and biological targets reach clinical development, potentially generating more genetically or molecularly defined populations. HTA systems may eventually face therapies developed for narrower mechanistic subgroups, increasing the importance of biomarker strategy, external validity and evidence on which patients genuinely benefit.
- Payers: The promise of AI-enabled target selection is partly economic. If better biological understanding reduces late-stage failure, development costs and wasted investment could fall. Whether those efficiencies ultimately translate into lower prices or better value for health systems is a separate question, and one that cannot be assumed simply because discovery becomes more computationally efficient.
- Industry / innovation partners: The deal reinforces a shift in the competitive value of AI platforms. Access to models alone is becoming less distinctive than access to high-quality human genetics, deeply phenotyped datasets, scalable experimental systems and the ability to connect computational predictions with drug development. The companies that can demonstrate that those systems generate therapeutically valid targets, rather than merely plausible ones, will have the stronger proposition.
Drug discovery has always had a target problem.
Before a molecule can be optimised, before a clinical trial can be designed and long before regulators or payers become involved, researchers have to decide which piece of biology is worth changing. Get that decision right and years of chemistry and clinical development have a chance to produce a medicine. Get it wrong and an apparently successful programme can progress deep into development before revealing that the underlying target never mattered enough in human disease.
AstraZeneca’s new collaboration with insitro is an attempt to intervene much earlier in that chain.
The agreement, announced on 15 October, centres on MASH and cardiomyopathies, two disease areas in which the clinical label can conceal substantial biological heterogeneity. Two patients may meet the same diagnostic criteria while reaching that diagnosis through different genetic, metabolic, inflammatory or structural pathways, which makes the search for targets that matter across meaningful groups of patients considerably more difficult.
Rather than beginning with a single hypothesis and testing it conventionally, insitro aims to use large quantities of human and cellular data to work backwards from disease biology. Its Virtual Human platform integrates population-scale genetics and clinical information with experimentally generated cellular datasets, applying machine learning to identify patterns that may point toward causal mechanisms rather than simple correlations.
That distinction is central to the partnership.
Pharmaceutical research does not suffer from a shortage of biological associations. Modern datasets can reveal thousands of genes, proteins and pathways that correlate with disease. The harder question is which of those relationships actually drive disease strongly enough that modifying them will produce a therapeutic effect.
This is where AstraZeneca believes AI can change the economics of discovery.
Sharon Barr, the company’s head of BioPharmaceuticals R&D, described target selection as one of the most critical decisions in pharmaceutical research. The partnership effectively places AI at that decision point, using computational analysis not merely to speed up tasks researchers already perform but to determine which biological hypotheses deserve to become drug programmes.
For insitro, the attraction is equally clear. Its models can generate hypotheses, but AstraZeneca brings the infrastructure needed to turn those hypotheses into medicines, including medicinal chemistry, clinical development, regulatory capabilities and global commercial scale. Under the agreement, insitro will work on identifying causal disease biology and advancing AstraZeneca-nominated targets toward development candidates, after which AstraZeneca will lead clinical development and commercialisation of selected programmes.
The economics of the deal reveal how much uncertainty remains.
Insitro receives $40 million upfront, while the much larger figure that has generated headlines, approximately $1.46 billion, is attached to future development and commercial milestones. The company could also receive single-digit royalties if resulting products ultimately reach the market.
In other words, the partnership may be worth close to $1.5 billion if programmes succeed, but most of that value lies on the other side of scientific and clinical hurdles that have not yet been crossed.
That is important in a pharmaceutical AI market where partnership values can easily create the impression that computational drug discovery has already proved its clinical value.
It has not.
AI can identify patterns in datasets, prioritise targets and help design molecules, but the ultimate experiment remains human biology. A target that looks compelling in genetics and cellular models still has to produce the expected pharmacological effect, demonstrate an acceptable safety profile and improve outcomes in clinical trials.
MASH illustrates the challenge particularly well.
The disease sits at the intersection of metabolic dysfunction, liver injury, inflammation and fibrosis, while progression varies substantially between patients. Therapeutic development has historically been difficult partly because reducing one feature of the disease does not necessarily prevent progression of another. Identifying which mechanisms drive fibrosis and clinically consequential liver disease, rather than simply accompanying them, is therefore particularly valuable.
Insitro already has independent work in this field. Its own MASH programme has used human genetics and machine-learning approaches to identify IRS1 as a target, with the company reporting preclinical evidence for its CTRO-1013 programme during the American Diabetes Association meeting in 2026. That programme is separate from the newly announced AstraZeneca collaboration, but it illustrates the kind of discovery approach the pharmaceutical group is buying into.
Cardiomyopathies present a parallel problem from a different direction. The term covers multiple disorders of the heart muscle with diverse genetic and biological causes, and molecular classification is becoming increasingly important as cardiovascular medicine shifts away from treating cardiomyopathy as a single structural phenotype.
For AstraZeneca, those diseases also fit a wider strategy that increasingly treats cardiovascular, renal and metabolic disease as interconnected rather than isolated therapeutic categories. Its current pipeline spans obesity, dyslipidaemia, heart failure, kidney disease and MASH, among other areas, making target discovery across shared mechanisms potentially useful beyond a single indication.
The partnership therefore reflects a wider evolution in pharmaceutical AI.
The first wave of excitement focused heavily on whether machine learning could design molecules faster or predict which chemical structures might bind a target. Those capabilities remain important, but companies are increasingly pushing AI further upstream, into the much messier question of human disease causality.
That is where the potential reward becomes greater because failure at the target-selection stage can contaminate everything that follows.
If AI helps companies abandon weak biological hypotheses before large clinical programmes begin, the value may come as much from the drugs that are never developed as from those that eventually reach patients.
But that also makes success harder to measure.
A platform can generate dozens of apparently novel targets and still fail to create a clinically useful therapy. The relevant performance metric is therefore not the number of hypotheses produced, the amount of data analysed or even the speed of discovery. It is the proportion of AI-enabled programmes that survive experimental validation and eventually demonstrate meaningful effects in humans.
The AstraZeneca-insitro collaboration will take years to answer that question.
For now, it provides another sign that large pharmaceutical companies are becoming more willing to place AI inside the core machinery of R&D rather than treating it as a peripheral technology experiment. Yet the payment structure also shows that enthusiasm has not removed traditional discipline: most of the potential financial value remains conditional on programmes moving successfully through development.
That may be exactly where the field needs to be.
AI drug discovery does not need another demonstration that computers can generate biologically interesting ideas. Pharmaceutical research has never lacked interesting ideas.
The harder test is whether AI can help companies choose the right biology early enough to prevent the wrong drugs from consuming years of development, and whether the targets it prioritises ultimately produce medicines that change outcomes for patients.
