Abstract
Artificial intelligence (AI) and machine learning (ML) are increasingly being applied across pharmaceutical development, from drug-property prediction and safety assessment to formulation design, model-informed development and clinical research. This perspective synthesizes discussions from Journées Galéniques 2024 and updates them in light of scientific and regulatory developments through 2026. ML enables prediction, pattern recognition, and data-driven decision-making, helping researchers identify drug targets, design new molecules, predict toxicity, and optimize drug properties. In formulation development, ML can improve solubility prediction, excipient selection, and formulation strategy by reducing trial-and-error approaches. AI-based models and imaging technologies can also enhance understanding of drug delivery systems and improve prediction of in vivo performance.
However, much of the published literature remains retrospective and is based on small or heterogeneous datasets, with limited external, prospective or clinically relevant validation. Recent regulatory frameworks place increasing emphasis on a clearly defined context of use, risk-based assessment of model credibility, data governance, generalizability, human oversight and lifecycle management. AI and ML should therefore be viewed not as replacements for pharmaceutical science, but as increasingly powerful tools that can strengthen and accelerate it. Their broader impact will depend on high-quality, interoperable data, rigorous validation, integration with pharmaceutical and clinical expertise, and clear evidence of meaningful clinical, manufacturing or regulatory value.
Introduction
Life and pharmaceutical sciences are in a constant state of scientific development and recently the extended application of data and computer science within the pharmaceutical field has attracted significant attention and interest in how this scientific addition can support pharmaceutical discovery and development. In particular, the application of machine learning (ML) has emerged as a powerful approach potentially capable of significantly impacting pharmaceutical and biomedical research. ML, is a subtype of artificial intelligence (AI) (see Fig. 1), centered around the development of algorithms based upon historical data that can support decision making without being explicitly programmed for specific tasks (Jordan and Mitchell, 2015). The formal mathematical foundations of ML date back several decades, with roots in statistics, optimization theory, and computer science.
The modern comeback of ML has been driven by advances in computational power, algorithmic innovation, and the availability of large-scale datasets. ML methods range from traditional algorithms, such as decision trees and support vector machines, to more recent and powerful models based on neural networks, including deep learning, which has significantly impacted the field over the last decade (e.g. LeCun et al., 2015; Kim et al., 2017; Cerchia and Lavecchia, 2023; Xu et al., 2024). While the conceptual roots of ML are longstanding, its recent revival in a substantially more powerful form has been enabled by advances in computational capacity, the generation of large-scale datasets through high-throughput screening (HTS), genomics, proteomics, and image analysis, and innovations in deep learning architectures. These developments have accelerated the adoption of ML in pharmaceutics. Today, ML is therefore not only a potential driver of innovation in fields like finance and marketing but also beginning to transform healthcare and pharmaceutical research.
Fig. 1. Introduction diagram of artificial intelligence and its subfields: machine learning and deep learning (Modified from Qi et al., 2024).

ML excels in pattern recognition, predictive modeling, and decision optimization, which are well aligned with the needs in the drug discovery and development process. In the early stages of drug discovery, ML can be used to mine large biomedical databases, such as genomics, transcriptomics, and proteomics datasets, with the aim to identify and validate new therapeutic targets (Ekins et al., 2019; Ferrero et al., 2017; Madhukar et al., 2019; Zhang et al., 2020; Mathai et al., 2020; Manen-Freixa and Antolin, 2024). Supervised learning algorithms can classify disease phenotypes, whereas unsupervised techniques help reveal previously undetected biological mechanisms and disease subtypes. The pharmaceutical industry has long and costly research and development programs with high attrition rates; on average it takes 10–15 years and costs upwards of $2.6 billion to bring a single novel drug from concept to market (DiMasi et al., 2016). Despite large investments in the pharmaceutical industry, many drug candidates fail during clinical trials, often due to a lack of efficacy or unforeseen safety issues. These challenges underscore the urgent need for more predictive, efficient, and data-driven approaches across the entire drug development pipeline.
In this context, the application of ML seems like an appealing addition to the scientific approach in the search for new pharmaceutical products, from both business and sustainability perspectives, as the ML-based algorithms may help us to learn more systematically from prior outcomes. The evolution of ML has coincided with exponential growth in available data within the pharmaceutical industry. Early ML applications focused primarily on supervised learning tasks, where models were trained on labeled datasets to make predictions. Over time, the field has expanded to include unsupervised learning, reinforcement learning, and generative modeling, each of which offers new capabilities for analyzing and interpreting complex systems. Deep learning has enabled the development of highly flexible, multi-layered networks capable of identifying hierarchical features from unstructured data such as images and genomic sequences (Goodfellow et al., 2016; Alom et al., 2018).
Beyond target identification, ML is increasingly applied to the design of drug-like molecules. Traditional compound screening methods relied on HTS of large chemical libraries, whereas ML-driven generative models, including variational autoencoders (VAEs), generative adversarial networks (GANs), and reinforcement learning approaches, are capable of designing novel molecules with optimized pharmacological properties based on in silico approaches (Zhavoronkov et al., 2019; Macalino et al., 2015; Schneider and Schneider, 2016; Bepari and Reza, 2021; Samad et al., 2023; Dara et al., 2022). These models not only accelerate lead identification but also facilitate the generation of compounds tailored to specific molecular targets or disease phenotypes. ML models have demonstrated the potential to outperform traditional rule-based systems in various tasks, such as predicting drug bioactivity, and forecasting adverse drug reactions (Vamathevan et al., 2019; Yang et al., 2020; Cronin et al., 2022; Jia et al., 2023; Zhao et al., 2021). Drug discovery can be seen as a sequential decision-making process, where learning algorithms are used to iteratively explore and refine compound structures based on predefined criteria, such as binding affinity or synthetic feasibility (Popova et al., 2018; Altae-Tran et al., 2017). ML may hence facilitate alternative routes and scientific discussions of which compounds to synthesize and screen – either as a supplement or as a replacement of the traditional medicinal chemical approach used in drug discovery.
Within the more supply- and development-oriented areas of the pharmaceutical industry (i.e. supply operation/manufacturing and CMC development), ML-based support has been explored in activities such as process and manufacturing development, physicochemical characterization, drug formulation, and chemical analysis. Here, process and manufacturing development refers to the optimization and scale-up of chemical processes for producing drug substances, rather than the design and synthesis of new molecular entities during drug discovery. In these downstream areas, much of the application of ML has focused on understanding and predicting properties such as solubility and permeability, which are critical for developing drug candidates into orally available products. A few examples of applying ML in the formulation development process of specific formulation systems have been described in the literature (Mäki-Lohiluoma et al., 2021; Kumar et al., 2013; Zulbeari et al., 2025; 2026; Bannigan et al., 2021; 2023), but this is clearly a field that is still at an early stage of ML implementation.
In the areas of pre-clinical and clinical development, ML has contributed to improvements in the predictability and efficiency of trials. For instance, physiologically based pharmacokinetic and ML models are used to predict pharmacokinetic and pharmacodynamic properties of candidate drugs, forecast toxicity, and model drug–drug interactions (Chen et al., 2018; Jia et al., 2025; Li et al., 2024a; Walter et al., 2025; Basile et al., 2019; Singh et al., 2024). In clinical trials, ML has been applied to improve patient recruitment through predictive modeling of electronic health record (EHR) data and to identify patients who are most likely to respond to a given therapy. Adaptive trial designs, guided by real-time ML analysis, have been described as enabling more flexible trial execution, including dosage adjustments and dynamic randomization, thereby reducing costs and potentially enhancing trial success rates (Liberali et al., 2025; Chow and Chang, 2008; Makutam et al., 2024; Zame et al., 2020).
An additional area in which AI and ML have been applied is drug metabolism and pharmacokinetics (DMPK), bridging early drug discovery with preclinical and clinical development. Data-driven models have been developed to predict a broad range of absorption, distribution, metabolism, and excretion (ADME) properties, including plasma protein binding, membrane permeability and transporter-mediated efflux, microsomal and hepatocyte metabolic stability, intrinsic clearance, cytochrome P450 (CYP) substrate and inhibition profiles, and drug–transporter interactions (Rodríguez-Pérez et al., 2023; Fang et al., 2023; Dudas and Miteva, 2024; Fluetsch et al., 2024; Han et al., 2025; Daood et al., 2025). Prospective industrial evaluations have supported this application for example, Fang et al. (2023) reported that ML models could be used to generate useful predictions across several ADME endpoints, including microsomal stability, efflux, solubility, and plasma protein binding, while also showing that predictive performance benefits from regular model retraining as new experimental data become available. Similarly, ML models have been reported for intrinsic clearance, plasma protein binding, and CYP3A4 time-dependent inhibition (Rodríguez-Pérez et al., 2023; Fluetsch et al., 2024; Han et al., 2025). Recent developments based on graph neural networks, multitask learning, chemical language models, and integrated ADMET platforms further expand the number of DMPK elements that may be evaluated computationally (Swanson et al., 2024; Handa et al., 2025; Tanihata and Iwata, 2026). Nevertheless, prediction of human pharmacokinetics from molecular structure remains challenging because experimental variability, limited or assay-specific datasets, differences in chemical space, and biological complexity can restrict model generalizability. Thus, as in formulation development, the greatest value of ML in DMPK is currently likely to arise from combining data-driven predictions with experimental measurements, mechanistic understanding, and established approaches such as in vitro–in vivo extrapolation and PBPK modeling rather than considering ML as a replacement for these methods (Dudas and Miteva, 2024; Handa et al., 2025; Tanihata and Iwata, 2026).
Applications of ML methodology in post-marketing surveillance and pharmacovigilance have also been described. In these settings, continuous monitoring of real-world data, including EHRs, insurance claims, and social media, has been applied to detect adverse drug reactions and monitor long-term efficacy (Tatonetti et al., 2012; Salas et al., 2022; Mishra and Gupta, 2025).
Despite its promise, the integration of ML into pharmaceutical workflows is not without challenges. Key concerns include the quality, heterogeneity, and bias of input data, as well as issues surrounding interpretability and explainability of complex models—particularly deep learning architectures. Regulatory bodies such as the Food and Drug Administration (FDA) and the European Medicines Agency (EMA) are actively exploring frameworks to evaluate ML-based tools, but questions about validation, transparency, and clinical applicability remain (FDA, 2021; 2025). Furthermore, the successful deployment of ML requires interdisciplinary collaboration among computational scientists, clinicians, pharmaceutical scientists, biologists, and regulatory experts. The evolution of ML reflects a dynamic interplay between algorithmic sophistication, data availability, and application-specific challenges. As ML continues to advance, it is increasingly tailored to meet the unique demands of the pharmaceutical domain. With innovations in model architecture, training paradigms, and interpretability, the field is steadily moving toward realizing the full potential of data-driven discovery and development.
To distinguish potential from demonstrated impact, the maturity of the evidence supporting each application needs to be considered. Retrospective model performance represents an important first step, but external and prospective validation, integration into real development workflows, and ultimately demonstration of clinical, manufacturing, or regulatory value provide progressively stronger evidence. Model performance should therefore be considered together with the context of use, uncertainty and the consequences of an incorrect prediction.
The recent rise of AI and ML has been significant, raising important questions about the extent to which these technologies may act as transformative forces in the pharmaceutical field. Despite their growing impact, their application in drug formulation remains relatively limited, highlighting the existing obstacles and constraints that hinder broader adoption. The emerging use of ML and AI in the field of pharmaceutical formulation—and the need thereof—was the topic of the Journées Galéniques meeting taking place in St Remy de Provence 2024 (Fig. 2). Journées Galéniques is a unique scientific meeting characterized by specialized themes, a collaborative atmosphere, and a carefully selected group of participants that fosters open scientific dialogue. Based on discussions held during the meeting, this review examines applications of AI and ML in the pharmaceutical sciences, summarizes key challenges and opportunities, and provides updated perspectives on future directions in the field.
Fig. 2. The theme of Journées Galéniques 2024 revolved around artificial intelligence and machine learning. This review capitalizes on the discussions held and provides perspectives from the participants taking part in the meeting.

Download the full article as PDF here Artificial intelligence and machine learning, a game changer in pharmaceutical development
or continue reading here on AI in pharmaceutical development
René Holm, Sanjay Konagurthu, Tausif Ahmed, Guillaume Louit, Brendan Griffen, Sune H. Holm, Phatsawee Jansook, Wei Wu, Yi Lu, Christine Allen, Panagiota Angeli, Christel A.S. Bergström, Anette Müllertz, Daniel Reker, Ferran Sanz, Martin Kuentz, Artificial intelligence and machine learning, a game changer in pharmaceutical development? – An updated perspective arising from Journees Galeniques 2024, European Journal of Pharmaceutical Sciences, Volume 227, 2026, 107688, ISSN 0928-0987, https://doi.org/10.1016/j.ejps.2026.107688.
Read more interesting articles on AI and ML here:
- Fine-tuned machine learning models for the discovery of dye nanoparticles with enhanced lung delivery
- Animal versus human oral drug bioavailability: machine learning says yes – they correlate
- Nitrites, scavengers and AI: What the FDA’s 2026 workshop revealed about nitrosamine control












































All4Nutra







