The real value of AI drug discovery lies in improving the probability of success, not merely reducing time and cost

Overview

The report responds to leading AI companies' optimistic claims that “most diseases will be cured within the next 5–10 years” by systematically reviewing actual progress in AI across drug discovery, clinical development, and regulatory approval. Bernstein concludes that AI is becoming a decision-support and automation layer for scientists, and that its greatest economic value may come from selecting better programs, avoiding costly failures, and addressing diseases that are difficult to tackle using traditional methods, rather than becoming a standalone, end-to-end pharmaceutical engine.

Core views

The report first distinguishes between the grand narratives of leading AI companies and the more pragmatic objectives of large pharmaceutical companies. Technology companies discuss using AI to cure most or even all diseases within the next 5–10 years, while pharmaceutical management teams disclose more specific operational and R&D metrics. In the second quarter of 2026, AI was mentioned a record 29 times on earnings calls among companies covered by Bernstein. Amgen stated that antibody lead optimization had become 50% faster, trial enrollment efficiency had increased by as much as threefold, and line-clearance time per batch on one production line had fallen from 30 minutes to 2 minutes. Pfizer stated that virtual screening accelerated computational chemistry for the Paxlovid program by 5–10 times and reduced computing time by 80%–90%, while AI-enabled manufacturing processes increased throughput by 20%. Bristol Myers Squibb aims to accelerate lead molecule identification by approximately 50% and shorten late-stage development cycles by 30% compared with several years ago. These results demonstrate that AI can improve specific steps, but they are not yet sufficient to support the systemic conclusion that it can rapidly “cure all diseases.” The fundamental constraint begins with humanity's limited understanding of disease biology. The report notes that only approximately 9% of the human proteome has currently been exploited for drug development, while roughly 91% remains partially characterized or completely unknown. Of the approximately 15,000–20,000 known diseases, only about 15% have FDA-approved therapies, and the proportion with therapies that truly modify disease progression is even lower. Many diseases lack reliable experimental models, and diseases often simultaneously involve multiple organs, aging, environmental factors, and interacting molecular pathways. AI can search for patterns in complex data, but the quality of its output remains constrained by foundational biological knowledge and the completeness of training data. The second constraint is proprietary data, expertise, and organizational culture. Data on interactions among drugs, proteins, and diseases are primarily retained by large pharmaceutical companies as trade secrets, and no large-scale open training corpus comparable to internet text exists. Effective use of AI also requires expertise in chemistry, biology, translational medicine, clinical development, and regulatory science; correctly selecting the problem itself is a core capability. The report therefore favors partnerships between AI companies and traditional pharmaceutical companies over direct disruption of incumbent pharmaceutical companies by AI companies. However, large pharmaceutical companies also face the risk of falling behind culturally because of legacy processes, infrastructure, and power structures. AI-native biotech companies benefit from organizational structures better suited to AI, but their breadth of expertise and depth of data are generally inferior to those of large pharmaceutical companies. Clinical evidence is the third unavoidable hurdle. Clinical trials are expensive, time-consuming, and prone to failure, with pivotal trials typically requiring hundreds of millions of dollars annually. AI can improve protocol design, site selection, patient identification, safety monitoring, and trial operations, but it cannot compress the natural course of a disease or eliminate the safety and efficacy evidence required by regulators, physicians, and patients. Even if AI significantly increases the number of drug candidates, trial funding and infrastructure, regulatory review resources, manufacturing capacity, healthcare staffing, and payer budgets cannot expand without limit. Consequently, a breakthrough in one R&D step will not automatically translate into large-scale patient benefits. The report argues that AI's most important economic role is to improve decision quality and the probability of R&D success, rather than simply reduce costs or shorten timelines. Its scenario analysis shows that a 20% quality improvement, measured by fewer failures or a higher probability of success, can create substantially more value than an equivalent improvement in speed or cost, with most of the value concentrated in the later stages of R&D. Traditional innovative drugs typically take 10–15 years from target hypothesis to launch, and after accounting for failed programs, the total cost per approved drug is approximately USD 2–3B. Therefore, terminating poor programs earlier and identifying more differentiated assets can be especially effective in avoiding costly late-stage failures. However, AI's effect on the probability of success cannot be separated from the difficulty of the problems being addressed. If AI is initially applied to diseases with well-understood biological mechanisms, better molecular design and development optimization may increase early success rates, amplifying market perceptions of AI's value. If AI helps the industry enter areas where biology is poorly understood or that have historically been undruggable or difficult to treat, overall success rates may remain low, but the value of solving these problems would instead be greater. Therefore, a lack of visible improvement in the industry's overall success rate does not necessarily mean that AI has failed to create value. The report also emphasizes that difficult problems lack training data. As AI moves deeper into these areas, it will increasingly require wet-lab experiments to generate new data. AI and laboratory biology are more likely to expand in parallel than replace each other, because conducting one additional wet-lab experiment is far cheaper than suffering one late-stage failure. Across the R&D process, AI already covers the four stages of discovery, preclinical development, clinical development, and approval, but the maturity of evidence varies. The most mature applications are concentrated in virtual screening, hit identification, and lead optimization, including Schrodinger's FEP+ physics-based simulations and generative chemistry tools. Insilico's PandaOmics platform discovered TNIK as a target for idiopathic pulmonary fibrosis and generated Rentosertib, which has entered Phase III clinical trials. At the preclinical stage, machine-learning ADMET models assess CYP450 metabolism, hERG cardiotoxicity, solubility, and oral bioavailability before synthesis. Accuracy on common endpoints in public benchmarks is approximately 80%–85%, but it remains unproven whether AI-supported preclinical work produces better clinical outcomes. The FDA Modernization Act 2.0 removed the legal requirement for animal testing before an IND in December 2022, and the FDA roadmap released in April 2025 further promoted the use of AI toxicity prediction, organ-on-a-chip systems, and organoid methods for monoclonal antibodies. Clinical-stage applications include adaptive dose exploration, patient stratification, electronic health record screening for enrollment, dropout prediction, and continuous safety monitoring. Approval-stage applications include large language model-assisted regulatory document drafting, cross-checking of CMC and labeling, and real-world evidence processing. The report's overall position is that AI should currently be viewed as a decision-support and automation layer surrounding scientists, rather than an end-to-end drug discovery engine that requires no scientists, experiments, or clinical validation. The business models of leading AI companies fall into enablement, direct drug development, and hybrid approaches. Through GPT-Rosalind, OpenAI provides scientific reasoning tools to companies including Amgen, Novo Nordisk, and Moderna, but has no proprietary drug pipeline. Alphabet established Isomorphic Labs in 2021 to design drugs using IsoDDE and has partnered with Lilly, Novartis, and Johnson & Johnson. Isomorphic Labs raised USD 2.1B in May 2026, making it the most extensive example of a leading AI company directly participating in drug creation. Anthropic lies between the two approaches: it acquired Coefficient Bio for a reported USD 400M in stock in April 2026 and announced the launch of a small number of early-stage drug programs in June of the same year, but has not yet disclosed a specific pipeline or AI drug discovery product. Large pharmaceutical companies have moved from “whether to adopt AI” to “how to combine capabilities.” Bernstein identified 183 drug discovery partnerships across 15 companies and believes that the number of deals, therapeutic-area priorities, and balance among external partnerships, internal models, and computing infrastructure are more indicative of strategy than nominal total deal value, because so-called “biobucks” often mix acquisitions, asset licensing, and clinical-stage transactions. Lilly has the broadest external portfolio, with 25 partnerships. Its disclosed aggregate upfront consideration totals USD 3.56B, but USD 3.2B of this came from the Morphic acquisition. Excluding that transaction, disclosed upfront consideration for the remaining 24 partnerships totals USD 353M. Lilly also partners with OpenAI and Isomorphic Labs and is building internal AI infrastructure based on NVIDIA technology. Roche has 20 partnerships, including 8 focused on oncology and 4 on neuroscience. Its partnership with Recursion includes an upfront payment of USD 150M and potential value of up to USD 12B. Roche's disclosed aggregate upfront consideration totals USD 1.906B, but USD 1.45B of this came from the Nurix clinical co-development and PathAI digital pathology transactions rather than discovery-stage partnerships. Amgen places greater emphasis on proprietary data, foundation models, deCODE genetics, and internal capabilities. Bernstein believes Lilly, Roche, and Amgen are currently among the leaders in large pharma, but cautions that this is only an external assessment, as the true quality of partnerships, internal integration, and innovation output remains difficult to determine. AI has already created a development pipeline of meaningful scale. The report identifies approximately 70 active AI-originated programs, primarily concentrated among leading AI-native biotech companies, with the vast majority of key validations expected over the next 3–5 years. Within its stock coverage, Bernstein rates LLY and GILD Outperform and BMY, PFE, MRK, ABBV, AMGN, and MRNA Market-Perform. The report does not suggest that operating-cost savings alone are sufficient to change the investment thesis for the pharmaceutical industry as a whole or for any individual company. What could truly change long-term value is the quality of innovation and future revenue potential.

Analysis framework

The report first compares the ambitious goals of leading AI companies with the quantifiable results disclosed by pharmaceutical companies, then identifies bottlenecks layer by layer across disease understanding, intervention design, clinical evidence generation, and healthcare system capacity. It subsequently compares the economic value of improvements in R&D probability of success, time, and cost through scenario analysis and assesses the maturity of AI applications across the four stages of discovery, preclinical development, clinical development, and approval. Finally, the report analyzes 183 partnerships across 15 pharmaceutical companies, comparing external partnerships, internal models, proprietary data, and computing infrastructure, and uses approximately 70 AI-originated programs to assess the path to future validation.

Methodology notes

• Competition and Strategy FrameworkValue chain analysis

  Breaking down AI applications by drug R&D stage

  The report divides the R&D chain into discovery, preclinical development, clinical development, and approval, assessing at each stage the problems AI can solve, the maturity of its applications, and the limitations that still require validation through human-led experiments or clinical evidence.

• Company Fundamentals and Financial Framework

  R&D probability-of-success sensitivity scenarios

  The report compares the effect on R&D returns when probability of success, speed, and cost each improve by 20%, demonstrating that reducing failures and improving program quality create greater economic value than merely accelerating development or reducing costs.

• Competition and Strategy FrameworkMoat / competitive advantage

  Comparison of data, expertise, organizational culture, and computing capabilities

  The report evaluates the relative advantages of large pharmaceutical companies, AI-native biotech companies, and leading AI companies based on proprietary experimental data, interdisciplinary expertise, internal models, computing infrastructure, and organizational adoption capabilities.

• Industry Analysis Framework

  Partnership portfolio and capability structure analysis

  The report counts 183 partnerships across 15 companies and argues that partnership numbers, therapeutic areas, and the mix of internal and external capabilities better reflect AI drug discovery strategies than total potential deal value.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Eli Lilly (LLY)

  Bernstein rates it Outperform and views it as one of the most advanced large pharmaceutical companies in AI drug discovery.

  Strengths

  It has the broadest external partnership portfolio, with 25 partnerships, while also working with OpenAI and Isomorphic Labs and developing NVIDIA-based internal computing infrastructure.

  Weaknesses

  Many partnerships have not disclosed specific therapeutic areas, and the aggregate upfront consideration is primarily attributable to the USD 3.2B Morphic acquisition.

  Comparison

  Its breadth of external partnerships leads the other large pharmaceutical companies in the sample.

  Risks

  The quality of external partnerships, degree of internal integration, and ultimate drug output remain difficult to assess externally.

• Roche (ROG)

  The report identifies it as one of the leaders in AI drug discovery among large pharmaceutical companies.

  Strengths

  It has 20 partnerships, with relatively concentrated therapeutic areas, and combines partnerships at scale with substantial computing infrastructure.

  Weaknesses

  USD 1.45B of its disclosed upfront consideration came from non-discovery-stage transactions, meaning nominal deal value may overstate its drug discovery investment.

  Comparison

  Its number of partnerships is second only to Lilly's, with a greater concentration in oncology and neuroscience.

  Risks

  Partnership valuation metrics are complex, and the true quality and output of R&D integration remain to be validated.

• Amgen (AMGN)

  Bernstein rates it Market-Perform and considers its AI strategy relatively advanced among large pharmaceutical companies.

  Strengths

  It places greater emphasis on proprietary data, foundation models, deCODE genetics, and internal capabilities; disclosed achievements include a 50% acceleration in antibody optimization and as much as a threefold increase in clinical enrollment.

  Weaknesses

  Its external partnership breadth is less extensive than Lilly's and Roche's.

  Comparison

  Compared with Lilly's broad external partnerships, Amgen is more oriented toward internal models and a data-driven approach.

  Risks

  It remains unproven whether operational acceleration can translate into higher clinical success rates and new-drug revenue.

• Alphabet (GOOGL) and Isomorphic Labs

  The report considers Alphabet the most advanced example of a leading AI company moving from tool enablement to direct participation in drug creation.

  Strengths

  Its products span research assistants, scientific tools, and drug design. Isomorphic Labs has IsoDDE and partnerships with Lilly, Novartis, and Johnson & Johnson.

  Weaknesses

  Direct participation in drug development means confronting traditional pharmaceutical constraints such as clinical validation, funding, and regulation.

  Comparison

  It participates more deeply in the direct creation of drug-related economic value than OpenAI and Anthropic.

  Risks

  Whether its platform capabilities can translate into clinical success and commercialized drugs remains to be validated.

• OpenAI

  It primarily participates in drug discovery as a provider of AI tools to pharmaceutical and biotech companies.

  Strengths

  GPT-Rosalind can support evidence review, data analysis, hypothesis generation, and experimental planning, and it partners with Amgen, Novo Nordisk, and Moderna.

  Weaknesses

  It has no proprietary drug pipeline and therefore cannot directly capture the full economic value of drug development.

  Comparison

  Compared with Alphabet, it is more focused on the enablement layer rather than direct drug development.

  Risks

  It remains unclear whether efficiency gains from scientific tools can improve the clinical probability of drug success.

• Anthropic and Coefficient Bio

  It follows a hybrid approach between tool enablement and direct drug development.

  Strengths

  It acquired Coefficient Bio in 2026 and announced the launch of a small number of early-stage drug programs to build first-hand experience using Claude in real-world drug discovery.

  Weaknesses

  It has not yet disclosed a specific drug pipeline or AI drug discovery product.

  Comparison

  Its level of direct participation is higher than OpenAI's but lower than Alphabet's through Isomorphic Labs.

  Risks

  The programs remain at an early stage and lack verifiable clinical and product evidence.

Key data

• Extent of human proteome exploitationApproximately 9% has been exploited, while roughly 91% remains partially characterized or completely unknownReflects that foundational biological knowledge remains a major constraint on AI drug discovery.
• Proportion of diseases with FDA-approved therapiesApproximately 15%Based on approximately 15,000–20,000 known diseases; the proportion with disease-modifying therapies is even lower.
• Typical innovative-drug R&D cycleApproximately 10–15 yearsTotal duration from target hypothesis to commercial launch.
• Total cost per approved drugApproximately USD 2–3BIncludes the cost of failed programs.
• Quality improvement scenario20%The report argues that the value created by a 20% increase in probability of success or reduction in failures is far greater than that from an equivalent improvement in time or cost.
• Active AI-originated programsApproximately 70Most key validations are expected over the next 3–5 years.
• Large pharma AI partnership sample15 companies and 183 partnershipsUsed to compare strategies involving external partnerships, internal models, computing power, and proprietary data.
• Lilly partnership portfolio25 partnershipsDisclosed aggregate upfront consideration totals USD 3.56B, of which USD 3.2B came from the Morphic acquisition; excluding it, the remaining 24 partnerships total USD 353M.
• Roche partnership portfolio20 partnershipsOf these, 8 are in oncology and 4 in neuroscience; USD 1.45B of the disclosed USD 1.906B in upfront consideration was unrelated to discovery-stage partnerships.
• Machine-learning ADMET accuracyApproximately 80%–85%Prediction accuracy for common preclinical safety and pharmacokinetic endpoints in public benchmarks.
• AI mentions on pharmaceutical earnings calls29 times in the second quarter of 2026A record within Bernstein's coverage universe.
• Isomorphic Labs financingUSD 2.1BCompleted in May 2026 to expand the platform and advance the pipeline into clinical development.

Impact & implications

The report argues that operating efficiencies generated by AI are not sufficient by themselves to change the investment thesis for the pharmaceutical industry or an individual company. What truly matters is whether AI can improve R&D program quality, reduce late-stage failures, and generate differentiated new-drug revenue. The proprietary data and development experience of large pharmaceutical companies remain important barriers, but organizational culture and internal integration capabilities will determine whether these assets can be converted into AI productivity. AI-native companies may develop advantages in specific difficult problems, but their ultimate value must still be validated through wet-lab experiments and clinical trials.

Risks

• Incomplete understanding of disease biology and protein function may limit the reliability of AI training data and conclusions.
• Drug discovery data are highly proprietary, and AI companies may struggle to replicate the data barriers and accumulated long-term expertise of large pharmaceutical companies.
• AI cannot eliminate the natural observation time, safety and efficacy evidence, or low R&D probability of success associated with clinical trials.
• Trial funding, regulatory review, manufacturing, healthcare staffing, and payer budgets may be unable to absorb the large number of drug candidates generated by AI.
• Large pharmaceutical companies may become laggards in AI adoption because of legacy processes and organizational culture.
• Partnership transaction values often mix acquisitions, asset licensing, and clinical-stage transactions, potentially distorting assessments of AI drug discovery investment intensity.
• AI-supported preclinical predictions and operational efficiencies have not yet been proven to systematically improve clinical-stage outcomes.

What to watch

• Whether approximately 70 active AI-originated programs can produce key clinical validation over the next 3–5 years.
• Whether AI can improve the probability of R&D success, rather than merely shorten individual processes or reduce operating costs.
• Whether Lilly, Roche, and Amgen can disclose actual innovation and pipeline output beyond speed-related metrics.
• Whether wet-lab investment increases in parallel as AI enters undruggable areas and fields with limited biological understanding.
• Actual adoption progress for the FDA's roadmap on non-animal testing methods, AI toxicity prediction, organ-on-a-chip systems, and organoids.
• The evolution of OpenAI's, Alphabet's, and Anthropic's business models between tool enablement and directly capturing the economic value of drugs.