The pharmaceutical industry is moving through a structural transition as artificial intelligence becomes embedded in the way new medicines are discovered and developed. Traditional drug R&D is slow, expensive and risky, often taking ten to fifteen years and several billion dollars to bring a single new therapy to market.¹ Artificial intelligence offers a different operating model that uses data, algorithms and simulation to identify promising targets, design higher quality molecules and anticipate outcomes long before they are tested in large human trials. Early experience indicates that AI can compress timelines, reduce cost and improve the probability of success at multiple points in the value chain.²

Executives increasingly view AI not as an isolated technology project but as a strategic capability that sits at the core of discovery and development. Analysts expect that by 2030 the convergence of AI technologies and human expertise will be a major contributor to rising R&D productivity and a reversal of the long-term decline in biopharma innovation return on investment.² For investors and boards, the question is shifting from whether to adopt AI to how quickly the organization can scale it and where it can create the most defensible advantage.

Key AI technologies transforming drug R&D

Several AI capabilities are now mature enough to materially change how discovery and development decisions are made. These technologies are complementary and increasingly used together on integrated platforms.

Generative models and deep learning

Generative models and deep learning systems learn the complex rules that govern chemistry and biology and then propose new solutions. For small-molecule drugs, deep generative models can design novel compounds that satisfy multiple objectives such as potency, selectivity and predicted safety. Companies such as Insilico Medicine have shown that generative deep learning can move from target discovery to a clinical candidate in under two years, well ahead of traditional timelines.³ Deep learning models are also used to analyze large multi-omics and phenotypic datasets, revealing disease mechanisms and targets that may not be obvious from human inspection alone.

Protein structure prediction and AlphaFold

Structure-based drug design depends on knowing the three-dimensional shape of proteins. Historically, that required years of experimental work for each protein. The AlphaFold program from Google DeepMind solved a central part of this challenge by predicting protein structures with near-experimental accuracy in minutes, and the AlphaFold Protein Structure Database now provides open access to predictions for over two hundred million proteins.⁴ Reviews of AlphaFold’s impact highlight its role in accelerating structure-based drug design, protein engineering and the investigation of new biological targets.⁵

Large language models in biomedical research

Large language models are starting to act as research copilots across the drug development lifecycle. Fine-tuned biomedical models can read and synthesize scientific literature, patents and clinical data, help identify mechanistic hypotheses and summarize safety or efficacy findings. In discovery, LLMs help teams navigate the rapidly expanding corpus of publications. In development, they assist with protocol drafting, patient eligibility criteria and safety narrative generation. Some groups are now combining language models with structured biomedical knowledge graphs to propose new links between genes, pathways and potential interventions.

Reinforcement learning and multi-objective optimization

Reinforcement learning provides a way to iteratively optimize molecules and experimental designs. In medicinal chemistry, reinforcement learning agents treat molecule design as a sequence of actions, modifying structures to maximize a reward function that encodes potency, selectivity, solubility and other properties. This approach is particularly powerful for multi-objective optimization, where trade-offs are complex and the design space is enormous. Similar methods are being applied to reaction planning and synthetic route optimization, reducing laboratory trial-and-error and shortening the path from hit to lead.

Digital twins and virtual control arms

AI-driven digital twins are emerging as a powerful tool in late-stage development. Companies such as Unlearn build disease-specific models that generate a virtual control patient for each real participant based on baseline data and historical clinical records.⁶ Regulators in Europe have formally qualified this approach for use in certain phase two and three trials, and US regulators have provided positive feedback on its use in covariate-adjusted analyses.⁶ Digital twin-based control arms can reduce the number of participants who need to receive placebo and shorten the time required to reach robust statistical conclusions, creating both ethical and economic benefits.⁷

Leading companies and emerging platforms

A diverse ecosystem of companies is building AI-native platforms for drug discovery and development, alongside internal initiatives in large pharmaceutical organizations.

Insilico Medicine operates an end-to-end generative AI platform that spans target discovery, molecule generation and preclinical optimization. Its most advanced program, rentosertib, is a first-in-class TNIK inhibitor for idiopathic pulmonary fibrosis that was discovered and designed with AI and has reported encouraging phase two data.³ Recursion combines high-throughput biology, imaging and machine learning on what it calls the Recursion Operating System, generating one of the largest proprietary biological datasets in the industry. Recursion has struck multibillion-dollar discovery collaborations with partners such as Roche and Genentech for neuroscience and oncology programs.⁸

BenevolentAI uses a biomedical knowledge graph and machine learning platform to identify targets and repurposing opportunities. Its AI systems identified the rheumatoid arthritis drug baricitinib as a potential treatment for COVID-19 in early 2020, which later demonstrated a significant mortality benefit and secured full regulatory approval for hospitalized patients.⁹ Exscientia pioneered AI-designed small molecules that entered human trials and secured a landmark research alliance with Sanofi covering up to fifteen oncology and immunology candidates with total deal value of up to 5.2 billion US dollars.¹⁰

Technology companies are also investing heavily. DeepMind’s work on AlphaFold underpins its sister company Isomorphic Labs, which recently raised a large external funding round and is partnering with global pharmaceutical companies to move AI-designed drugs into the clinic.¹¹ Other specialist platforms include Atomwise for structure-based virtual screening, Schrödinger for physics-driven simulation coupled with machine learning, insitro for machine learning on human disease models, Valo Health for AI-enabled human data-centric discovery, and PathAI for AI-powered pathology that feeds into target and biomarker discovery.

Almost every global pharmaceutical company now has either an internal AI group, external partnerships or both. AstraZeneca, Sanofi, Roche, Novartis, Pfizer, Merck, GSK, Bayer, Takeda and Amgen all report multiple active collaborations with AI-native biotechs as well as internal programs focused on chemistry, biology and clinical applications. Analysts tracking industry activity count close to one hundred substantial AI–pharma partnerships signed in the last decade, with the pace rising in the past few years.¹²

Early proof points and case studies

The most convincing evidence that AI is changing drug R&D comes from real programs that have advanced into the clinic and, in some cases, to regulatory approval.

The rentosertib program from Insilico Medicine is one of the clearest examples. A Nature Medicine paper and accompanying company disclosures describe how a generative AI platform identified the fibrosis-linked TNIK pathway, designed novel small-molecule inhibitors and optimized a development candidate in a fraction of the usual time, moving from project inception to a clinical candidate in under eighteen months.³ Phase two results in idiopathic pulmonary fibrosis reported statistically significant improvements in lung function at the optimal dose with a manageable safety profile, positioning rentosertib as one of the first AI-discovered drugs to demonstrate clear clinical benefit in humans.³

The AlphaFold project from Google DeepMind illustrates AI’s impact on enabling science rather than individual products. In 2020 AlphaFold achieved near-experimental accuracy for protein structure prediction and has since provided structures for almost all catalogued proteins, dramatically accelerating structure-based research across therapeutic areas.⁴ Reviews in the scientific literature emphasize that AlphaFold has already influenced discovery efforts in areas such as malaria vaccines, oncology targets and enzyme design by making structural information available almost instantly where it previously took years of experimental work.⁵

BenevolentAI’s identification of baricitinib for COVID-19 created an early and visible proof of concept for AI-driven drug repurposing. Using a biomedical knowledge graph and AI reasoning over approved drugs and disease mechanisms, the company highlighted baricitinib as a promising candidate within days of the first reports of the virus, targeting both viral entry and inflammatory pathways.⁹ Subsequent clinical trials demonstrated a statistically significant reduction in mortality for hospitalized patients, and regulators converted emergency authorization into full approval, with BenevolentAI’s role in the identification acknowledged in scientific and regulatory commentary.⁹

A growing body of analysis now examines the clinical performance of AI-discovered molecules more systematically. One recent review of AI-native biotech pipelines found that molecules discovered with AI support achieved an estimated 80 to 90 percent success rate in phase one trials, significantly higher than historical averages for traditionally discovered drugs.¹³ Commentary from oncology groups and professional associations echoes this finding, noting that a material share of AI discovery effort is directed at cancer indications and that early-phase outcomes appear encouraging.¹⁴

Impact on R&D timelines, costs and success rates

Drug development economics have been under pressure for years as costs rose and success rates stagnated. AI offers levers to improve each side of this equation. At the front of the pipeline, better target identification and in silico screening reduce the number of dead-end programs that fail late and expensively. During lead optimization and preclinical testing, AI models help teams prioritize experiments and predict safety liabilities earlier. In the clinic, digital tools improve trial design, site selection and patient recruitment.

Analysts estimate that across discovery and preclinical stages, well-executed AI programs could cut time and cost by double-digit percentages, in some scenarios close to half, by reducing cycles of trial-and-error and narrowing the set of compounds that must be synthesized and tested in vivo.² For development, digital twin control arms and AI-assisted trial optimization can shorten study durations by months and allow more participants to receive active treatment rather than placebo, which improves both patient experience and sponsor economics.⁶⁷

The most powerful effect may come from improving success rates. Industry-wide data historically show that only a small fraction of candidates entering human trials eventually reach approval. By using AI to prioritize more biologically plausible targets and to eliminate compounds with poor predicted safety or pharmacokinetics before they reach the clinic, companies can raise the quality of the portfolio. The phase one success rates reported for AI-discovered molecules, at roughly 80 to 90 percent, compare favorably with traditional benchmarks and suggest that AI-based selection is beginning to influence downstream attrition.¹³ A Deloitte analysis projects that if AI technologies are scaled effectively, industry-wide R&D return on investment could recover and rise through 2030 after a decade of decline.²

Market growth and investment dynamics

The business environment around AI in drug discovery has matured rapidly. Only a few years ago, AI-drug discovery was a niche segment; it is now a multi-billion-dollar global market that continues to grow at high double-digit rates.

Market research from Grand View Research estimates that the global artificial intelligence in drug discovery market was approximately 1.5 billion US dollars in 2023 and could reach more than 20 billion US dollars by 2030, implying a compound annual growth rate close to thirty percent.¹⁵ In the United States alone, the segment is expected to grow from roughly 0.8 billion US dollars in 2023 to well over 4 billion US dollars by the end of the decade.¹⁶ Those figures exclude broader AI investments in life sciences data infrastructure and analytics, which further amplify demand for specialized discovery platforms.

Venture and growth investors have financed a substantial wave of AI-native biotechs and platform companies, with individual financings in the hundreds of millions of dollars and several high-profile initial public offerings over the past five years. Strategic investors from the pharmaceutical and technology sectors have participated alongside financial investors. At the same time, large pharmas have committed multi-billion-dollar budgets through discovery collaborations, as illustrated by Sanofi’s alliance with Exscientia and Roche’s collaboration with Recursion.¹⁰⁸ These commitments signal that AI is now regarded as a core enabler of future pipeline growth rather than a speculative technology experiment.

Regulatory and ethical considerations

As AI moves closer to the center of decision-making in discovery and development, regulators and ethics bodies are placing greater emphasis on transparency, validation and patient protection. The leading regulators are not trying to slow AI adoption; instead, they are working to ensure that AI-derived evidence meets the same standards of rigor as traditional methods.

In early 2025 the US Food and Drug Administration released a draft guidance titled “Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products,” which lays out expectations for how sponsors should document and validate AI tools used to generate data for submissions.¹⁷ The FDA’s Center for Drug Evaluation and Research has also published a broader AI strategy that covers drug development, manufacturing and pharmacovigilance and has created internal structures to coordinate AI policy and technical expertise.¹⁸ European regulators and international consortia are issuing complementary guidance, including the qualification of digital twin methods for use in certain trials.

Ethical debates focus on algorithmic bias, explainability, data privacy and accountability. AI models trained on non-representative datasets risk encoding biases that could overlook the needs of under-served populations or misestimate risks in certain demographic groups. Reviews in the regulatory and bioethics literature argue for more representative training data, systematic bias audits and human-in-the-loop oversight for high-impact decisions.¹⁹ Privacy regulations such as GDPR and HIPAA require careful handling of patient-level data used to train models, which has encouraged the use of approaches such as federated learning and advanced anonymization techniques.

Transparency is becoming a practical requirement. Sponsors are expected to be able to explain how critical AI systems work at a conceptual level, describe their training data and demonstrate performance across relevant subgroups. Many organizations address this by publishing methods, participating in external benchmarks and building internal model governance processes modeled on more traditional validation frameworks. Regulators have signaled that AI should augment rather than replace human judgment, with clear lines of accountability for decisions that affect patient safety.¹⁷¹⁹

Pharma partnerships and collaboration models

Partnership structures have become the primary way that large pharmaceutical companies access cutting-edge AI capabilities while AI-native firms gain access to data, therapeutic expertise and late-stage development capabilities. These alliances take a variety of forms, from target-specific discovery projects to multi-year platform collaborations that span entire therapeutic areas.

The collaboration between Sanofi and Exscientia is often cited as a landmark in this space. Sanofi agreed to pay 100 million US dollars upfront and up to 5.2 billion US dollars in milestones for access to Exscientia’s end-to-end AI platform to design up to fifteen novel small-molecule candidates in oncology and immunology, with options for co-investment and royalties.¹⁰ The scale of the agreement reflects a strategic choice to embed AI systematically rather than trial it on isolated pilot projects.

In another example, Roche and its subsidiary Genentech formed a broad discovery collaboration with Recursion with an upfront payment of 150 million US dollars and the potential for billions in milestones across as many as forty programs in neuroscience and oncology.⁸ The partners are using Recursion’s high-content imaging and machine learning platform to explore disease biology and identify targets that might not emerge through conventional approaches.

Similar alliances connect many other AI platforms with pharmaceutical companies. Atomwise has signed multi-target deals with several global pharmas for structure-based virtual screening.²⁰ Generate Biomedicines collaborates with large companies to apply generative biology to protein therapeutics.²¹ Valo Health has partnered with Novo Nordisk on cardiometabolic disease, with potential milestone payments measured in billions of dollars as the collaboration expands.²² Meanwhile Unlearn partners with sponsors across neurology and other areas to incorporate digital twins into ongoing clinical trials.⁶

Alongside one-to-one deals, consortia models are also emerging to pool data and expertise without compromising intellectual property. Some initiatives use federated learning so that AI models can learn from distributed datasets across multiple companies while data remain behind each organization’s firewall. This collaborative infrastructure can help mitigate data fragmentation and enable more robust models while respecting commercial and regulatory constraints.

Future outlook and strategic implications

The next decade is likely to see AI move from a set of powerful tools to an organizing principle for how discovery and development are planned and executed. Forecasts from consulting and industry groups suggest that by 2030 AI will be deeply embedded in target identification, lead optimization, trial design and real-world evidence generation, supporting a more continuous and data-driven approach to R&D.²

For pharmaceutical and biotech leaders, the implications are strategic. AI capabilities will increasingly differentiate organizations on the dimensions that matter most: the speed with which they can move from concept to clinic, the quality and diversity of their pipelines and the economics of their R&D portfolios. Companies that scale AI thoughtfully, invest in the right talent and data foundations and build strong external partnerships are positioned to benefit from rising R&D productivity, stronger competitive moats and more resilient portfolios.

Successful strategies will treat AI as a core part of the innovation system rather than an overlay. That means aligning AI initiatives with therapeutic area strategies, ensuring that discovery and development teams co-design use cases with data scientists and embedding responsible-AI principles into day-to-day operations. It also means revisiting governance and risk management to address issues such as model validation, bias monitoring and compliance with evolving regulations.

Most importantly, the industry will be judged by how AI-enabled innovation translates into better outcomes for patients. Shorter timelines matter because they bring needed therapies to people faster. Higher success rates matter because they increase the odds that truly novel mechanisms reach the market. When AI is combined with strong scientific judgment, rigorous validation and a patient-centric mindset, it can help deliver on the long-stated ambition of more precise, effective and accessible medicines.