AI in Pharmaceutical R&D: Evidence from Drug Discovery to Approval

Introduction

On June 3, 2025, Nature Medicine published a Phase IIa trial result that had been anticipated across AI drug discovery for years. Rentosertib (formerly ISM001-055), a TNIK inhibitor whose target and molecule were both identified with AI assistance, was evaluated in 71 IPF patients across 21 sites in China in a randomized, double-blind, placebo-controlled, 12-week trial. The primary endpoint was safety: the rate of treatment-emergent adverse events, which was comparable across all arms [1]. Among the secondary endpoints, the 60 mg once-daily group (n=18) showed a mean forced vital capacity (FVC) change of +98.4 mL (95% CI: 10.9 to 185.9) over 12 weeks, compared to a decline of 20.3 mL (95% CI: -116.1 to 75.6) in the placebo group (n=17) [1].

The paper's conclusions are appropriately measured. The authors describe an exploratory FVC signal that "warrants further investigation in larger-scale clinical trials of longer duration" [1]. The trial was not powered to establish efficacy. What it does represent is the first peer-reviewed Phase IIa result for a drug whose target and molecular structure were both found using generative AI, a meaningful milestone that the field had not previously cleared.

This result sits inside a broader transformation. AI is now embedded across the pharmaceutical value chain: from protein structure prediction and early discovery, through clinical operations and site selection, to regulatory submissions and supply chain management. Understanding where the evidence is strong, where it is early-stage, and where oversight frameworks are still developing is essential for anyone working in drug development today.

Evidence Maturity at a Glance

The table below separates applications where peer-reviewed or regulatory evidence is strong from those where evidence is still accumulating or primarily based on vendor and trade reporting.

Evidence status definitions: Established = strong peer-reviewed primary evidence with substantial external validation and adoption; formal recognition (such as a Nobel Prize) may serve as supporting context. Emerging = peer-reviewed data exists but is preliminary, context-specific, or aggregate across heterogeneous interventions. Early-stage = mechanistic logic and operational adoption documented; controlled peer-reviewed evidence limited or absent.

The Scale of AI Adoption in Pharma: What the Numbers Show

Before examining specific domains, it is worth distinguishing between market projections and operational performance data, since these are routinely conflated in AI-pharma coverage.

On the market side, the AI-enabled drug discovery and clinical trials market is projected to exceed $3 billion in global revenue in 2025, with sustained growth forecast through 2035 [2]. McKinsey's 2025 analysis estimated that AI could generate $60 to $110 billion in annual value across the pharmaceutical sector, primarily through R&D acceleration and commercial operations optimization [3]. These projections reflect investment appetite, not confirmed operational gains.

On the activity side, the data is more concrete. Applications incorporating AI or machine learning (ML) components submitted to the FDA grew from 29 in 2019 to 132 in 2021 [4], and the agency reported reviewing more than 500 such submissions between 2016 and 2023 [5]. Academic output has tracked alongside: clinical trial AI publications increased by 444% from 2019 to 2024, at a compound annual growth rate of 40% [4]. As of April 2024, 31 drugs developed by AI-discovery companies were in human clinical trials across eight companies, with nine in Phase II or III [6].

These figures reflect sustained institutional commitment. They also bracket the honest operational question: how much of that investment is producing demonstrable results?

Drug Discovery: Where the Most Documented Progress Has Emerged

AlphaFold Changes the Starting Conditions

For target identification and structural biology, the single most consequential AI development to date is AlphaFold2. In October 2024, the Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper of Google DeepMind for developing AlphaFold2, and to David Baker of the University of Washington for his work in computational protein design [7]. The AlphaFold2 system, described in a 2021 Nature paper demonstrating highly accurate protein structure prediction from amino acid sequence alone, predicted the structure of over 200 million proteins and attracted more than three million users across 190 countries by 2025 [8, 29]. Before it, determining a protein's three-dimensional structure required months of experimental work with X-ray crystallography or cryo-electron microscopy. AlphaFold shifted that bottleneck: researchers who previously spent months acquiring structural data can now access predictions in hours and redirect effort toward downstream validation.

Isomorphic Labs, the DeepMind spin-out applying Isomorphic's proprietary AI platform to drug design, formed collaborations with Eli Lilly and Novartis announced in January 2024, involving AI-driven de novo small-molecule discovery for targets selected by those companies; Lilly committed $45 million upfront with up to $1.7 billion in performance milestones, and Novartis committed $37.5 million upfront [28]. The Novartis collaboration was expanded in February 2025 to add additional research programs [9]. These commitments signal pharmaceutical companies' confidence in AI-assisted structural biology for early-stage discovery.

What AI Has and Has Not Changed in Discovery

An honest accounting published in Drug Target Review in February 2026 describes the state of AI drug discovery after 2025: AI can compress early discovery timelines by 30 to 40% and reduce preclinical candidate development to 13 to 18 months versus the traditional three to four years [10]. That is a real operational gain. But the same analysis notes that "AI has not demonstrably improved the pharmaceutical industry's approximately 90% clinical failure rate. While AI can compress early discovery timelines, clinical trial duration, regulatory review timelines and manufacturing scale-up remain bound by biology, patient enrolment and regulatory requirements, none of which AI can bypass" [10].

Multiple AI-assisted drug programs were deprioritized, shelved, or showed no efficacy signal after Phase II in 2025, alongside the Rentosertib result [10]. The clinical attrition pattern looks broadly similar to conventionally developed compounds at the same stage. AI accelerates the early phases; it has not yet changed what ends most programs in the clinic.

Clinical Trial Operations: Evidence Across the Lifecycle

A narrative review by Olawade et al., published online in October 2025 and assigned to International Journal of Medical Informatics volume 206 (2026), synthesized studies from PubMed, Embase, IEEE Xplore, and Google Scholar covering January 2015 to December 2024 [11]. The review reported that AI patient recruitment tools improved enrollment rates by 65% across the included studies, predictive analytics models achieved 85% accuracy in forecasting trial outcomes, and AI integration accelerated trial timelines by 30 to 50% while reducing costs by up to 40% [11]. Digital biomarkers integrated into monitoring systems achieved 90% sensitivity for adverse event detection in the reviewed literature [11].

These figures represent aggregate trends across heterogeneous study designs and contexts. The authors appropriately note that implementation barriers include regulatory uncertainty and algorithmic bias [11]. Reported effects should not be assumed to generalize uniformly across therapeutic areas, trial phases, or implementation maturity levels without local validation.

Enrollment: A Structural Problem AI Is Beginning to Address

Enrollment delays are among the most reliably documented sources of trial cost and timeline overrun. Tufts CSDD research examining over 15,000 investigative sites across Phase II and III global trials found that 41% of activated sites failed to achieve their target enrollment numbers [12]. A 2020 Tufts CSDD study of 87 trials found that enrollment was completed within its planned timeline in 77% of studies, up from 48% in 2012 [13], though significant variability persists across therapeutic areas and geographies.

AI addresses enrollment challenges through two distinct mechanisms. First, predictive site selection tools analyze historical site performance, geographic patient demographics, and investigator experience to rank and select sites before activation, reducing the risk of activating underperforming sites. The AI-powered clinical trial site feasibility market was valued at $1.24 billion in 2024, with a compound annual growth rate of 23.8% and projections reaching $3.55 billion by 2029 [14]. Second, AI-driven patient matching tools apply natural language processing to parse electronic health record data against eligibility criteria, identifying potential participants who would not otherwise present through conventional screening. These tools can reduce screening failures and accelerate pre-enrollment identification, though their performance depends heavily on EHR data completeness and local validation against the specific eligibility framework.

Protocol Design and Amendment Prevention

Protocol complexity is a documented driver of delays and cost escalation. Tufts CSDD research covering thousands of protocols showed a near-doubling of trial endpoints and tripling of data points collected between 2010 and 2020 [15]. AI tools now applied to protocol design can flag inclusion/exclusion criteria that have historically generated high amendment rates across similar indications, simulate enrollment projections under different eligibility configurations, and benchmark protocol complexity against historical comparators.

ICH E6(R3), finalized at Step 4 on January 6, 2025, articulates stronger expectations for protocol quality, risk-proportionate monitoring, and built-in quality from the outset [16]. AI-assisted protocol review tools are increasingly positioned as one mechanism for building that quality in before site activation, rather than correcting it through costly post-activation amendments. Any such tool would itself require validation for the specific protocol type and therapeutic area in which it is applied.

Kitsa's KScout platform applies predictive analytics to site selection, drawing on site performance patterns most predictive of successful activation and enrollment. KScreener connects to EHR and FHIR pipelines to support patient pre-screening against protocol eligibility criteria, reducing the manual screening burden at the site level.

AI at the Regulatory Interface: Policy Is Moving on Both Sides

The FDA's Draft Guidance Framework

In January 2025, the FDA published draft guidance FDA-2024-D-4689, "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products" [5]. This is the FDA's first attempt to articulate a framework for how AI-generated data and information will be evaluated in regulatory submissions. Several features of the guidance deserve careful reading.

First, it is draft, not final, and its recommendations are nonbinding. It is not legislation or a final rule. Second, its stated scope is AI used "to produce information or data intended to support regulatory decision-making regarding safety, effectiveness, or quality for drugs" [5]. AI tools used in operational document drafting, project management, or internal workflows are outside its explicit scope. Sponsors should not apply this guidance as a universal regulatory requirement for all AI use across drug development.

Third, the guidance introduces a risk-based credibility assessment framework, recommending that sponsors document and justify AI model use in submissions where those models generate data or information supporting regulatory decisions. It was informed by more than 800 public comments on earlier discussion papers and the FDA's experience reviewing more than 500 AI-containing submissions from 2016 to 2023 [5].

The FDA's Internal AI Deployment: Elsa

On June 2, 2025, the FDA launched Elsa (Electronic Language System Assistant), an agency-wide generative AI tool for internal staff [17]. According to FDA and trade reporting, Elsa was initially built using Anthropic's Claude model within an Amazon Web Services FedRAMP-accredited GovCloud environment, and was designed to help reviewers with tasks including adverse event summarization, label comparisons, clinical protocol review, and internal database development [17]. FDA Commissioner Makary stated at launch that tasks previously requiring two to three days could be completed in six minutes [18].

Trade reporting indicates that early use of Elsa produced accuracy concerns among some agency reviewers, including instances of fabricated citations or misrepresented research [18]. FDA's Chief Artificial Intelligence Officer Jeremy Walsh stated in April 2026 that the agency's generative AI adoption had grown from roughly 1% of staff in early 2025 to over 80%, and that the FDA was using both Google's Gemini and Anthropic's Claude models across its operations at that time [19]. Following a February 27, 2026 Trump administration directive requiring federal agencies to cease using Anthropic's Claude, the FDA directed staff to transition Elsa's primary model to Google's Gemini; reporting from March 2026 described this as a politically forced and faster-than-planned migration, with Gemini set to become the primary Elsa model [20]. A Government Executive report from April 2026 indicated the FDA was using both Gemini and Claude across its operations during this period [19]. The practical implication for sponsors: FDA reviewers interact with AI tools in submission review workflows, and the tools and their underlying models are subject to change through administrative and policy channels.

The EMA's Lifecycle Framework

In September 2024, the EMA published its Reflection Paper on the Use of Artificial Intelligence in the Medicinal Product Lifecycle (EMA/CHMP/CVMP/83833/2023), adopted jointly by CHMP and CVMP [21]. This is a reflection paper, meaning it expresses EMA's current thinking and provides recommendations, not binding obligations. It covers AI use across drug discovery, clinical trials, manufacturing, and post-authorization activities, and emphasizes a risk-based, human-centered approach at each stage.

The EU AI Act entered into force on August 1, 2024, providing a binding legislative framework for AI across sectors. However, the Act is sector-agnostic and covers only a subset of AI/ML systems used in life sciences [22]. For pharmaceutical-specific AI applications, the EMA reflection paper provides the most relevant regulatory context in Europe, supplemented by existing GCP, GMP, and computerized system validation requirements (including Annex 11 and 21 CFR Part 11 in relevant jurisdictions).

A key principle running through both the FDA guidance and EMA paper is proportionality: the rigor of documentation, validation, and oversight expected for AI should be commensurate with the regulatory impact of the AI's output. Not all AI applications in drug development carry equal regulatory weight.

AI in Regulatory Document Generation: Capabilities and Limits

Generative AI applied to regulatory writing (CTD module drafting, Clinical Study Reports, Investigator Brochures, safety summaries, and consent form development) represents one of the most practically significant developments in medical writing over the past two years. The operational case is real.

Large language models trained on regulatory content can produce first drafts of Module 2 summaries, perform semantic consistency checks between Module 5 safety data and Module 2 summaries, flag eCTD hyperlink errors, and assist with global submission harmonization across regional formats [23]. Industry reporting cites up to 40% reduction in CSR authoring time for teams using AI writing tools within structured review workflows [24]; these figures come from consultancy and vendor sources rather than peer-reviewed studies, and should be read accordingly.

The limitations are equally important. Generative AI models can produce regulatory text that sounds authoritative but contains factual errors, misrepresents source data, or cites studies that do not exist. The FDA's own Elsa deployment encountered accuracy concerns of this type within its first weeks [18]. The Department of Health and Human Services' May 2025 public health report drew criticism for citing nonexistent studies, a consequence of unchecked AI-generated content in a high-stakes public document [18].

Neither FDA draft guidance FDA-2024-D-4689 nor the EMA reflection paper creates universal requirements for expert review of every AI-generated document in pharma operations. Where AI supports regulatory submissions or benefit-risk decisions, both frameworks recommend that organizations treat AI output as traceable, validated against the specific context of use, and reviewed by qualified experts, proportionate to the AI's regulatory role and not as a blanket requirement across all pharmaceutical AI use. For regulatory writing teams, the practical standard is: AI as a drafting tool within an expert-led review and verification workflow, not as a replacement for it.

Kitsa's KScribe platform applies AI to the generation of regulatory and clinical documents including protocols, Investigator Brochures, DSURs, CSRs, and Informed Consent Forms, with cross-document consistency checking built into the generation process to catch discrepancies that propagate across multi-document packages. The platform is designed to support expert review workflows, not to substitute for them.

Manufacturing and Supply Chain: Where AI Delivers More Immediate Results

AI's application in pharmaceutical manufacturing and supply chain logistics is advancing in specific documented contexts and carries less regulatory complexity than its clinical or submission counterparts. Most published figures on productivity gains come from vendor analyses, syndicated market content, or trade reporting rather than peer-reviewed pharmaceutical studies; the magnitudes cited vary considerably by process type, plant maturity, and AI implementation scope, and should be benchmarked against independent pharmaceutical evidence before being used in planning [25].

Predictive maintenance tools that model equipment sensor data to anticipate failure before it occurs can reduce unplanned downtime, one of the more reliably documented cost drivers in pharmaceutical production. Computer vision applied to visual inspection tasks can support consistency at throughput rates that challenge manual review; performance depends on training data quality, defect type, and validated operating parameters [27]. Demand forecasting tools that improve the alignment between production volumes and anticipated distribution needs can reduce waste from overproduction and expired inventory [26].

These manufacturing and supply chain applications generally do not require a separate AI-specific approval pathway, but they remain subject to applicable GMP validation, change-control, and data-integrity requirements depending on the system's function and the nature of the records it generates. Reduced pathway complexity relative to clinical AI applications is part of why adoption in manufacturing and supply chain contexts has progressed more quickly.

Regulatory and Documentation Considerations

Four overlapping regulatory frameworks are relevant to AI use in pharmaceutical development, and distinguishing them matters for compliance planning.

ICH E6(R3) (Step 4, January 6, 2025) is an ICH Good Clinical Practice guideline whose regulatory force depends on the specific mechanism each jurisdiction uses to implement it. Implementation instruments vary by country and may include binding regulations, guidance documents, or both; the resulting legal effect depends on the particular instrument applied and any underlying statutory framework. In the United States, for example, the FDA typically incorporates ICH guidelines through FDA guidance documents, which carry recommendations but are not themselves legally binding absent a separate statutory or regulatory requirement. Sponsors and investigators should verify the applicable requirements in each country where a trial is conducted. E6(R3) introduced a strengthened protocol quality framework and risk-proportionate monitoring expectations [16]. AI tools applied to protocol design or clinical monitoring must be fit for purpose within this framework.

FDA draft guidance FDA-2024-D-4689 (January 2025) is nonbinding draft guidance applicable to AI used to support regulatory decision-making. It recommends a risk-based credibility assessment framework and recommends that sponsors document AI model use and justify model fitness for the context of use in submissions [5]. It does not apply uniformly to all AI use in pharmaceutical operations.

EMA Reflection Paper EMA/CHMP/CVMP/83833/2023 (September 2024) is a nonbinding reflection paper covering AI across the medicinal product lifecycle. It recommends risk-proportionate validation, explainability commensurate with regulatory impact, and documented human oversight for high-impact applications [21]. It does not create uniform requirements across all AI deployments.

EU AI Act (entered into force August 2024) is binding legislation but sector-agnostic. Its applicability to specific pharmaceutical AI systems is use-case dependent. High-risk classifications under the Act apply to specific categories including AI in medical devices and safety-critical systems. Many pharma operational AI applications, such as manufacturing process control, demand forecasting, and document drafting tools, are unlikely to fall within the highest-risk tiers under the Act's current classification framework, but this determination requires system-by-system analysis and cannot be assumed in advance [22].

Alongside these, existing computerized system validation requirements (Annex 11 in the EU, 21 CFR Part 11 in the US) apply to AI systems used in GxP-regulated contexts where they generate or modify electronic records subject to regulatory review; applicability depends on the system's role and its handling of regulated data. GMP requirements for data integrity and GCP obligations for trial oversight similarly condition AI governance on the function the system performs. Sponsors should assess AI tool governance requirements under each applicable framework rather than assuming a single document covers all use cases.

What AI Has Not Yet Resolved: The Case for Measured Expectations

The pharmaceutical industry's approximately 90% clinical failure rate has not changed in the current generation of AI drug discovery [10]. Accelerating early discovery work does not compensate for inadequate translational biology, poorly selected targets, or patient populations where the drug simply does not work. The Rentosertib Phase IIa trial, while encouraging, was small and exploratory on efficacy measures; the biology will have to hold up in a powered Phase IIb. Most AI-derived candidates face the same challenge.

AI-assisted regulatory writing carries genuine accuracy risks that parallel what has been observed in broader generative AI deployments. Hallucination, where models produce plausible but incorrect information, is a documented failure mode that is particularly dangerous in regulatory documents where claims trace to specific studies and datasets. Expert review is not a bureaucratic formality; it is the control that makes AI-assisted drafting acceptable in high-stakes submissions.

Site selection models trained on historical data carry biases that reflect past trial conduct: therapeutic areas, geographies, and patient populations that were overrepresented in training data. Applying a site selection model to a novel therapeutic area or underrepresented geography without local validation is a meaningful risk, not a hypothetical one. EHR-based patient matching similarly depends on data completeness and coding consistency that varies substantially across sites and systems.

None of these limitations argue against AI adoption in drug development. They argue for implementing AI within governance frameworks that treat validation, transparency, and human oversight as design requirements rather than optional additions.

Key Takeaways

• Rentosertib, developed with AI-assisted target discovery and molecular design, showed an exploratory FVC signal in a small safety-primary Phase IIa trial published in Nature Medicine in June 2025; a powered Phase IIb trial is needed to evaluate efficacy [1].
• AlphaFold2, recognized with the 2024 Nobel Prize in Chemistry, has predicted over 200 million protein structures and is used by more than three million researchers across 190 countries, materially changing the starting conditions for structure-based drug discovery [7, 8].
• A narrative review by Olawade et al. (published online October 2025, International Journal of Medical Informatics 206, 2026) reported AI patient recruitment tools improved enrollment rates by 65% and AI integration reduced trial costs by up to 40% across the reviewed literature; these are aggregate findings from a narrative review and should not be assumed to generalize uniformly across all trial contexts [11].
• The FDA received more than 500 drug submissions containing AI components between 2016 and 2023, published its first draft guidance on AI in regulatory decision-making in January 2025, and deployed an internal AI review tool across the agency by June 2025 [5, 17].
• AI has not changed the pharmaceutical industry's approximately 90% clinical failure rate; clinical timelines remain governed by biology, patient enrollment, and regulatory requirements [10].
• FDA draft guidance and the EMA reflection paper establish a risk-based, proportionality framework for AI in regulated contexts, not universal requirements; the applicable framework depends on what the AI is doing and what regulatory impact its output carries [5, 21].
• AI in pharmaceutical manufacturing and supply chain operations is being applied to predictive maintenance, quality control, and demand forecasting; most published performance figures come from vendor or trade sources rather than peer-reviewed pharmaceutical studies, and gains vary by context [25, 26, 27].

Frequently Asked Questions

What is the current regulatory status of AI in pharmaceutical submissions?+

The FDA published draft guidance FDA-2024-D-4689 in January 2025, covering AI use to support regulatory decision-making for drug and biological products [5]. This is draft, nonbinding guidance. It recommends a risk-based credibility assessment framework and recommends that sponsors document and justify AI model use in submissions where those models generate data or information supporting regulatory decisions. The EMA's Reflection Paper (EMA/CHMP/CVMP/83833/2023, September 2024) covers a broader lifecycle scope and recommends risk-proportionate validation and human oversight for AI with significant regulatory impact [21]. Neither document creates universal requirements for all AI use in pharmaceutical operations; their application depends on context.

Has any AI-designed drug received regulatory approval?+

As of early 2026, no drug developed using AI-assisted discovery had received regulatory approval, based on available pipeline reporting [6, 10]. Rentosertib, developed with AI-assisted target and molecule design by Insilico Medicine, showed an exploratory secondary endpoint signal in a 71-patient Phase IIa safety trial published in Nature Medicine in June 2025, and is progressing toward Phase IIb [1]. As of April 2024, 31 AI-developed candidates were in human trials across eight companies, with none having reached approval at that time [6]. The term "AI-designed" lacks a stable industry definition; different programs describe varying degrees of AI involvement from target identification through lead optimization.

Does AI improve clinical trial success rates?+

Available evidence shows that AI accelerates early discovery timelines and improves certain trial efficiency metrics, but it has not demonstrably improved the pharmaceutical industry's approximately 90% clinical failure rate. Multiple AI-assisted drug programs were shelved after Phase II in 2025. Clinical failures reflect biology, patient population selection, and endpoint design; AI currently has limited influence over those factors [10].

What does AI do in clinical trial site selection?+

AI tools for site selection analyze historical site performance, patient recruitment track records, geographic disease burden, and investigator experience to rank sites before activation. The AI-powered clinical trial site feasibility market was valued at $1.24 billion in 2024 and is projected to reach $3.55 billion by 2029 [14]. These tools reduce the risk of activating underperforming sites, which is a primary driver of trial delays. Performance depends on data quality and local validation; models trained primarily on Phase III oncology trials may not transfer directly to rare disease or first-in-human programs.

How is the FDA using AI in its own review processes?+

The FDA launched Elsa, an internal generative AI assistant, on June 2, 2025, to support scientific reviewers with adverse event summarization, label comparisons, clinical protocol review, and database development [17]. FDA Commissioner Makary cited efficiency gains at launch, stating that tasks previously requiring two to three days could be completed in six minutes [18]. Early use surfaced accuracy concerns in some outputs [18]. FDA's Chief Artificial Intelligence Officer reported in April 2026 that agency-wide generative AI adoption had grown from roughly 1% to over 80%, with the FDA using both Claude and Gemini models across operations at that time [19]. Following a February 2026 federal directive to cease using Anthropic's Claude, the FDA began transitioning Elsa's primary model to Gemini; reporting from March 2026 described this as a politically forced migration with Gemini set to become the primary model [20].

What safeguards apply to AI-generated regulatory documents?+

Neither FDA draft guidance FDA-2024-D-4689 nor the EMA reflection paper creates a universal requirement for expert review of every AI-generated document in pharmaceutical operations. Where AI-generated content supports regulatory submissions or benefit-risk decisions, both frameworks recommend that organizations treat such content as traceable, validated for its specific context of use, and reviewed by qualified experts, proportionate to the regulatory role of the AI and not as a mandate across all use cases [5, 21]. For regulatory writing teams, this means treating AI as a drafting tool within a structured expert review workflow, with factual verification against primary sources before any AI-generated content enters a regulatory package.

References

1. [1] Zhavoronkov A et al. "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial." Nature Medicine. 2025;31:2602-2610. DOI: 10.1038/s41591-025-03743-2 https://www.nature.com/articles/s41591-025-03743-2
2. [2] Research and Markets / Globe Newswire. "AI-Enabled Drug Discovery and Clinical Trials Research Report 2025-2035." June 9, 2025. https://www.globenewswire.com/news-release/2025/06/09/3095666/28124/en/AI-Enabled-Drug-Discovery-and-Clinical-Trials-Research-Report-2025-2035
3. [3] McKinsey and Company. Referenced in Eularis. "The Future of AI and Innovation in Pharma in 2026 and Beyond." November 2025. https://eularis.com/the-future-of-ai-and-innovation-in-pharma-in-2026-and-beyond/
4. [4] Niazi SK. "A Critical Review of the FDA's Draft Guidance on Artificial Intelligence in Drug and Biological Product Regulation." Journal of Chemistry (Wiley Online Library). February 2026. https://onlinelibrary.wiley.com/doi/10.1155/joch/5202999
5. [5] FDA. "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products." Docket FDA-2024-D-4689. January 2025. https://www.regulations.gov/docket/FDA-2024-D-4689
6. [6] AION Labs. "The AI Revolution in Pharma: Will 2025 Be the Breakthrough Year?" March 2026. https://aionlabs.com/the-ai-revolution-in-pharma-will-2025-be-the-breakthrough-year/
7. [7] Nature News. "Chemistry Nobel Goes to Developers of AlphaFold AI that Predicts Protein Structures." October 9, 2024. https://www.nature.com/articles/d41586-024-03214-7
8. [8] Google DeepMind. "AlphaFold: Five Years of Impact." March 2026. https://deepmind.google/blog/alphafold-five-years-of-impact/
9. [9] Isomorphic Labs. "Partnerships." https://www.isomorphiclabs.com/partnerships
10. [10] Drug Target Review. "AI in Drug Discovery: 2025 in Review." February 10, 2026. https://www.drugtargetreview.com/article/192951/ai-in-drug-discovery-2025-in-review/
11. [11] Olawade DB, Fidelis SC, Marinze S, Egbon E, Osunmakinde A, Osborne A. "Artificial intelligence in clinical trials: A comprehensive review of opportunities, challenges, and future directions." International Journal of Medical Informatics. 2026;206:106141. PMID 41075423. DOI: 10.1016/j.ijmedinf.2025.106141 https://pubmed.ncbi.nlm.nih.gov/41075423/
12. [12] Getz K et al. Tufts CSDD. "Enrollment Performance: Weighing the Facts." Applied Clinical Trials. 2012. https://www.appliedclinicaltrialsonline.com/view/enrollment-performance-weighing-facts
13. [13] Tufts Center for the Study of Drug Development. Referenced in: "A Machine Learning Approach for Recruitment Prediction in Clinical Trial Design." arXiv:2111.07407. 2021. https://arxiv.org/pdf/2111.07407
14. [14] Research and Markets / Globe Newswire. "AI-Powered Clinical Trial Site Feasibility Research Report 2025: Market to Reach $3.55 Billion by 2029." November 4, 2025. https://www.globenewswire.com/news-release/2025/11/04/3180613
15. [15] Tufts CSDD. Referenced in: Clinical Trial Vanguard. "Tufts CSDD: New Insights on the Clinical Trial Industry." March 2024. https://www.clinicaltrialvanguard.com/conference-coverage/tufts-csdd-new-insights-on-the-clinical-trial-industry/
16. [16] ICH. "E6(R3): Guideline for Good Clinical Practice." Step 4, January 6, 2025. https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf
17. [17] Clinical Leader. "AI at the FDA: Legal Implications and Strategic Considerations for Drug Developers." January 2026. https://www.clinicalleader.com/doc/ai-at-the-fda-legal-implications-and-strategic-considerations-for-drug-developers-0001
18. [18] Medical Economics. "FDA's New AI Tool 'Elsa' Faces Accuracy Concerns Despite Commissioner's High Hopes." May 2026. https://www.medicaleconomics.com/view/fda-s-new-ai-tool-elsa-faces-accuracy-concerns-despite-commissioner-s-high-hopes
19. [19] Government Executive. "FDA to Pilot Real-Time Clinical Drug Trials Through Cloud and AI." April 29, 2026. https://www.govexec.com/technology/2026/04/fda-pilot-real-time-clinical-drug-trials-cloud-ai/413199/
20. [20] Chew K, Yang M (Clinical Leader). "FDA's Elsa AI Switches From Claude To Gemini: What Sponsors Need To Know." March 12, 2026. https://www.clinicalleader.com/doc/fda-s-elsa-ai-switches-from-claude-to-gemini-what-sponsors-need-to-know-0001
21. [21] EMA. "Reflection Paper on the Use of Artificial Intelligence (AI) in the Medicinal Product Lifecycle." EMA/CHMP/CVMP/83833/2023. September 2024. https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf
22. [22] GoodLifeSci / Sidley. "European Regulator Clarifies Guidance on the Use of AI in the Medicinal Product Lifecycle." October 22, 2024. https://goodlifesci.sidley.com/2024/10/22/european-regulator-clarifies-guidance-on-the-use-of-ai-in-the-medicinal-product-lifecycle/
23. [23] DDReg Pharma. "AI-Powered Regulatory Submissions: A New Era for IND, NDA, and BLA Documentation." July 2025. https://resource.ddregpharma.com/blogs/ai-powered-regulatory-submissions-a-new-era-for-ind-nda-bla-documentation/
24. [24] IntuitionLabs. "AI and the Future of Regulatory Affairs in the U.S. Pharmaceutical Industry." January 2026. https://intuitionlabs.ai/articles/ai-future-regulatory-affairs-pharma
25. [25] IntuitionLabs. "AI in Pharma Supply Chain: Forecasting and Logistics." 2026. https://intuitionlabs.ai/articles/ai-pharma-supply-chain-forecasting-logistics
26. [26] Pharmaphorum. "Reducing Drug Shortages: The Power of AI in Pharma Supply Chain Management." https://pharmaphorum.com/rd/reducing-drug-shortages-power-ai-pharma-supply-chain-management
27. [27] Markets.FinancialContent / TokenRing. "AI Revolutionizes Pharma Supply Chains: A New Era of Localized Resilience and Efficiency." November 1, 2025. https://markets.financialcontent.com/wral/article/tokenring-2025-10-31-ai-revolutionizes-pharma-supply-chains-a-new-era-of-localized-resilience-and-efficiency
28. [28] Isomorphic Labs. "Isomorphic Labs Kicks Off 2024 With Two Pharmaceutical Collaborations." January 7, 2024. https://www.isomorphiclabs.com/articles/isomorphic-labs-kicks-off-2024-with-two-pharmaceutical-collaborations
29. [29] Jumper J, Evans R, Pritzel A, et al. "Highly accurate protein structure prediction with AlphaFold." Nature. 2021;596:583-589. DOI: 10.1038/s41586-021-03819-2 https://www.nature.com/articles/s41586-021-03819-2

Suggested internal links: KScribe: Regulatory Document Generation | KScout: Site Selection Intelligence | KScreener: FHIR-Connected Patient Pre-Screening | AI in Patient Recruitment | ICH E6(R3) Compliance Checklist