Generative artificial intelligence in pharmaceutical drug development: A systematic review of time and cost efficiency across discovery, preclinical, and clinical phases

Aaron Riemer , Victoria Freund

Intelligent Pharmacy ›› 2026, Vol. 4 ›› Issue (2) : 145 -158.

PDF (3973KB)
Intelligent Pharmacy ›› 2026, Vol. 4 ›› Issue (2) :145 -158. DOI: 10.1016/j.ipha.2025.12.006
Review article
Generative artificial intelligence in pharmaceutical drug development: A systematic review of time and cost efficiency across discovery, preclinical, and clinical phases
Author information +
History +
PDF (3973KB)

Abstract

The pharmaceutical drug development process is notoriously complex, lengthy, and costly, often requiring more than a decade and billions of dollars to bring a single compound to market. Generative artificial intelligence (AI)has emerged as a disruptive technology with the potential to accelerate this process by designing novel molecules, predicting biological interactions, and generating synthetic data. This systematic review, conducted according to PRISMA 2020 guidelines, analyzes 100 peer-reviewed studies published between 2018 and early 2025 to examine how generative AI contributes to time and cost efficiency across the discovery, preclinical, and clinical phases of drug development. The findings reveal that AI demonstrates the greatest reported impact in early-stage discovery, significantly reducing timelines for molecule identification and lead optimization while lowering experimental costs, although much of this evidence is based on in-silico proof-of-concept studies.In preclinical development, AI shows potential for predictive toxicology and resource-efficient study design, though applications remain limited and prospective validation is sparse. Clinical adoption is still emerging, with efficiency gains primarily seen in cost reduction rather than time savings, and no robust evidence yet that computational acceleration translates into reduced attrition or regulatory progress. Overall, the review highlights hidden costs as well as methodological, regulatory, and ethical barriers that currently hinder broader adoption. Future research should focus on expanding AI applications to later stages of development, standardizing evaluation metrics, and ensuring responsible, explainable integration into pharmaceutical R&D workflows.

Graphical abstract

Keywords

Generative artificial intelligence / Drug discovery / Pharmaceutical development / Time and cost efficiency / Systematic review

Cite this article

Download citation ▾
Aaron Riemer, Victoria Freund. Generative artificial intelligence in pharmaceutical drug development: A systematic review of time and cost efficiency across discovery, preclinical, and clinical phases. Intelligent Pharmacy, 2026, 4 (2) : 145-158 DOI:10.1016/j.ipha.2025.12.006

登录浏览全文

4963

注册一个新账户 忘记密码

1 Introduction

The development of new pharmaceutical compounds is a complex, lengthy, and resource-intensive process. Estimates suggest that bringing a single new drug to market can cost up to 2.6 billion USD and may take 10–15 years to complete.1,2 This is largely due to high attrition rates during the various stages of drug development, including discovery, preclinical testing, and clinical trials, where more than 90% of drug candidates ultimately fail to reach approval.3 These financial and temporal burdens are further exacerbated by the rising prices of novel therapies, particularly specialty drugs, which can cost hundreds of thousands of dollars annually per patient.4 As a result, pharmaceutical companies and healthcare systems face increasing pressure to accelerate research and development (R&D) processes and to improve their efficiency.

Artificial intelligence (AI) has emerged as a disruptive technology with the potential to transform drug discovery and development pipelines. Unlike traditional computational approaches, generative AI methods such as variational autoencoders (VAEs), generative adversarial networks (GANs), transformer-based large language models, and diffusion models are capable of learning complex data distributions to design novel molecules, predict biological interactions, and generate synthetic data for pharmaceutical research.5–8 These capabilities allow exploration of vast chemical spaces beyond traditional high-throughput screening and may significantly reduce experimental cycles in early-stage R&D. Reports from industry and academia suggest that these technologies could enable pharmaceutical companies to optimize timelines and decision-making, ultimately lowering costs and improving access to innovative treatments.9–13 However, much of the available evidence originates from in-silico design studies rather than prospective experimental validation, meaning that reported efficiency gains should be interpreted cautiously in light of biological and regulatory constraints.

Although several studies have highlighted the transformative potential of AI in pharmaceutical research, most existing reviews are either focused on technical aspects of model development or address AI applications in healthcare more broadly without a structured, phase-specific analysis of its impact on R&D efficiency.14–16 There is currently no systematic synthesis that maps how generative AI contributes to improving time and cost efficiency across the entire drug development pipeline and what barriers limit its adoption in practice.

Against this backdrop, the objective of this review is to evaluate whether generative AI contributes measurable efficiencies rather than simply describing technological capabilities.

The contribution of this review is to close this gap by systematically analyzing 100 peer-reviewed studies published between 2018 and early 2025. It provides a consolidated, evidence-based overview of the current applications of generative AI across the discovery, preclinical, and clinical phases of drug development. Specifically, the review examines mechanisms through which AI enhances time and cost efficiency, identifies barriers and hidden costs associated with its implementation, and highlights opportunities for future research and methodological innovation.

This review offers novelty by providing a phase-specific synthesis of reported time- and cost-efficiency, whereas existing reviews primarily focus on technical architectures or general AI applications in healthcare. The central research question guiding this analysis is:

“What contributions does generative artificial intelligence make to improving time and cost efficiency across the different phases of pharmaceutical drug development, and what barriers currently hinder its adoption?”

By answering this question, the review delivers actionable insights for researchers, industry stakeholders, and regulators seeking to leverage AI responsibly to enhance efficiency and innovation in pharmaceutical R&D.

2 Drug development and generative AI

The pharmaceutical drug development process is designed to ensure that new compounds are both safe and effective for human use. It consists of multiple stages that collectively span over a decade and require significant financial investments.17–21 These extended timelines reflect not only procedural requirements but also fundamental biological uncertainty, as preclinical success rarely predicts safety or efficacy in humans.

2.1 Drug development phases

The process begins with the discovery phase, where researchers identify potential biological targets and although high initial hit rates rarely translate into downstream clinical efficacy. This phase involves target validation, hit identification, and lead optimization, typically requiring large chemical libraries and extensive high-throughput screening methods.22,23 Following discovery, preclinical development focuses on evaluating the pharmacological properties, safety profiles, and toxicological risks of candidate compounds in laboratory and animal studies. Finally, successful candidates enter clinical trials, which are divided into Phase I (safety and dosage), Phase II (efficacy and side effects), and Phase III (confirmation of efficacy and monitoring of adverse reactions in large patient populations). Clinical trials are particularly resource-intensive and face high failure rates, making them a significant contributor to overall R&D costs and timelines.3 As a result, efficiencies achieved in early discovery rarely compensate for high attrition further downstream (Fig. 1).

2.2 Generative AI

Generative AI introduces novel opportunities to address these challenges by transforming various aspects of the R&D pipeline. Unlike classical machine learning, which primarily focuses on analyzing existing datasets for pattern recognition, generative AI can create entirely new molecular structures, biological hypotheses, and synthetic data relevant to drug discovery and development.5–7 However, many generative applications remain at the proof-of-concept level and require biological or regulatory validation before they can inform real-world decision-making.

Several types of generative models have been applied to pharmaceutical research: Fig. 2

• Variational Autoencoders (VAEs): Used to generate new molecules with desired pharmacological properties by learning continuous chemical latent spaces.5

• Generative Adversarial Networks (GANs): Consist of two neural networks, a generator and a discriminator, trained in opposition to each other. In drug discovery, GANs can generate chemically valid and diverse molecular structures that mimic the distribution of real compounds. This adversarial setup enables the exploration of new chemical spaces and supports early-stage drug design by proposing novel candidates that traditional methods might overlook.7

• Transformer-based large language models: Capable of understanding complex structure–activity relationships and predicting drug–target interactions with high accuracy.6

• Diffusion models: Enable high-quality generative simulations of molecular dynamics and protein-ligand interactions, representing a new frontier in computational drug design.8

These models support a range of tasks across different development phases, including de novo drug design, lead optimization, predictive toxicology, virtual screening, and simulation of clinical trial outcomes.11,15,24–26 By enabling researchers to explore vast chemical spaces computationally and rapidly generate testable hypotheses, generative AI holds the potential to shorten experimental cycles, reduce costly laboratory work, and improve decision-making throughout the drug development pipeline. Yet these benefits are not evenly distributed across the pipeline, as evidence remains concentrated in early-stage discovery while later phases still lack prospective or clinically validated use cases.

3 Methods

This review was designed and conducted according to the PRISMA 2020 statement to ensure methodological transparency and reproducibility. The methodology consisted of a structured literature search, machine learning-assisted screening, and qualitative content analysis following established frameworks. To acknowledge methodological limitations, we report the search boundaries and the potential for database-related omissions, as well as uncertainty introduced by machine-learning-assisted screening. As recommended by current AI-assisted review practice, screening decisions are contextualised with caution, given that heuristic stopping rules may increase false-negative risk.

3.1 Research design

Given the novelty of the research field and the limited availability of empirical data, a qualitative desk research approach was chosen to explore the contributions of generative AI to improving time and cost efficiency in pharmaceutical drug development. An exploratory research design was appropriate because the phenomenon under investigation has only recently gained scientific attention.27 The objective was to capture, systematize, and synthesize the emerging body of knowledge to provide a foundation for future research. Because of the limited empirical maturity of the field, the findings should be interpreted as exploratory rather than confirmatory. This approach ensures that early claims about efficiency are mapped cautiously without implying causal or generalizable effects.

3.2 Literature search strategy

The systematic literature search was conducted using the Web of Science database, selected for its comprehensive and high-quality coverage of scientific publications in the life sciences and pharmaceutical research. The search covered studies published between 2018 and early 2025 to reflect recent technological advancements in AI and the corresponding rise in patent applications.28,29 Because Web of Science does not index all domain-relevant literature (e.g., PubMed, Scopus, IEEE Xplore), the search boundaries are explicitly acknowledged as a limitation regarding database recall.

The search strategy was informed by an initial exploratory search10 and refined into five thematic categories addressing different stages and efficiency aspects of drug development:

1. Discovery phase

2. Preclinical development

3. Clinical development

4. Optimization of research data

5. Time and cost efficiency

Keywords for each category combined terms from the domains of generative AI and pharmaceutical research(see Fig. 3). To support methodological transparency, all search strings and field commands are fully documented and can be replicated.

Boolean operators (AND, OR) and the field command “TS = ” (Search in Topic) were applied. Across all categories, the search identified 11,178 records. After deduplication, 2976 unique records remained(Fig. 4). Search results were exported, duplicates removed, and the dataset managed using Zotero reference software.

3.3 Inclusion and exclusion criteria

To ensure that only high-quality and relevant studies were included in the review, strict eligibility criteria were applied:

Inclusion criteria:

• Studies addressing applications of AI in pharmaceutical drug development

• Peer-reviewed journal articles or reviews

• Published between 2018 and early 2025

• Written in English

Exclusion criteria:

• Non-peer-reviewed sources such as conference abstracts, preprints, or white papers

• Studies focusing solely on diagnostics without direct relevance to drug development

• Retracted articles or studies without available full text

We acknowledge that applying these exclusion rules, particularly in cases of unavailable full texts, may introduce a risk of false-negative omissions, which is treated as a methodological limitation rather than a confirmatory filtration. Retracted or inaccessible studies were excluded to maintain research integrity, and their removal is transparently reported.

3.4 Screening process with ASReview

To identify publications relevant to the research question, a relevance screening was conducted. The open-source tool ASReview was used for this purpose, which applies active machine learning to support the screening process. The aim of this step was the systematic evaluation of titles and abstracts in order to create a focused body of literature for the subsequent content analysis. Algorithmic prioritization enabled a more efficient and largely objective selection of relevant contributions.30,31

In the first step, titles and abstracts were fed into the database. Subsequently, a manual assessment of four relevant and four irrelevant texts was conducted to establish the initial training dataset. The model was then automatically trained to prioritize additional relevant contributions. These were displayed iteratively and, after manual review of the abstract, were labeled as relevant or irrelevant, allowing the model to be continuously optimized.

Prior to the screening, a data-driven stopping strategy was defined. Due to time constraints, the process was terminated once 110 publications had been marked as relevant. This target was reached after screening 193 publications. Because the stopping threshold was determined heuristically, the procedure may have resulted in false-negative omissions, which we acknowledge as a methodological limitation rather than a definitive indicator of saturation. illustrates the change in hit rate across the last ten screened publications. The trend line shows a downward trajectory, indicating a decreasing rate of relevance (see Fig. 5).

Future replications could include a sensitivity check or validation subset to empirically quantify missed studies, which was not feasible within the present scope.

3.5 Overview of the literature corpus

The systematic search initially identified 11,178 records. After deduplication, 2976 unique studies were screened. Using ASReview,193 articles were manually assessed, resulting in 110 publications for full-text review. Of these, ten articles lacked full-text availability, and three were retracted, leaving 100 studies for qualitative analysis. We acknowledge that the exclusion of retracted or unavailable full-text studies may introduce a minor risk of omission bias, which is documented transparently rather than interpreted as a substantive filter. These cases are noted as methodological limitations rather than determinants of the evidence base.

The included studies cover all major phases of drug development—discovery, preclinical, and clinical—reflecting the increasing application of AI to optimize timelines, reduce costs, and enhance decision-making in pharmaceutical innovation. A lists all included studies with author, year, AI method, drug development phase, reported efficiency gains, and identified barriers (see Fig. 6).

3.6 Qualitative content analysis

The final dataset of 100 studies was analyzed using MAXQDA software (Version 24.9.1), a widely used tool for qualitative data analysis in social and health sciences. The qualitative content analysis followed Mayring's methodological framework,32 enabling systematic and rule-based examination of text segments.

A deductive coding approach was applied. Categories and codes were predefined before the coding process because no established coding systems existed for this research field. Two main coding dimensions were developed:

1. Drug development phase (discovery, preclinical, clinical)

2. Mechanisms of time and cost efficiency

Each relevant text passage was coded along these dimensions. Additional inductive subcodes were created to capture emerging themes such as barriers, limitations, and organizational factors influencing AI adoption. Each passage received two codes, enabling a differentiated analysis of efficiency contributions across phases.

To ensure coding reliability, a sample-based validation was conducted. Selected passages were re-coded and checked against predefined categories. Ambiguous codings were revised to enhance reliability and minimize subjective bias. Visualizations such as code matrices and cross-tabulations were generated to identify coding patterns and support interpretation of findings. Given the interpretive nature of qualitative synthesis and the novelty of the field, these coding results should be understood as exploratory patterns rather than definitive causal inferences. Accordingly, the approach aims at structuring emerging insights rather than establishing generalizable effects Fig. 7.

3.7 Risk of bias and study quality

This review is subject to several methodological limitations that may affect the validity and generalizability of its findings. First, the literature search was exclusively conducted via the Web of Science database. Although Web of Science offers high-quality indexing and broad coverage, this restriction may have led to the omission of relevant publications indexed in other databases such as PubMed, Scopus, or IEEE Xplore. Furthermore, the use of a fixed set of search terms introduces the risk of selection bias by excluding studies that employ alternative terminology or keywords.

The screening process was supported by the active-learning tool ASReview, which relies on user-labeled training data to predict article relevance. While this method improves efficiency, it introduces model dependency and may result in the under-representation of relevant studies, particularly in later screening iterations. The stopping rule, based on a perceived saturation of the relevance curve, was heuristically determined and might have excluded relevant publications that ranked lower in ASReview's prioritization.

Another important limitation lies in the qualitative coding approach. The category system was constructed deductively based on the research question and refined inductively during analysis. While this approach follows established principles of qualitative content analysis, it entails a degree of subjectivity, both in the formulation of categories and in the coding decisions. The analysis was conducted by a single coder, which increases the risk of individual bias and reduces inter-coder reliability. No formal inter-coder agreement was calculated, and coding consistency was not independently verified.

Additionally, this review only included peer-reviewed articles published in English between 2018 and 2025. Grey literature, internal industry reports, and non-English sources were excluded. This limits the scope of the review and may introduce publication bias, particularly in a fast-evolving field such as generative AI, where innovations are some-times disseminated via preprints, conference proceedings, or proprietary platforms.

Finally, the classification of AI contributions into the two efficiency dimensions (time and cost) is a conceptual abstraction that may not fully reflect the nuances of practical implementation. Reported efficiency gains are often not quantified in a comparable manner across studies, which hampers direct benchmarking or meta-analytic synthesis. Moreover, reported improvements frequently stem from in-silico simulations or computational benchmarks rather than biological or translational validation, limiting inferences about real-world performance.

Taken together, these methodological limitations do not invalidate the findings of this review but highlight areas for caution in their interpretation. Accordingly, this review should be understood as an early-stage mapping exercise rather than a confirmatory evaluation of effect sizes. Future work—especially meta-analytical or quantitative approaches—should aim to address these limitations by adopting broader inclusion criteria, involving multiple coders, and applying formal bias-assessment tools adapted to AI research.

4 Results

To provide a structured understanding of the findings, the results section begins with an overall analysis of how generative AI is currently applied throughout the drug development pipeline, followed by a more detailed, phase-specific examination of its contributions to time and cost efficiency. The results presented here reflect reported effects in the reviewed studies, many of which rely on in-silico evaluations rather than prospective biological validation.

4.1 Overall results

The analysis of the 100 included studies revealed a broad spectrum of reported applications of artificial intelligence across the drug-development pipeline. AI technologies are increasingly utilized in three main phases: discovery, preclinical development, and clinical development.

The qualitative coding identified two overarching dimensions: the drug-development phase and the mechanisms of time and cost efficiency. Each relevant text segment was assigned two codes, enabling a differentiated understanding of how AI may contribute to efficiency gains within each development stage.

The coding results indicate that the majority of applications reported to date focus on the early stages of drug development. A substantial number of studies address the discovery phase, where AI supports tasks such as target identification, molecular design, and lead optimization. Preclinical development also shows proposed benefits through predictive toxicology models, virtual screening, and data-driven optimization of experimental processes. In clinical development, AI assists in patient recruitment, trial design, and adaptive trial monitoring, although evidence at this stage remains limited to fewer studies compared to the discovery and preclinical phases.

Visual representations of code frequencies and co-occurrence analyses were generated using MAXQDA to illustrate these findings. These heatmaps and frequency tables show the distribution of coded segments across phases and efficiency mechanisms, highlighting where AI-driven improvements are most frequently reported rather than prospectively validated Fig. 8, Fig. 9.

4.2 Overview of phases and coding

The qualitative analysis of the 100 included studies revealed distinct patterns of how artificial intelligence is applied across the three main phases of pharmaceutical drug development. Using the predefined two-dimensional coding framework, each relevant text segment was assigned both a development phase code and an efficiency mechanism code (time efficiency or cost reduction). This dual coding allowed for a structured examination of how AI may contribute to each stage of drug development based on the available literature.

4.2.1 Phase I – discovery

Most AI applications were identified in the discovery phase. Studies in this phase focus on target identification, molecular design, and lead optimization. Generative AI methods, such as deep-learning-based molecular generation and virtual screening techniques, are commonly used to explore large chemical spaces efficiently. The coding results indicate that AI is reported to reduce the time required for compound identification and early-stage hypothesis testing, with several studies suggesting that this may lower costs associated with high-throughput screening Fig. 10, Table 1.

4.2.2 Phase II – preclinical development

In preclinical development, AI supports predictive toxicology, in-silico pharmacokinetics modeling, and the optimization of laboratory experiments. Many studies report that AI-driven simulations and virtual testing may help reduce the number of necessary animal experiments and shorten experimental timelines. Reported cost efficiencies are primarily based on claims of reduced trial-and-error processes and improved prioritization of candidate molecules before clinical testing Fig. 11, Table 2.

4.2.3 Phase III – clinical development

Clinical development shows fewer, but growing, applications of AI. The included studies highlight the use of AI in patient recruitment, clinical trial design, and adaptive trial monitoring. AI is also used for analyzing large patient datasets to improve dose-finding and safety assessments. Although fewer studies have reached this phase compared to discovery and preclinical stages, the evidence remains limited and largely retrospective, and current findings suggest that AI may have the potential to reduce the duration of clinical trials and related costs by improving participant selection and trial efficiency Fig. 12, Table 3.

4.3 Potentials across development phases

The analysis of 153 double-coded segments shows that the current scientific discourse attributes the greatest potential for generative AI to Phase I, the discovery stage. With 139 coded segments, this early stage accounts for the majority of mentions, where time efficiency (114 segments) is emphasized more than cost efficiency (25 segments). This highlights expectations that generative AI can potentially accelerate the lengthy initial processes of drug discovery.

Within Phase I, the “target molecule identification (de novo)” process stands out with 66 mentions in the context of time efficiency. Ambreen et al. (2025)24 report that the GENTRL model can shorten lead optimization from several months to just a few weeks by generating novel molecular structures. Mak et al. (2022)13 and Lu et al. (2023)26 report that, using this model, the entire process from data collection to validation of potent inhibitors was reduced to less than two months, compared to the traditional two to three years. This acceleration is attributed to AI's ability to explore vast chemical spaces and generate novel structures beyond classical screening libraries.

Additional mechanisms contributing to time efficiency in Phase I include the “prediction of mechanisms or toxicities” (12 mentions) and the “automated analysis of large datasets” (7 mentions).

Cost efficiency in Phase I is primarily linked to “AI-supported de novo development” (10 mentions) and “drug repurposing” (7 mentions). Economic benefits are expected as AI reduces costly experimental cycles and the risk of failure by leveraging virtual compound testing and repurposing strategies.13,24,26

In Phase II (preclinical development), fewer coded segments (10)were found. The focus here is on cost reduction (8 segments), particularly through “virtual predictive models” and “optimization of resource allocation.” Tripathi et al. (2024)25 highlight how large language models can assist researchers in designing experiments, optimizing assay selection, sample sizes, and control conditions. Askr et al. (2023)15 describe the “Dr. VAE” model, which predicts drug responses based on transcriptomic data, enabling more targeted experimental planning and reducing costly development efforts. Only 2 segments addressed time efficiency, largely related to predicting toxicological properties.

Phase III (clinical development) shows the least representation in the literature, with only 4 coded segments, all referring to cost reduction. Three of these segments discuss the potential of “virtual predictive models,” such as the DeepVCT approach described by Mishra and Gupta(2025),33 which aims to optimize clinical trial design and patient safety. Another coding relates to safety in the context of model-based predictions. No explicit time-efficiency potentials were identified for Phase III. This suggests that the duration of clinical studies—characterized by long patient recruitment times, treatment durations, follow-up requirements, and complex approval processes—has not yet been mean-ingfully shortened with current generative AI applications, or that such potentials are not yet well addressed in the literature.33

Overall, generative AI currently shows reported potential in early drug discovery, with strong emphasis on accelerating molecule identification and reducing costs via de novo design and repurposing. Later phases (preclinical and clinical development) are discussed less frequently, with a focus primarily on cost reduction rather than time savings.13,15,24–26,33

4.4 Synthesis of findings and answer to the research question

This qualitative desk research, based on a systematic analysis of 100 scientific publications, investigated the central research question: What contribution can generative AI make to improving time and cost efficiency in the various phases of pharmaceutical drug development? The evaluation of 153 double-coded literature segments provides a differentiated picture of the current scientific discourse.

Generative AI shows phase-specific and unevenly distributed potentials. The most prominent reports are observed in Phase I (discovery), where 114 of the 139 codings in this phase refer to time efficiency. Mechanisms such as accelerated target-molecule identification and lead optimization—reportedly reducing timelines from years to a few weeks—improved prediction of mechanisms and toxicities, and automated analysis of large datasets are highlighted. The ability of generative models to explore vast chemical spaces and design novel molecules de novo is portrayed in the literature as a key driver of such prospective efficiency gains. Cost-efficiency potentials through AI-supported de novo design and drug repurposing are also identified (25 codings), albeit less frequently than time-related improvements.11,13,24,26

In Phase II (preclinical development), the potential of generative AI is discussed less prominently (10 codings). The discourse primarily revolves around cost reduction (8 codings), achieved by optimizing resource allocation and utilizing virtual predictive models that enable more targeted experimental planning and may help reduce the number of animal studies. Time efficiency is only marginally addressed (2 codings), and the evidence remains largely in silico.15,25

Phase III (clinical development) shows the lowest representation in the literature (4 codings), all related to cost reduction. Virtual predictive models, such as Deep Virtual Clinical Trials, aim to optimize clinical-study design and improve patient safety, which could lead to cost savings.33 No explicit mentions of time efficiency were found in this phase, suggesting that current AI approaches have not yet demonstrated meaningful effects on clinical timelines or that this aspect has not been sufficiently explored scientifically.

Across phases, a recurring theme is the challenge of “hidden costs” of AI implementation, encompassing not only initial investments in technology and expertise but also data management, extensive validation processes, regulatory requirements, ethical considerations, and improving the interpretability of black-box models.34

In summary, according to the current scientific literature, generative AI currently shows its strongest reported potential to improve time and cost efficiency in early drug development, particularly the discovery phase. Its ability to rapidly generate and analyze hypotheses and explore large data spaces is viewed as transformative for prospective efficiency gains. In later, more complex and highly regulated phases (preclinical and clinical development), these potentials are discussed less frequently and primarily focus on cost reduction. Realizing significant time savings in these stages appears unlikely at present or remains underrepresented in the literature.

5 Discussion

The findings of this systematic review indicate a reported yet still emerging potential for generative AI to improve time and cost efficiency in pharmaceutical drug development, while also revealing substantial gaps between early-stage computational promise and later translational maturity. Compared to existing narrative reviews, which primarily emphasize technical architectures6,8,14 or discuss AI in healthcare without examining phase-specific efficiency effects,15,16,26 this study offers an early structured mapping of reported efficiency signals across discovery, preclinical, and clinical stages based on 100 peer-reviewed publications.

A central observation is the uneven distribution of evidence across development phases. Published studies most frequently report time-saving effects in Phase I, particularly through computational molecule generation, target identification, and lead optimization. Processes traditionally requiring months or years are reported in several studies to be shortened to weeks via in-silico hypothesis testing.13,24,26 These efficiency claims are closely linked to the ability of variational autoencoders, GANs, and diffusion models to explore chemical space beyond traditional high-throughput screening. However, these signals remain largely retrospective and computational, with few examples of biological or regulatory validation.

Moving beyond discovery, the available literature shows markedly diminishing evidence density. Reported efficiencies in preclinical development focus primarily on cost reduction through virtual toxicology, prioritization, and experimental planning.15,25 Time-reduction effects are rarely quantified and remain dependent on modeling assumptions rather than validated outcomes. In Phase III, evidence is sparse and largely speculative, with only four coded mentions referring to cost-related applications such as Deep Virtual Clinical Trials.33 No published evidence has yet demonstrated meaningful shortening of clinical study duration, suggesting that biological complexity, patient recruitment, and regulatory oversight continue to constrain the impact of generative AI at this stage. The observed concentration of reported efficiency gains in Phase I should therefore not be interpreted solely as an indicator of technological superiority, but also as a reflection of publication patterns within the current literature. Early-stage, in-silico studies are more readily publishable, while later-phase applications are constrained by regulatory, ethical, and proprietary barriers, resulting in a structural underrepresentation of Phase II and III evidence.

These patterns highlight several biological, translational, and regulatory constraints. First, many generative models operate as “black boxes,” complicating mechanistic interpretability and limiting their integration into pharmacological reasoning.35 Second, claims of computational efficiency depend on high-quality, harmonized datasets spanning chemistry, biology, and clinical metadata,24,26 which are unevenly available across organizations. Third, prospective validation frameworks remain under-developed. Robust benchmarking, laboratory confirmation, multi-center reproducibility studies, and standardized evaluation criteria are prerequisites before computationally generated candidates can progress safely toward clinical evaluation.16 Regulatory authorities have begun to acknowledge these challenges—visible in developments such as the EU AI Act and recent ICH guidance updates29,36—yet expectations regarding explainability, documentation standards, risk control, and model provenance remain in flux.

Realizing these potentials therefore requires strategic and organizational adaptation. Pharmaceutical firms need investments in data governance, computational infrastructure, and interdisciplinary expertise to connect modeling outputs with operational decision-making. Hybrid workflows that combine in-silico prediction with high-throughput laboratory testing may help bridge the gap between computational generation and biological validation.13,15,26 Academic institutions play an essential role in method development and early exploration, but translation into regulated spaces will require collaborative research programmes linking academia, industry, and regulatory scientists. Ethical and safety issues—including bias detection, risk transparency, and patient protection—must be addressed to support societal trust and regulatory approval.35,36

These findings must be interpreted with caution due to methodological limitations. Restriction to Web of Science may have excluded relevant literature indexed elsewhere, and predefined search terms may have missed alternative terminology. The sample reflects studies published up to early 2025, thereby excluding the most recent work. By focusing exclusively on peer-reviewed sources, potentially informative industry insights and grey literature were omitted. Finally, the deductive coding approach, while systematic, introduces interpretive subjectivity and may have limited detection of countervailing themes.32

Overall, this review provides an initial, structured mapping exercise of how generative AI is reported to influence time and cost efficiency across pharmaceutical development. Current evidence suggests that the most substantial and repeatedly claimed impacts remain concentrated in the discovery phase, whereas preclinical and clinical applications show limited and uneven maturity. Bridging this gap will require coordinated efforts in biological validation, regulatory standard-setting, dataset harmonization, and interdisciplinary collaboration before generative AI can materially alter downstream clinical outcomes or decision pathways.

6 Future directions

The current literature on generative AI in pharmaceutical drug development is still in its early stages, with a noticeable focus on exploratory case studies and model prototypes. As the field matures, several research gaps become apparent that should be addressed to fully leverage the reported potential of AI in pharmaceutical innovation.

First, the disproportionate concentration of studies in the discovery phase points to a lack of research in the preclinical and especially the clinical domains. While early drug discovery benefits from the high degree of computational control and the availability of large molecular datasets, later phases involve greater biological and regulatory complexity, which generative AI has yet to sufficiently address, reflecting the uneven maturity of evidence across phases.24,25,33 There is a clear need for interdisciplinary research efforts that integrate AI-generated insights into experimental workflows in preclinical testing and validate their predictive power in real-world settings.

Another major gap lies in the limited methodological standardization across studies. Many publications employ proprietary datasets or describe in-house developed generative models with limited reproducibility, limiting comparability across studies.15,26 Comparative benchmarking of generative architectures—such as GENTRL, Dr. VAE, and DeepVCT—remains scarce, making it difficult to evaluate their relative strengths, generalizability, or limitations across different stages of drug development.15,24,25,29,33 To advance the field, future research should emphasize transparent reporting, use of open datasets, and adoption of standardized performance metrics that facilitate systematic comparisons between models.

Beyond pharmaceutical R&D, similar limitations have been documented in adjacent clinical domains. A recent review of artificial intelligence across diagnostic imaging, clinical decision support, pathology, surgery, and drug discovery highlights persistent challenges including limited explainability, data bias, retrospective validation, and unclear regulatory pathways. These cross-domain findings reinforce that performance gains demonstrated in silico do not readily translate into clinically actionable improvements without prospective evidence and biological verification.37

Validation strategies also require further refinement. Currently, many studies rely on retrospective datasets or in silico benchmarks to assess the utility of AI-generated outputs.30,34 While these approaches offer preliminary insights, they do not replace experimental validation, which remains essential for establishing biological plausibility and translational potential. Prospective studies that integrate AI outputs into actual experimental designs—such as predictive toxicology models or virtual trial simulations—are needed to test the reliability and real-world impact of generative approaches.28,34 Such work is required to bridge the gap between in-silico performance and biological translation.

Finally, responsible AI adoption must be a key guiding principle in future work. Although generative models offer unprecedented creative power, their “black-box” nature, potential for bias, and lack of interpretability remain serious concerns.34,35 To ensure safe and ethical deployment, future research should prioritize explainability, uncertainty quantification, and risk mitigation strategies, especially in high-stakes applications such as dose prediction, patient stratification, or trial design. Moreover, regulatory frameworks must keep pace with technological advancements, providing clear guidance for the approval of AI-assisted methods in pharmaceutical R&D.36

In summary, while the current scientific discourse highlights the promise of generative AI in early drug development, substantial work is needed to expand its application to later stages, standardize evaluation methods, and implement ethical safeguards. These steps are essential to move from proof-of-concept studies to robust, clinically relevant AI systems that can be responsibly integrated into pharmaceutical innovation pipelines.

7 Conclusion

This review systematically examined the application of generative artificial intelligence across the pharmaceutical drug development process, with a specific focus on time and cost efficiency. Drawing on a comprehensive analysis of 100 peer-reviewed publications, the study revealed that the majority of research efforts to date are concentrated in the early phases of drug discovery. In particular, Phase I applications—such as de novo molecule generation, target identification, and lead optimization—benefit from generative models that can rapidly explore vast chemical spaces. These technologies have reported the potential to shorten timelines and reduce associated costs by streamlining experimental processes and minimizing trial-and-error cycles.

In contrast, the preclinical and clinical development phases remain comparatively underexplored. In preclinical research, generative AI shows promise in predictive toxicology, resource-efficient study design, and transcriptomics-based modeling. While the overall number of publications in this phase is limited, the reported contributions suggest that AI may enhance precision in experiment planning and could reduce the need for extensive in vivo testing. Clinical applications are still in their infancy, with only a few studies investigating AI-assisted trial design, patient stratification, and safety monitoring. These contributions primarily reflect cost-saving mechanisms, while concrete evidence for significant time reductions in clinical timelines remains scarce and requires further validation.

The findings of this review suggest that generative AI offers emerging potential to influence pharmaceutical R&D, but its current impact remains unevenly distributed along the development pipeline. The discovery phase is clearly the most advanced in terms of implementation and reported benefits. Later phases face methodological, regulatory, and ethical challenges that likely contribute to slower adoption and limited representation in the scientific literature. Moreover, the hidden costs of AI integration.

8 Graphical summary

To provide a concise visual overview of the findings, Table 4 summarizes the identified applications, efficiency gains, and barriers of generative AI across the three main phases of pharmaceutical drug development. The synthesis highlights that current research predominantly focuses on the early discovery stage, where AI demonstrates the strongest time- and cost-saving potential. In preclinical development, applications remain limited and mainly target cost reduction, while clinical-stage adoption is still emerging and primarily constrained by regulatory complexity and methodological challenges.

References

[1]

Avorn J. The $2.6 billion Pill—methodologic and policy considerations. N Engl J Med. 2015;372(20):1877–1879.

[2]

DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry:new estimates of R&D costs. J Health Econ. 2016;47:20–33.

[3]

Sun D, Gao W, Hu H, Zhou S. Why 90% of clinical drug development fails and how to improve it? Acta Pharm Sin B. 2022;12(7):3049–3062.

[4]

Tufts study finds big rise. cen.acs.org/articles/92/web/2014/11/Tufts-Study-Finds-Big-Rise.html. Accessed March 13, 2025.

[5]

Ochiai T, Inukai T, Akiyama M, et al. Variational autoencoder-based chemical latent space for large molecular structures with 3D complexity. Commun Chem. 2023;6(1):1–14.

[6]

Yenduri G, M R, G CS, et al. Generative pre-trained transformer: a comprehensive review. arXiv https://doi.org/10.48550/arXiv.2305.10435; 2023 May 21.

[7]

Schmidt-Colberg M. Generative Adversarial Networks (GAN): Eine Übersicht Und Beispiele Zum Thema GAN.

[8]

Yang L, Zhang Z, Song Y, et al. Diffusion models: a comprehensive survey of methods and applications. arXiv https://doi.org/10.48550/arXiv.2209.00796; 2024 Dec 2.

[9]

Shah B, Viswa CA, Zurkiya D, Leydon E, Bleys J. Generative AI in the Pharmaceutical Industry: Moving from Hype to Reality.

[10]

Doron G, Genway S, Roberts M, Jasti S. Generative AI: driving productivity and scientific breakthroughs in pharmaceutical R&D. Drug Discov Today. 2025;30(1):104272.

[11]

Ocana A, Pandiella A, Privat C, et al. Integrating AI in drug discovery: a transformative approach. Biomark Res. 2025;13(1):45.

[12]

Chen W, Liu X, Zhang S, Chen S. AI for drug discovery: resources, methods, and applications. Mol Ther Nucleic Acids. 2023;31:691–702.

[13]

Mak K-K, Balijepalli MK, Pichika MR. Success stories of AI in drug discovery. Expet Opin Drug Discov. 2022;17(1):79–92.

[14]

Tang X, Dai H, Knight E, et al. A Survey of Generative AI for de Novo Drug Design:new Frontiers in Molecule and Protein Generation. arXiv. 2024 Jun 26.

[15]

Askr H, Elgeldawi E, Ella HA, Elshaier YAMM, Gomaa MM, Hassanien AE. Deep learning in drug discovery: an integrative review. Artif Intell Rev. 2023;56(7):5975–6037.

[16]

Hasselgren C, Oprea TI. AI for drug discovery: are we there yet? Annu Rev Pharmacol Toxicol. 2024;64:527–550.

[17]

Deore AB, Dhumane JR, Wagh R, Sonawane R. The stages of drug discovery and development process. Asian J Pharmaceut Res Dev. 2019;7(6):62–67.

[18]

Bobbarala V, ed. Drug Discovery - Concepts to Market. InTech; 2018.

[19]

Hinkson IV, Madej B, Stahlberg EA. Accelerating therapeutics for opportunities in medicine: a paradigm shift in drug discovery. Front Pharmacol. 2020;11:770.

[20]

Paul SM, Mytelka DS, Dunwiddie CT, et al. How to improve R&D productivity: the pharmaceutical industry's grand challenge. Nat Rev Drug Discov. 2010;9(3):203–214.

[21]

Tamimi NAM, Ellis P. Drug development: from concept to marketing. Nephron Clin Pract. 2009;113(3):c125–c131.

[22]

Sundberg S. High-throughput and ultra-high-throughput screening: solution- and cell-based approaches. Curr Opin Biotechnol. 2000;11(1):47–53.

[23]

Kaitin KI. Deconstructing the drug development process: the new face of innovation. Clin Pharmacol Ther. 2010;87(3):356–361.

[24]

Ambreen S, Umar M, Noor A, Jain H, Ali R. Advanced AI and ML frameworks for drug discovery. Eur J Med Chem. 2025;284:117164.

[25]

Tripathi S, Gabriel K, Tripathi PK, Kim E. Large language models in drug development. Chem Biol Drug Des. 2024;103(6):e14568.

[26]

Lu M, Yin J, Zhu Q, et al. Artificial intelligence in pharmaceutical sciences. Engineering. 2023;27:37–69.

[27]

Flick U. Qualitative Sozialforschung: Eine Einführung. 11. Aufl. Reinbek Bei Hamburg:Rowohlt. 2025.

[28]

Maslej N. Artificial intelligence index Report 2025. Artif Intell. 2025.

[29]

Doerr P, Ward M, Tominaga T. ICH 30th Anniversary Publication. ICH; 2021.

[30]

Asreview Software Documentation.

[31]

Van De Schoot R, De Bruin J, Schram R, et al. An open source machine learning framework for efficient and transparent systematic reviews. Nat Mach Intell. 2021;3(2):125–133.

[32]

Mayring P. Qualitative inhaltsanalyse. Grundlagen Und Techniken. Weinheim: Beltz; 2022.

[33]

Mishra HP, Gupta R. Leveraging generative AI for drug safety. CRCEP. 2025;20(2):89–97.

[34]

Gangwal A, Lavecchia A. AI-Driven drug discovery for rare diseases. J Chem Inf Model. 2025;65(5):2214–2231.

[35]

Adadi A, Berrada M. Peeking inside the black-box: a survey on XAI. IEEE Access. 2018;6:52138–52160.

[36]

Rodriguez A, Meier C, Serazin E, et al. Unlocking the potential of AI. Drug Dis. 2023.

[37]

Ogut E. Artificial intelligence in clinical medicine: challenges across diagnostic imaging, clinical decision support, surgery, pathology, and drug discovery. Clin Pract. 2025;15(9):169.

Rights & permissions

2026 The Authors. Publishing services by Elsevier B. V. on behalf of Higher Education Press and KeAi Communications Co. Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

PDF (3973KB)

0

Accesses

0

Citation

Detail

Sections
Recommended

/

〈 〉