Digital technologies are radically transforming the pharmaceutical industry, reshaping not only manufacturing processes [1] but also the very approach to drug development and application [2]. Artificial intelligence (AI), machine learning technologies, and deep data analytics algorithms play a key role in this rapid transformation [3]. In the Russian Federation, digitalization is considered by the state a priority element of technological sovereignty and is enshrined in Decree of the President of the Russian Federation No. 490 (as amended on 15.02.2024) "On the Development of Artificial Intelligence in the Russian Federation" [4].

Turning to global trends, one can note a rapid growth in investments in digital technologies: in 2023, the volume amounted, according to various estimates, to between 1.5 and 3.5 billion US dollars, and by 2030, investments are projected to increase to 8–12 billion [5].

Despite the impressive successes of AI solutions in many processes of the pharmaceutical industry (for example, automation of regulatory procedures [6], optimization of drug supply logistics chains [7], quality control, and algorithmic support for pharmacoeconomic decisions [8]), their full-scale implementation into real-world practice still remains a serious challenge. These difficulties are largely due to a number of systemic barriers closely related to the specific nature of the pharmaceutical industry.

To address the identified gaps, a multi-aspect analysis of national and international approaches to the use of digital technologies in healthcare was conducted. The research focused on the strategies of the European Union (European Medicines Agency; EMA), the United Kingdom (Medicines and Healthcare products Regulatory Agency; MHRA), the USA (U.S. Food and Drug Administration; FDA), and the countries of the Eurasian Economic Union.

The objective of our study is to identify the role of digital technologies in the pharmaceutical sector and to form an understanding of the key challenges.

Artificial intelligence as a technological driver of progress

The pharmaceutical industry is entering an era of deep digital transformation, with artificial intelligence (AI) becoming its key driver [9].

In essence, AI is a set of mathematical methods and computing systems capable of identifying hidden patterns in large datasets and making decisions based on self-learning or adaptive algorithms (see Fig. 1) [10]. AI includes a wide range of tools: machine learning algorithms and neural networks [11]; "expert" systems, which are formalized sets of rules (e.g., in the form of an "IF-THEN" chain) and use logical inference mechanisms to simulate expert reasoning [12]; as well as symbolic logic methods that allow describing and analyzing cause-and-effect relationships in complex processes (e.g., in pharmacokinetics or toxicology) (see Fig. 2) [13]. All these tools are aimed at extracting knowledge from data and supporting decision-making under conditions of high complexity and uncertainty.

Fig. 1. The principle of operation of neural network architecture as part of artificial intelligence modules (AI modules)The diagram illustrates the architecture of a multilayer neural network as one of the central elements of deep learning. Input data passes through hidden layers, forming connections between them that are controlled by the central AI module.

Fig. 2. Architecture of the main components and methods of artificial intelligence (AI)

The paradox of introducing digital tools into the pharmaceutical sector

The practical application of artificial intelligence in the pharmaceutical industry illustrates the paradox of modern digitalization: local successes coexist with serious barriers to scaling. The greatest progress has been recorded in three key areas of drug development: virtual screening of molecules (significantly accelerating the search for potential candidates), modeling of pharmacokinetic profiles, and prediction of compound toxicity. All of this potentially reduces the time and cost of R&D stages [14, 15].

One of the most striking examples is the experience of Insilico Medicine, which in 2019, using the GENTRL (Generative Tensorial Reinforcement Learning) platform, designed, synthesized, and experimentally confirmed the activity of DDR1 kinase inhibitors in just 46 days, whereas traditionally a similar process takes from several months to several years [16].

However, the potential of digital technological solutions is manifested not only in the area of development and search for new molecules. It is also beginning to play an increasingly prominent role in the field of pharmacovigilance, primarily in the analysis of large arrays of heterogeneous data.

Traditionally, the detection of "safety signals" (reports from consumers or professionals about adverse events (AEs) and adverse drug reactions (ADRs)) required lengthy statistical analysis and expert evaluation. For example, in 2015, this very approach made it possible to detect the association between the use of SGLT2 inhibitors and the development of euglycemic diabetic ketoacidosis (euDKA) based on FAERS data [17]. It is precisely such tasks that are becoming a priority area for AI application – due to its ability to automatically process text data, find atypical patterns, and accelerate the detection of rare or weak risk signals.

In addition to optimizing individual stages of preclinical development, AI is also being actively integrated into more nuanced processes of working with target molecules, including hit discovery and lead compounds. For example, when using de novo design methods, AI algorithms not only analyze known structures but also generate new, potentially active compounds. A study by Yu et al. showed how a model based on a variational autoencoder with a graph structure successfully used the scaffold hopping method to generate 30,000 molecules targeting JAK1 kinase (a key target of inflammatory signaling pathways). As a result, seven active molecules were synthesized, demonstrating biological activity in in vitro tests [18].

Another illustrative example is the work by Jang et al., focused on the receptor tyrosine kinase FLT-3, which plays a key role in the pathogenesis of acute myeloid leukemia. Using a deep generative model trained on structural elements of known active compounds, the researchers generated over 10,000 new molecules. After ranking by target binding energy (based on free energy calculations), one of them was synthesized and demonstrated high selectivity for the mutant form of FLT-3. Despite the lack of full validation on healthy cells and limited experimental validation, the work demonstrates the applied potential of generative AI algorithms in the development of selective targeted inhibitors [19].

Computer-aided prediction of biological activity spectra allows for the preliminary selection of molecules that are most likely to interact with on-target targets (causing the required pharmacotherapeutic effect) and not interact with off-target targets, already at the early stages of pharmaceutical research [20].

Based on digital technologies, in addition to predicting target pharmacological activity, the assessment of possible toxic effects and the calculation of pharmacokinetic parameters of promising substances not yet synthesized are performed [21].

Digital technologies today already make it possible not only to accelerate key stages of pharmaceutical development but also to provide a more individualized approach to solving complex problems: from structural modification of the molecular scaffold to the selection of chemical substituents, taking into account unique biological targets and profiles of pharmacological activity and safety. However, despite impressive local achievements, only a few drugs discovered or developed using AI tools have reached human clinical trials, and none have received clinical approval [22].

Local successes and global limitations

Despite significant progress, the potential of artificial intelligence in drug development is still limited by a number of fundamental barriers.

One of the most discussed challenges in the literature today is the "depth of chemical space" within which trained models operate. Generative algorithms typically rely on already studied structures and molecules, which reduces their potential ability to propose truly novel compounds beyond known pharmacological and chemical classes [23, 24]. These circumstances are particularly important to consider when working with new biological targets, where the lack of sufficient initial data is critical.

Other obstacles include the low synthesizability of proposed compounds: even with theoretically promising molecules, the problem of their synthesis remains unresolved. The problem of "synthesizability" requires the use of retrosynthetic systems and the development of new metrics that take into account both the availability of reagents and the cost of synthesis [25]. A number of recent studies emphasize the need to consider the real conditions of organic synthesis already at the early stages of drug molecule design. One such approach is based on the preliminary breakdown of future compounds into structural fragments – synthons – that can be assembled using known and reproducible chemical reactions. This approach can significantly increase the synthetic accessibility of the created compounds and bring virtual modeling closer to the practical capabilities of the laboratory [26].

A separate difficulty is the formulation of criteria for evaluating the effectiveness of artificial intelligence methods. To date, there are no universal indicators to compare the results of such models with data obtained using traditional approaches (e.g., high-throughput screening methods). This complicates the objective assessment of the advantages of digital technologies in practical application [27]. Furthermore, the question remains open as to how many compounds need to be screened to obtain, with an acceptable probability, at least one substance with confirmed biological activity, a so-called hit.

Infrastructural challenges cannot be ignored either. Many academic and commercial organizations still have limited access to the computing resources necessary to fully implement AI methods. Modern algorithms consume significantly more computing power compared to classical virtual screening approaches, which limits their application under budget and technical constraints [28].

From the perspective of public and professional acceptance, a wary attitude from the medical community remains a serious barrier. This is largely due to the low transparency of most algorithms, especially in clinically significant situations where patient safety is at stake. This "opacity" (often referred to in foreign literature using the metaphor of a "black box" [29]) means the impossibility of accurately explaining on the basis of which factors the system made a particular decision. Additional difficulties are caused by the validation of AI algorithms, which further complicates the passage of regulatory review [15].

Despite these barriers, the pharmaceutical industry is currently showing a very positive trend towards overcoming them. One of the current development directions is the so-called hybrid architecture, combining machine learning algorithms with heuristic (i.e., rule-oriented) methods. This approach allows, at the initial stage, to quickly select potentially active compounds, then refine their chemical structure based on synthon analysis (structural fragments suitable for synthesis), and in the final phase, perform fine "tuning" of the molecule using deep learning technologies [23, 27].

Of particular interest is the active learning strategy, in which experimental data are sequentially included in training datasets. This can significantly reduce the number of necessary laboratory studies, both in vitro and in vivo [11]. Artificial intelligence is actively used in the design and conduct of preclinical and clinical studies of new drugs, on which the prospects of any pharmaceutical development depend [30–32].

Thus, the sustainable development of digital technologies in the pharmaceutical field is impossible without a systematic approach. For the reliable implementation of artificial intelligence into routine practice, not only technological readiness is needed, but also systemic institutional support. Key conditions should include: the development of unified regulatory standards, the inclusion of thematic blocks on digital technologies in training and continuing education programs for specialists, the organization of collaborative research to verify the effectiveness of algorithms, as well as the creation of international archives of anonymized medical data suitable for training models in compliance with ethical norms. Only under such conditions can digital technologies become a full-fledged element of pharmaceutical practice: from the stage of searching for a molecule with specified pharmacological properties, obtaining evidence of drug quality, safety, and efficacy, organizing the drug supply system, and the rational medical use of medicines.

Digital technologies, primarily solutions based on artificial intelligence, are ceasing to be just an auxiliary tool and are beginning to shape new principles for the functioning of the pharmaceutical industry. Already today, AI is being used at key stages of the drug lifecycle. However, the widespread implementation of such solutions is constrained by a number of factors: a shortage of quality data, limited model reproducibility, the lack of a regulatory framework, and the general attitude of the professional community towards new technologies.

To move from isolated achievements to sustainable digital transformation, comprehensive measures are needed (including the development of a regulated infrastructure, the integration of practical modules into professional education, as well as international cooperation on AI issues). Only with such an approach can digitalization become not a technological experiment, but an integral part of pharmaceutical practice, increasing its efficiency, transparency, and patient orientation.