AI can support decision-making, enable rational drug design, determine the best course of a patient’s treatment with personalized medications, accomplish the clinical data produced, and utilised that data to create new drugs in the future [7]. From the lab to the bedside, it is logical to assume that AI will contribute to creating pharmaceutical products. Eularis created the E-VAI analytical and decision-making AI platform, which employs ML algorithms and a user-friendly interface to create analytical roadmaps based on rivals, crucial stakeholders, and the market share currently held to forecast critical factors in pharmaceutical sales [8]. This boosts sluggish sales and gives marketing directors the ability to foresee where to make expenditures. It also helps them allocate resources for optimum market share growth. Figure 2 presents an overview of several AI uses in drug discovery and development.

Fig. 2

Applications of AI in various pharmaceutical business subfields, including pharmaceutical product management and drug development

AI in drug discovery

The research and development of new drugs is a challenging, expensive, and lengthy task. On average, the R&D cycle spans around 10–15 years. Despite the significant financial investment made by the pharmaceutical industry, pursuing the next blockbuster drug persists. This R&D is since only one in every ten potential drug candidates completes phase I clinical trials and attains regulatory approval [9, 10]. The cost and time constraints associated with developing newer therapeutic compounds may be a contributing factor in the pharmaceutical industry's acceptance of AI [11].

The tools and technologies employed by AI are valuable in rapidly identifying hit and lead materials, validating drug targets, and optimizing drug structure design, potentially benefiting the healthcare industry by reducing the cost and timeline associated with discovering novel molecules. However, despite these advantages, AI must still overcome significant data hurdles, including the data's complexity, growth, diversity, and ambiguity [12, 13].

The chemical structure that would elicit the desired reaction at the target location may be predicted using a variety of in silico approaches. This structure can then be improved to meet a variety of criteria, such as potency, safety, solubility, permeability, and synthetic tractability. These methods also make it possible to plan the production of the compound and anticipate the molecule's physicochemical characteristics [13, 14].

By utilizing both structure- and ligand-based methods, along with all available data, it is feasible to hasten the elimination of non-lead compounds. Recently, researchers have employed the quantitative structure–activity relationship (QSAR) modelling device for screening potential pharmacologically active compounds from a pool of one million candidates. Moreover, the deep learning approach, an evolution of the earlier ML approach, can now handle the massive amount of data gathered throughout the drug discovery and development procedure [15, 16].

Using a computer model based on the QSAR, large quantities of compounds or certain physicochemical qualities, such as log P or log D, may be swiftly predicted. These models, however, are far from being able to forecast with any degree of accuracy complex biological traits like a compound's efficacy and undesirable side effects. Additionally, QSAR-based models have issues with limited exercise groups, erroneous investigational facts, and a need for more trial validations. To address these problems, researchers can employ newly developed AI methodologies, such as Deep Learning (DL) and pertinent modelling lessons, to assess the safety and effectiveness of pharmaceutical molecules through extensive data showing and study [17, 18].

DL models beat traditional ML techniques in 15 drugs candidate-related absorption, distribution, metabolism, excretion, and toxicity (ADMET) data sets regarding predictability. Drug metabolism sites are identified using artificial intelligence (AI) techniques like XenoSite, FAME, and SMARTCyp. By displaying molecule distributions and properties, the huge virtual chemical space suggests the existence of a molecular topographic map. Chemical space visualization's idea is to collect positional information on nearby molecules to hunt for bioactive compounds; thus, virtual screening (VS) helps choose appropriate molecules for future investigation. PubChem, ChemBank, DrugBank, and ChemDB are a few open-access chemical databases.

For the purpose of locating prospective novel drugs, AI-based QSAR approaches, such as decision trees, support vector machines, random forests, and linear discriminant analysis (LDA), have evolved from QSAR modelling tools [15, 19, 20].

We have included a list of a few AI technologies used throughout the drug development phase in Table 1 to help readers understand. Figure 3 summarizes the various AI models used during drug development methods. Physicochemical characteristics, bioactivity, toxicity, target proteins, drug interactions, drug-protein binding interactions, and de novo synthesis of certain organic synthetic compounds are all predicted by these models [21].

Table 1 The AI techniques/tools used in the drug discovery processFig. 3

Different applications of AI in drug discovery

AI in drug development

An acceptable dosage form with the essential delivery qualities must then include a unique medicinal component. In this case, AI can take the place of the conventional approach of trial and error [22]. With the use of QSPR, a variety of computational techniques may resolve issues in the formulation design area, such as instability issues, dissolving, porosity, and many more [23]. Decision-support technologies use rule-based algorithms to choose the kind, nature, and amount of the excipients depending on the physicochemical properties of the drug. They also use a feedback loop to keep an eye on and occasionally tweak the entire process [24].

Piroxicam direct-filling hard gelatin capsules were designed using a hybrid method that combines expert systems (ES) and ANN in order to achieve the necessary dissolving profile. Based on the input parameters, the Model Expert System (MES) delivers judgements and recommendations for formulation development. Contrarily, ANN make use of backpropagation learning to link the formulation parameters to the desired outcome, enabling trouble-free formulation creation. The control module collaboratively manages this process [22].

Using a variety of mathematical methods, including computational fluid dynamics (CFD), discrete element modelling (DEM), and the finite element method (FEM), researchers have investigated the effects of the powder's flow property on the die-filling and tablet compression processes [25, 26]. CFD may also be used to examine how tablet shape affects the profile of the tablet's disintegration [27]. Integrating these mathematical models with AI may have a huge positive impact on the rapid manufacturing of pharmaceutical products. Technologies incorporating AI have evolved into versatile tools that find wide application in various stages of drug development. These stages include identifying and validating drug targets, designing new drugs, repurposing existing drugs, enhancing R&D efficiency, aggregating and analysing biomedicine data, and making informed decisions regarding patient enrolment in clinical trials [17, 28, 29]. These prospective applications of AI offer the chance to mitigate bias and human interference while addressing the inefficiencies and uncertainties resulting from traditional drug development approaches [30].

Drug repurposing [31], pharmacological features [32], protein characteristics and efficacy [33], drug combination, drug-target interaction [34], and prediction of potential synthetic methods for drug-like molecules [35] are other uses of AI in the pharmaceutical industry. In addition, the identification of associations between drugs and illnesses and the development of novel biomarkers and therapeutic targets allow for the identification of new pathways and targets utilizing omics analysis [36, 37].

AI in drug formulation

Pharmaceutical sciences have seen various formulations, for example solid dispersions, extrudates, pellets, nanoparticles, and liposomes, arise in addition to standard dosage forms. The name "formulation techniques" is given to these techniques because they empower the development of formulations or incorporate functionality into common dosage forms such as tablets. AI applications in formulation techniques are even more worthwhile to investigate in order to create next-generation drug products with desired efficacy and health outcomes because these methods can successfully address a variety of API issues, such as low solubility, stability, bioavailability, and production capability [6].

Controlled-release tablet formulation

Researchers utilize pharmacokinetic simulations and ANN to develop controlled-release formulations [5]. The ANN model learns sophisticated and specialized abilities from the input and output data units with the use of Chem software. In order to anticipate the best tablet formulations based on two ideal in vitro dissolution-time profiles and two desirable in vivo release profiles, researchers use a sophisticated ANN model. Dissolution is the rate-limiting step in the in vivo absorption of the drug since it is linearly proportional to the amount of the drug taken in vivo. In vitro release patterns are often detected using the difference factors (f1), and similarity factor (f2) [38].

Immediate-release tablets formulation

To boost tablet strength, Turkoglu developed a direct compression tablet formulation utilizing hydrochlorothiazide [39]. In a different study, Kesavan and Peck developed a model of a caffeine tablet formulation to describe the diluting agent and binder content in each formulation, processing variables (type of granulator, method of adding binder), and granule and tablet properties (disintegration time, hardness, and friability). These two analyses demonstrated that neural networks performed better than traditional statistical methods. Kesavan and Peck's findings have so been re-evaluated by academics employing a variety of genetic algorithms and neural networks [40]. This presentation illustrated how the relative relevance of the output attributes and the restrictions placed on the several tiers of components and processing factors determined the ideal formulation [41]. Researchers used neuro-fuzzy computing to analyse the same data and frequently created helpful rules that highlighted the most important aspects of any item [5].

Hard gelatin capsule shell formulation

Developing hard gelatin capsule formulations involves using executive tools like ANN and expert systems (ES). ANN stimulate human mental processes, such as generalization, learning, prediction, and abstraction from domain knowledge. With ANNs, the data and statistics collected during investigative work may be transformed into knowledge very quickly, enabling the manufacturer to generate few domain-specific strategies for forthcoming occurrences or forecast the theoretical preparation’s characteristics [22]. By extending the Expert Network and conducting analysis, Wendy I. Wilson in 2005, created a capsule shell manufacturing of Biopharmaceutical Classification System II drugs, such as carbamazepine, ketoprofen, naproxen, and ibuprofen. Capsugel's expert system, for the formulation of powders in hard gelatin capsules, was used all over the world despite the drawback of just providing a proposed composition. During the initial test, researchers discovered that the system exhibited low prediction accuracy and a significant error rate. Researchers retrained the ANN using a new dataset, resulting in models with an R2 of less than 70%. Lastly, for the model drugs, the smart hybrid system predicted the quantity of drug soluble around 5%. By using only 10% of the newly generated data for cross-validation, the researchers showed that the system was capable of creating a formulation that satisfied its performance requirements. Researchers presented the system's ability to analyse several BCS class II drugs by considering wettability and intrinsic dissolving properties [42].

Solid dispersions (SD)

One or more APIs dispersed in a solid matrix describe solid dispersions [43, 44]. They are currently a practical and affordable approach for enhancing solubility and bioavailability [45]. They have been extensively employed in academics and industry to overcome concerns with API poor solubility. Many AI-based SD studies have used ANNs to optimize the formulations [46,47,48]. Researchers utilized ANNs to enhance the floating and drug release characteristics of SD of Nimodipine prepared with PEG and effervescent mixtures [46]. ANNs were employed to elucidate the relationship between variables such as API concentration, the molar mass of PEG, and temperature in a SD formulated with PVP [49]. Researchers recently developed a model using ML approaches to expect the stability of SD. They employed twenty molecular descriptors to compare eight ML methods. Among these methods, the RF model exhibited the highest estimate precision and provided insights into every input. The top five contributing parameters among the twenty descriptors they picked were the drug loading ratio, relative ambient humidity, storage temperature, preparation temperature, and molecular weight of polymers [50].

Emulsions, microemulsions, and nanoemulsions

Emulsions are biphasic systems with water and oil phases spread over each other and stabilized by an emulsifier [51]. The utilization of micro- and nanoemulsions has the potential to provide a variety of advantages, including increased API bioavailability, superb optical clarity, and improved long-term stability [52,53,54,55]. Researchers have published studies on these systems that utilize AI approaches. Kumar et al. regulated the fatty alcohol content with the use of ANNs to produce a steady o/w emulsion. Particle size, zeta potential, conductance, and viscosity were among the emulsion product properties that the ANNs could accurately predict. They also made it possible to quantify the relative significance of the inputs [56]. Gasperlin et al. successfully predicted the structures of microemulsions by creating two ANNs that can determine the kind of microemulsion from the desired composition or a differential scanning calorimetry (DSC) curve, respectively [57]. Additionally, Agatonovic-Kustrin et al. developed a stable microemulsion formulation for the oral administration of rifampicin and isoniazid using ANN model data for treating the ongoing stage of TB [58]. Amani et al. used ANNs to study potential influences on nanoemulsion particle size and discovered that the final particle size's most important factor was the total energy provided during preparation [59]. In addition, Seyed et al. looked into the component concentrations of nanoemulsion to catch the most stabilized structure with minimum cytotoxicity. They found that emulsifier concentration, which was shown to be the primary determinant of nanoemulsion stability, had no effect on cytotoxicity [60].

Self-emulsifying drug delivery systems (SEDDS)

Drugs, oils, surfactants, and occasionally cosolvents are combined in isotropic ways to create SEDDS [61]. SEDDS offer several advantages due to their physical stability, ease of production, and ability to address concerns regarding low drug bioavailability [62]. SEDDS can effectively tackle various API concerns, including enzymatic degradation, gut wall efflux, solubilization, and bioavailability [63]. Fatouros et al. utilized AI techniques such as neuro-fuzzy networks to create a dynamic lipolysis model that simulates medication absorption and predicts the IVIVC. Without requiring complex settings, the model showed significant prediction skills, indicating its potential for application in forecasting the in vivo behaviour of formulations made of lipids [64]. Utilizing ANNs coupled with I-optimal design, Parikh and Sawant optimized the crucial elements that determine the droplet size of SEDDS. When compared to the quadratic model based on I-optimal design, the ANN-coupled replicas showed the comparative contributions of every factor and were more accurate [65]. Li et al. used multiple linear regression (MLR) and ANN approaches to create quantitative structure-property relationship (QSPR) models that relate the molecular structures of the surfactant, co-surfactant, oil, and drug used in SEDDS with the drug solubility. The researchers found that key factors influencing drug solubility were the ratios of surfactant and oil, as well as the dipole moment and energy of the highest occupied molecular orbital [66].

Other formulation techniques

In addition to these formulation techniques, researchers have applied AI methods to beads and pellets [67,68,69,70,71], microparticles and nanoparticles [72,73,74,75,76,77,78,79,80,81,82,83,