Revolutionizing Drug Discovery and Pharmacovigilance Through Artificial Intelligence:A Comprehensive Systematic Review of Machine Learning Architectures, Real? ? World Evidence Integration, and Regulatory Compliance

The pharmaceutical industry faces a profound productivity crisis. Despite exponential increases in research expenditure, the number of new molecular entities (NMEs) approved annually has remained stagnant for decades—a phenomenon termed 'Eroom's Law,' the ironic inverse of Moore's Law [1]. The average cost to bring a single drug to market now exceeds $2.6 billion, with a development timeline of 10–15 years, and late‎ ‎ stage attrition rates exceeding 90% [2]. Fundamental bottlenecks persist at every stage of the pipeline: the identification of biologically validated targets, the generation of lead compounds with optimal pharmacological profiles, the prediction of clinical outcomes from preclinical data, and the post‎ ‎ market surveillance of safety signals across heterogeneous patient populations [3].

Artificial intelligence and machine learning have emerged as transformative technologies capable of addressing these bottlenecks at unprecedented scale and speed [4]. The exponential growth in biomedical data—from genomics, proteomics, structural biology, electronic health records (EHRs), and wearable sensors—provides the substrate upon which sophisticated ML models can learn complex biological patterns that elude conventional statistical approaches [5]. Simultaneously, advances in computational hardware (graphics processing units, tensor processing units), open‎ ‎ source deep learning frameworks (TensorFlow, PyTorch), and cloud‎ ‎ based scientific computing infrastructure have dramatically reduced the barrier to deploying large‎ ‎ scale AI models in pharmaceutical research [6].

The intersection of AI and drug discovery now spans the entire translational continuum. In the

yearly discovery phase, deep learning models trained on structural and biochemical databases can screen virtual compound libraries of billions of molecules in hours, predict binding affinities with high accuracy, and generate de novo molecular structures with user‎ ‎ defined pharmacological properties [7,8]. In clinical development, ML algorithms analyze multi‎ ‎ omic patient profiles to identify predictive biomarkers, stratify trial populations, and flag safety signals in real time [9]. In the post‎ ‎ marketing phase, natural language processing (NLP) systems mine spontaneous adverse event reports, EHRs, and social media to detect pharmacovigilance signals months or years before they would be identified through traditional passive surveillance [10,11].

This systematic review provides a comprehensive and critical synthesis of the literature on AI/ML in drug discovery and pharmacovigilance, with particular attention to: (i) the mechanistic architecture and performance characteristics of leading ML models; (ii) the integration of real‎ ‎ world data (RWD) sources into AI‎ ‎ driven pharmacovigilance systems; (iii) explainable AI (XAI) methodologies and their relevance to scientific trust and regulatory decision‎ ‎ making; and (iv) the evolving international regulatory landscape for AI‎ ‎ based pharmaceutical evidence. By mapping the current state of the field, identifying gaps, and projecting future directions, this review aims to serve as a canonical reference for computational pharmacologists, regulatory scientists, and pharmaceutical innovators.

METHODS

Search Strategy and Databases

A systematic literature search was conducted in accordance with PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta‎ ‎ Analyses) guidelines [12]. Five electronic databases were searched: PubMed/MEDLINE, Embase, Web of Science (Core Collection), IEEE Xplore, and Scopus. The search period spanned January 1, 2015 to March 31, 2025, encompassing the primary epoch of deep learning application to biomedicine. Supplementary hand‎ ‎ searching of reference lists of included studies and gray literature sources (WHO pharmacovigilance reports, FDA guidance documents, EMA reflection papers) was performed to minimize publication bias.

The search strategy employed a combination of MeSH terms and free‎ ‎ text keywords using Boolean operators. Primary terms included: 'artificial intelligence,' 'machine learning,' 'deep learning,' 'neural network,' 'drug discovery,' 'pharmacovigilance,' 'adverse drug event,' 'real‎ ‎ world evidence,' 'natural language processing,' 'explainable AI,' 'regulatory submission.' Secondary terms included specific model architectures (GNN, transformer, VAE, GAN, LSTM, CNN) and application domains (ADMET prediction, target identification, de novo design, signal detection, NLP‎ ‎ based ADE extraction).

 Inclusion and Exclusion Criteria

Studies were included if they: (1) reported original research or systematic review data on AI/ML applications in drug discovery or pharmacovigilance; (2) were peer‎ ‎ reviewed and published in English; (3) provided quantitative performance metrics (e.g., AUC, F1, sensitivity, specificity, RMSE) or qualitative mechanistic insights; (4) addressed human pharmaceutical agents (small molecules or biologics). Studies were excluded if they: focused exclusively on veterinary pharmaceuticals without human relevance; used AI only as a statistical preprocessing tool without pharmacological interpretation; lacked methodological transparency; or were retracted, duplicated, or conference abstracts without full‎ ‎ text availability.

Regulatory guidance documents and white papers from major agencies (FDA, EMA, ICH, PMDA, CDSCO) were included if they directly addressed AI/ML use in pharmaceutical development or pharmacovigilance. Following deduplication and full‎ ‎ text screening, 214 primary studies, 18 regulatory guidance documents, and 11 systematic reviews met inclusion criteria and formed the evidence base for this review.

MACHINE LEARNING ARCHITECTURES IN DRUG DISCOVERY

Foundations of Modern Pharmaceutical AI

The application of ML to pharmaceutical problems is not novel; quantitative structure–activity relationship (QSAR) modeling has been practiced since the 1960s [13]. However, the deep learning revolution, catalyzed by the landmark achievements of convolutional neural networks in image recognition (AlexNet, 2012) and subsequent advances in recurrent, attention‎ ‎ based, and graph‎ ‎ based architectures, has fundamentally transformed the predictive capability and applicability of computational models in drug research [14]. Modern deep learning architectures excel in three pharmaceutical domains: (i) property prediction—learning the relationship between molecular structure and biological/physicochemical properties; (ii) generative design—synthesizing novel molecular entities with specified property profiles; and (iii) knowledge extraction—mining large‎ ‎ scale heterogeneous datasets for biological insights.

Graph Neural Networks for Molecular Modeling

Graph neural networks represent perhaps the most architecturally appropriate deep learning framework for molecular science, given that chemical compounds are naturally represented as graphs—atoms as nodes and bonds as edges [15]. GNNs operate through message‎ ‎ passing neural network (MPNN) algorithms that iteratively aggregate neighborhood information across graph nodes, learning molecular representations that encode both local and global structural context [16].

Gilmer et al. (2017) formalized the MPNN framework and demonstrated superior performance over fingerprint‎ ‎ based methods across QM9 quantum chemistry benchmarks [17]. The subsequent literature has expanded GNN application to drug–target interaction prediction, protein–ligand binding affinity estimation, and multi‎ ‎ target polypharmacology profiling. A landmark study by Stokes et al. (2020) deployed a GNN trained on Escherichia coli growth inhibition data to screen the Drug Repurposing Hub—a library of approximately 6,000 compounds—identifying halicin, a molecule with a novel structural scaffold and potent broad‎ ‎ spectrum antibiotic activity including against Mycobacterium tuberculosis and carbapenem‎ ‎ resistant Acinetobacter baumannii [18].

Subsequent GNN variants including graph attention networks (GATs), GraphSAGE, and direction‎ ‎ sensitive molecular GNNs have further improved the handling of stereochemical information, long‎ ‎ range dependencies, and scaffold generalization. DimeNet++ and SphereNet incorporate 3D geometric information through distance and angular embeddings, approaching ab initio quantum chemical accuracy at a fraction of the computational cost [19]. The integration of heterogeneous knowledge graphs—linking genes, proteins, diseases, and compounds—through GNNs (e.g., RotatE, KG‎ ‎ BERT) enables multi‎ ‎ hop reasoning across biological networks for target deconvolution and mechanism elucidation [20].

Transformer Models and Chemical Language Processing

The transformer architecture, introduced by Vaswani et al. (2017) through the seminal 'Attention Is All You Need' paper, has catalyzed revolutionary progress in natural language processing and, by extension, in chemical and biological sequence modeling [21]. The self‎ ‎ attention mechanism enables transformers to capture non‎ ‎ local dependencies within sequences—a property of particular relevance to SMILES (Simplified Molecular Input Line Entry System) notation and protein amino acid sequences, where distant positions frequently determine critical interactions.

ChemBERTa, developed by Chithrananda et al. (2020), applied the BERT (Bidirectional Encoder Representations from Transformers) pre‎ ‎ training paradigm to SMILES strings, demonstrating that a model pre‎ ‎ trained on 77 million unlabeled molecules from PubChem could be fine‎ ‎ tuned on small labeled datasets to achieve state‎ ‎ of‎ ‎ the‎ ‎ art performance on molecular property prediction benchmarks [22]. MolBERT, SELFormer, and Uni‎ ‎ Mol have further refined chemical transformer architectures by incorporating 2D/3D structural priors and atom‎ ‎ level masking strategies. In protein science, ESM‎ ‎ 2 and AlphaFold2 represent the apex of transformer application—AlphaFold2's multi‎ ‎ sequence attention mechanism achieving near‎ ‎ experimental accuracy in protein structure prediction, fundamentally reshaping structure‎ ‎ based drug design [23,24].

Large language models (LLMs) have more recently been adapted for pharmaceutical research through retrieval‎ ‎ augmented generation (RAG) and tool‎ ‎ augmented paradigms. GPT‎ ‎ 4‎ ‎ based systems have demonstrated capacity for synthesizing multi‎ ‎ step retrosynthetic pathways, interpreting clinical pharmacology literature, and engaging in structured drug–drug interaction reasoning [25]. However, concerns regarding hallucination, factual inconsistency, and non‎ ‎ calibrated uncertainty remain active research and regulatory challenges.

Generative Models for De Novo Drug Design

Generative AI has arguably been the most disruptive application of deep learning in early‎ ‎ phase drug discovery. Unlike discriminative models that classify or predict properties of existing molecules, generative models explore the vast chemical space—estimated at 10^60 drug‎ ‎ like molecules—to propose novel structures with targeted pharmacological profiles [26]. Three primary generative architectures have gained prominence: variational autoencoders, generative adversarial networks, and diffusion models.

Variational autoencoders encode molecules into a continuous latent space in which interpolation and gradient‎ ‎ based optimization can be performed. The junction‎ ‎ tree VAE (JT‎ ‎ VAE) of Jin et al. (2018) represented a significant advance by decomposing molecules into tree‎ ‎ structured scaffolds and rings prior to encoding, ensuring chemical validity throughout decoding [27]. Subsequent work by Gómez‎ ‎ Bombarelli et al. (2018) integrated Bayesian optimization over the VAE latent space to navigate toward molecules with optimized QED (quantitative estimate of drug‎ ‎ likeness) scores and predicted biological activities [28].

Generative adversarial networks produce molecules through an adversarial training process between a generator and discriminator. MolGAN by De Cao and Kipf (2018) operated directly on molecular graphs using reinforcement learning reward signals to optimize for properties including drug‎ ‎ likeness, solubility, and synthetic accessibility [29]. Diffusion models, which have achieved state‎ ‎ of‎ ‎ the‎ ‎ art performance in image generation, have been adapted to molecular design through Geometric Diffusion Models (GDM) and DiffSBDD, which condition molecular generation on 3D protein pocket geometries to enable true structure‎ ‎ based generative design [30,31]. The DiffDock platform by Corso et al. (2022) applied diffusion over the space of ligand poses, outperforming traditional docking algorithms on several pose prediction benchmarks [32].

Reinforcement Learning for Multi‎ ‎ Objective Optimization

Reinforcement learning (RL) addresses a core challenge in drug design: optimizing molecules across multiple potentially competing objectives—binding potency, selectivity, metabolic stability, aqueous solubility, blood–brain barrier permeability, and synthetic accessibility—simultaneously. RL formulates molecule generation as a sequential decision process in which an agent learns a policy that maximizes a scalar reward signal encoding the multi‎ ‎ property objective [33].

REINVENT (Olivecrona et al., 2017) established RL‎ ‎ guided RNN generation as a paradigm for lead optimization, training an agent with a prior chemical language model and reward function based on predicted biological activity [34]. REINVENT 2.0 extended this to multi‎ ‎ component scoring functions incorporating ADMET constraints [35]. DrugEx deployed transformer‎ ‎ based RL agents for scaffold‎ ‎ constrained generation, enabling lead optimization that preserves a defined pharmacophore while diversifying peripheral substituents [36]. Beyond molecule generation, RL has been applied to clinical dose optimization, with pharmacokinetic/pharmacodynamic (PK/PD) RL agents demonstrating superior individualized dosing for vancomycin and warfarin versus guideline‎ ‎ based dosing in retrospective analyses [37]

Comparative Performance Benchmarking

Systematic benchmarking of ML architectures across standardized pharmaceutical datasets enables rigorous assessment of relative performance. The MoleculeNet benchmark suite (Wu et al., 2018) provides a canonical framework encompassing quantum mechanics, physical chemistry, biophysics, and physiology datasets [38]. Across the BBBP (blood–brain barrier permeability), BACE (beta‎ ‎ secretase 1 inhibition), HIV (replication inhibition), and ClinTox (clinical trial toxicity) datasets, GNNs and transformer‎ ‎ based models consistently outperform molecular fingerprint‎ ‎ based approaches, with graph attention networks achieving AUC improvements of 3–12% across most tasks [39]. For 3D property prediction tasks, GNNs incorporating conformer‎ ‎ level information (e.g., SchNet, DimeNet++) outperform 2D‎ ‎ only models, highlighting the importance of 3D structural priors in property prediction. Table 1 summarizes the primary ML architectures surveyed in this review.

Table 1. Machine Learning Architectures in Drug Discovery: Applications, Advantages, and Representative Studies

AI IN TARGET IDENTIFICATION AND VALIDATION

Multi‎ ‎ Omics Integration for Target Discovery

Target identification—determining the biological macromolecule whose modulation will produce a desired therapeutic effect with acceptable safety—has historically been among the highest‎ ‎ risk stages of drug development [40]. Multi‎ ‎ omics data integration using ML offers a systematic approach to target prioritization by synthesizing evidence across genomics, transcriptomics, proteomics, epigenomics, and metabolomics layers to construct a systems‎ ‎ level causal model of disease biology [41].

Network medicine approaches leverage protein–protein interaction (PPI) networks and knowledge graphs to identify disease‎ ‎ associated modules—dense subnetworks of co‎ ‎ regulated genes—that represent actionable intervention points [42]. Mendelian randomization (MR), when implemented with ML‎ ‎ based confounder adjustment and pleiotropy correction, provides causal genetic evidence for target–disease associations that is less susceptible to confounding than observational epidemiological data [43]. The Open Targets Platform (Ochoa et al., 2021) integrates human genetic, somatic mutation, expression, and clinical evidence into a unified scoring framework; ML‎ ‎ based meta‎ ‎ analysis of this evidence has been shown to predict clinical success rates with significantly higher accuracy than any single evidence type alone [44].

AlphaFold2 and Structure‎ ‎ Based Drug Design

The release of AlphaFold2 (Jumper et al., 2021) by DeepMind—which predicted protein structures with median backbone RMSD below 1.0 Å for CASP14 targets—represents one of the most consequential scientific advances of the past decade for drug discovery [23]. By making high‎ ‎ confidence structural predictions available for essentially the entire human proteome (approximately 20,000 proteins) through the AlphaFold Protein Structure Database (AlphaFold DB), AlphaFold2 has eliminated a principal bottleneck to structure‎ ‎ based drug design: the laborious, expensive, and frequently unsuccessful process of experimental structure determination by X‎ ‎ ray crystallography or cryo‎ ‎ electron microscopy [45].

Downstream applications include AI‎ ‎ guided virtual screening against AlphaFold structures, allosteric binding site prediction, and cryptic pocket identification using molecular dynamics coupled with ML classifiers [46]. ESMFold (Lin et al., 2022) built upon the ESM protein language model to achieve comparable structural prediction accuracy at substantially lower computational cost, enabling rapid proteome‎ ‎ scale screening [47]. RoseTTAFold All‎ ‎ Atom (2024) extended the paradigm to predict protein structures in complex with small molecules, nucleic acids, and metal cofactors—a capability with profound implications for covalent drug design and PPI modulator discovery [48].

Drug Repurposing Through Knowledge Graph Embedding

Drug repurposing—identifying new therapeutic indications for approved drugs—offers an accelerated development pathway by leveraging established safety profiles and pharmacokinetic data [49]. Knowledge graph embedding methods, which learn low‎ ‎ dimensional representations of entities (drugs, diseases, proteins, pathways) and relations (binds, causes, treats), enable link prediction across biomedical knowledge graphs to surface repurposing hypotheses with supporting mechanistic rationale [50].

During the COVID‎ ‎ 19 pandemic, several AI‎ ‎ based repurposing analyses identified baricitinib and dexamethasone as candidate anti‎ ‎ SARS‎ ‎ CoV‎ ‎ 2 therapeutics through network proximity analyses and ML‎ ‎ predicted target engagement, with both subsequently validated in clinical trials [51]. BioKG, Hetionet, and DRKG (Drug Repurposing Knowledge Graph) provide standardized, curated knowledge graph resources that have been used as training substrates for embedding models including TransE, RotatE, ComplEx, and ERNIE [52].

ADMET PROPERTY PREDICTION AND TOXICOLOGY SCREENING

Predictive ADMET Modeling

Unfavorable absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles account for approximately 40% of drug candidate failures in clinical development [53]. AI‎ ‎ based ADMET prediction platforms enable rapid in silico screening of compound libraries, allowing medicinal chemists to prioritize synthetic efforts toward candidates with acceptable predicted profiles before committing to expensive experimental assays [54].

DeepTox (Mayr et al., 2016), a deep neural network ensemble that won the Tox21 Data Challenge, demonstrated that deep learning substantially outperforms traditional random forest and SVM classifiers on nuclear receptor and stress response pathway activity prediction across 12 toxicity assays [55]. PKSmart and DeepPK use LSTM and GNN architectures trained on clinical PK data to predict clearance, volume of distribution, half‎ ‎ life, and oral bioavailability directly from molecular structure, achieving Pearson correlation coefficients of 0.75–0.88 on held‎ ‎ out test sets [56]. CYP450 interaction prediction—critical for anticipating drug–drug interactions—has been addressed by several transformer models fine‎ ‎ tuned on human microsomal incubation data, with state‎ ‎ of‎ ‎ the‎ ‎ art models achieving balanced accuracy exceeding 85% [57].

In Vitro to In Vivo Translation

The translation of in vitro ADMET data to in vivo predictions remains a significant challenge due to species differences, protein binding variability, and the complexity of hepatic and renal extraction processes. Mechanistic PK/PD models integrated with ML components (hybrid PBPK‎ ‎ ML models) offer a promising solution by combining the biological interpretability of physiologically‎ ‎ based pharmacokinetic modeling with the pattern‎ ‎ recognition capabilities of neural networks [58]. Deep learning models trained on in vitro–in vivo correlations from large proprietary compound libraries have demonstrated improved prediction of human PK parameters compared to allometric scaling alone [59].

AI APPLICATIONS IN CLINICAL TRIAL DESIGN AND EXECUTION

Biomarker Discovery and Patient Stratification

The failure of many drugs in Phase II and III trials stems from inadequate patient stratification—enrolling broad, heterogeneous populations when efficacy or safety is confined to biologically defined subgroups [60]. ML‎ ‎ based biomarker discovery integrates multi‎ ‎ omic and clinical data to identify predictive and prognostic biomarkers that enable precision enrollment. Regularized machine learning methods (LASSO, elastic net), random survival forests, and gradient‎ ‎ boosted trees have been applied to identify genomic and proteomic signatures predictive of response in oncology, autoimmune disease, and CNS trials [61].

Deep learning survival models such as DeepSurv (Katzman et al., 2018) and Cox‎ ‎ nnet use neural networks to model time‎ ‎ to‎ ‎ event endpoints while accommodating high‎ ‎ dimensional covariates, outperforming the conventional Cox proportional hazards model on several cancer genomics benchmarks [62]. Multi‎ ‎ modal integration—combining imaging, EHR, genomics, and digital biomarker data—using attention‎ ‎ based fusion architectures has demonstrated superior outcome prediction compared to single‎ ‎ modality models in early Alzheimer's disease and treatment‎ ‎ resistant depression trials [63].

Adaptive Trial Design and AI‎ ‎ Driven Dose Finding

Adaptive clinical trial designs—which allow pre‎ ‎ specified modifications to trial parameters (sample size, dose levels, endpoint definitions) based on interim data—benefit substantially from ML‎ ‎ based decision algorithms that provide optimal adaptations while maintaining statistical validity [64]. Bayesian adaptive designs integrated with Gaussian process regression models enable continuous dose–response surface estimation and model‎ ‎ guided dose escalation in Phase I/II oncology trials, reducing the number of patients exposed to subtherapeutic or toxic doses compared to the 3+3 rule‎ ‎ based approach [65].

Synthetic control arms—constructed from real‎ ‎ world patient data using causal ML methods (propensity score matching, inverse probability weighting, targeted learning)—are increasingly accepted by regulatory agencies as historical control comparators in rare disease and oncology settings, reducing placebo arm size without sacrificing statistical rigor [66]. The FDA has issued specific guidance on the use of synthetic control arms under the RWE Framework, conditional on demonstrating exchangeability between real‎ ‎ world and trial populations through standardized covariate adjustment [67].

Digital Endpoints and Decentralized Trials

Wearable sensors and digital health technologies are generating continuous, high‎ ‎ frequency patient data streams that constitute novel digital endpoints with potential biomarker value. ML algorithms applied to actigraphy, photoplethysmography, electrocardiography, and speech data have demonstrated sensitivity to pharmacodynamic drug effects in CNS, cardiovascular, and respiratory disease trials [68]. The FDA's Digital Health Center of Excellence and the EMA's qualification procedures for digital endpoints under the Biomarker Qualification Program provide evolving regulatory frameworks for the incorporation of ML‎ ‎ processed digital biomarkers into regulatory submissions [69].

AI‎ ‎ DRIVEN PHARMACOVIGILANCE AND POST‎ ‎ MARKET SAFETY

Traditional Signal Detection and Its Limitations

Conventional pharmacovigilance relies primarily on spontaneous adverse event reporting systems (SAERS) such as the FDA Adverse Event Reporting System (FAERS), the WHO VigiBase, and the EMA EudraVigilance database. Signal detection in these systems employs disproportionality measures including the proportional reporting ratio (PRR), reporting odds ratio (ROR), and Bayesian confidence propagation neural network (BCPNN) [70]. While these methods remain regulatory standards, they suffer from well‎ ‎ documented limitations: massive under‎ ‎ reporting (estimated at 5–10% of all ADEs), reporting bias (Weber effect—early surge in reporting after drug launch), duplicate reports, absence of denominator data, and inability to detect signals confounded by concomitant medications or comorbidities [71].

7NLP‎ ‎ Based ADE Detection from Biomedical Text

Natural language processing offers a transformative capability for pharmacovigilance: automated extraction of drug–ADE relationships from the vast corpus of biomedical text clinical notes, discharge summaries, published literature, regulatory submissions, and patient‎ ‎ reported outcomes [72]. The development of transformer‎ ‎ based biomedical NLP models, particularly BioBERT (Lee et al., 2020), ClinicalBERT (Alsentzer et al., 2019), and PubMedBERT (Gu et al., 2021), has established a new performance frontier for named entity recognition (NER), relation extraction, and event detection in clinical text [73].

ADE extraction from clinical notes using fine‎ ‎ tuned BERT models has achieved F1 scores of 0.81–0.93 on established benchmarks including the 2018 n2c2 NLP Challenge datasets, substantially exceeding rule‎ ‎ based and traditional ML baselines [74]. The ADE‎ ‎ Corpus‎ ‎ V2 and the BioCreative VI Track 5 datasets provide standardized evaluation resources for drug–ADE extraction models. Beyond individual document processing, large‎ ‎ scale temporal analysis of clinical note repositories enables detection of emerging pharmacovigilance signals with population‎ ‎ level sensitivity, providing days‎ ‎ to‎ ‎ weeks lead time over spontaneous reporting [75].

Social Media and Patient‎ ‎ Reported ADE Mining

Social media platforms and online patient communities represent an underutilized pharmacovigilance resource. Patients increasingly report drug side effects through Twitter, Reddit, patient forums (PatientsLikeMe, WebMD), and health apps, often using lay language and colloquialisms that are not mapped to standard MedDRA terminology [76]. ML‎ ‎ based normalization of colloquial ADE expressions to standardized ontologies, combined with temporal trend analysis, enables the detection of safety signals from patient‎ ‎ generated content that may not yet appear in formal reporting systems.

Nikfarjam et al. (2015) published a pioneering study extracting ADEs from Twitter using CRF and maximum entropy models, establishing the feasibility of social media pharmacovigilance [77]. Subsequent transformer‎ ‎ based systems applied to Twitter and Reddit ADE detection have achieved precision of 72–85% and recall of 65–80% on manually annotated test sets, with significant improvements in low‎ ‎ resource languages through cross‎ ‎ lingual transfer learning [78]. The WHO Uppsala Monitoring Centre's VigiGram platform integrates social media signals with VigiBase spontaneous reports to provide a hybrid surveillance dashboard, representing an early regulatory application of AI‎ ‎ driven social media pharmacovigilance [79].

Real‎ ‎ World Data Integration

Real‎ ‎ world data encompasses all information routinely collected outside of controlled clinical trials, including EHRs, administrative claims, patient registries, wearable sensor data, and genomic biobanks [80]. The integration of these diverse data sources through AI/ML substantially amplifies the power and scope of pharmacovigilance. Table 2 summarizes the primary RWD sources, AI methods applied,

Pharmacovigilance Use Cases, And Associated Limitations.

Federated Learning for Multi‎ ‎ Institutional Safety Surveillance

A fundamental tension in AI‎ ‎ driven pharmacovigilance is the requirement for large, diverse patient populations to train robust signal detection models versus the privacy, regulatory, and institutional constraints that prevent centralization of patient‎ ‎ level data. Federated learning (FL)—a distributed ML paradigm in which models are trained across multiple institutions without sharing raw data—offers a principled resolution to this tension [81]. In FL, each participating institution trains a local model on its own data; only model parameters (gradients or weights) are shared with a central aggregation server, which combines them to produce a global model.

TriNetX, OMOP‎ ‎ based federated networks (PCORnet, OHDSI), and the European Health Data Space represent emerging federated data infrastructure for pharmacovigilance [82]. Studies applying FL to ADE detection have demonstrated that federated models approach the performance of centralized models trained on pooled data, with only 2–5% AUC reduction, while maintaining privacy guarantees through differential privacy augmentation [83]. The MELLODDY consortium demonstrated FL‎ ‎ based molecular property prediction across 10 pharmaceutical companies without sharing proprietary compound data, providing proof‎ ‎ of‎ ‎ concept for cross‎ ‎ industry federated pharmaceutical AI [84].

EXPLAINABLE AI AND SCIENTIFIC INTERPRETABILITY

The Black‎ ‎ Box Problem in Pharmaceutical AI

Deep learning models, despite their superior predictive performance, are fundamentally black boxes: they produce predictions without transparent justification of the features and reasoning processes that led to them [85]. This opacity poses profound challenges in pharmaceutical applications, where mechanistic understanding is inseparable from scientific validity, and where regulatory submissions require traceable, reproducible, and interpretable evidence pathways. A model that predicts a compound will cause hepatotoxicity but cannot specify which structural motifs or pharmacophoric features drive the prediction provides limited actionable insight to medicinal chemists.

XAI Methodologies

The field of explainable AI (XAI) has produced a diverse toolkit for post‎ ‎ hoc interpretation of black‎ ‎ box models [86]. SHapley Additive exPlanations (SHAP), derived from cooperative game theory, quantifies the marginal contribution of each feature to a model's prediction across all possible feature subsets, providing both global (population‎ ‎ level) and local (sample‎ ‎ level) feature importance [87]. LIME (Locally Interpretable Model‎ ‎ Agnostic Explanations) approximates complex models locally with interpretable surrogates. Gradient‎ ‎ based attribution methods—Integrated Gradients, GradCAM, SmoothGrad—compute attribution scores by backpropagating prediction gradients to input features.

For molecular models, attention visualization in graph neural networks provides chemically meaningful atom‎ ‎  and bond‎ ‎ level attribution, identifying pharmacophoric contributions to predicted properties [88]. Concept‎ ‎ based explanations (TCAV, ConceptSHAP) provide higher‎ ‎ level mechanistic interpretations aligned with chemical knowledge, enabling validation of model reasoning against established structure–activity relationships (SAR). The integration of XAI with medicinal chemistry visualization tools (3D pharmacophore mapping, SAR heat maps) bridges the gap between computational predictions and experimental design decisions.

Regulatory Expectations for Model Explainability

Regulatory agencies have increasingly articulated expectations for AI model interpretability in pharmaceutical submissions. The FDA's Artificial Intelligence/Machine Learning‎ ‎ Based Software as a Medical Device (AI/ML‎ ‎ based SaMD) Action Plan (2021) identifies transparency and explainability as core components of trustworthy AI, emphasizing the need for clear documentation of model architecture, training data characteristics, performance evaluation, and known limitations [89]. The EMA's Reflection Paper on AI (2023) requires that AI models used in clinical development be supported by appropriate validation evidence and that their outputs be interpretable in the context of clinical decision‎ ‎ making [90].

The EU AI Act (2024) establishes a risk‎ ‎ tiered regulatory framework in which AI systems classified as high‎ ‎ risk—including those used in clinical diagnosis, treatment recommendation, and pharmaceutical decision support—are subject to mandatory conformity assessments encompassing transparency, explainability, robustness, and bias mitigation requirements [91]. Compliance with these requirements necessitates systematic XAI documentation as part of the model lifecycle management process, creating a direct regulatory demand for explainability in pharmaceutical AI.

REGULATORY COMPLIANCE AND INTERNATIONAL FRAMEWORKS

 US FDA Framework for AI in Drug Development

The FDA has established a progressively sophisticated regulatory framework for AI in pharmaceutical development, spanning device software, drug development tools, and real‎ ‎ world evidence. The Real‎ ‎ World Evidence Framework (2018) established the conditions under which RWE from observational studies, pragmatic trials, and administrative data could support regulatory decisions, including new indications, labeling updates, and post‎ ‎ market surveillance [92]. The PDUFA VII commitment letter (2022) included specific provisions for developing an AI/ML action plan encompassing algorithmic bias assessment, iterative learning oversight, and pre‎ ‎ determined change control protocols (PCCPs) for adaptive AI models [93].

The FDA's Biomarker Qualification Program and Drug Development Tool (DDT) qualification pathway provide regulatory pathways for AI‎ ‎ derived biomarkers and digital endpoints, requiring demonstration of analytical validity, clinical validity, and context of use appropriateness. The FDA's emerging framework for AI in drug manufacturing (PAT—Process Analytical Technology) enables real‎ ‎ time adaptive control of manufacturing processes using AI, with implications for quality by design and continuous manufacturing [94].

 EMA and EU Regulatory Framework

The European Medicines Agency has published a series of regulatory science initiatives addressing AI in medicines development. The EMA Regulatory Science Strategy to 2025 (2020) explicitly identifies AI as a priority area, emphasizing the need for qualification procedures, scientific advice frameworks, and capacity building within the Agency for AI evaluation competency [95]. The EMA's Guidance on the Use of Real‎ ‎ World Data in Benefit‎ ‎ Risk Assessment provides a structured framework for the regulatory use of observational data, emphasizing design‎ ‎ based confounding control and causal inference methodology.

The EU AI Act, adopted in 2024, creates the world's first comprehensive AI regulatory framework with direct implications for pharmaceutical AI. Medical device AI systems are classified as high‎ ‎ risk per Annex III, requiring conformity assessment, CE marking, and registration in the EU AI database. The Act mandates human oversight provisions, algorithmic transparency, and data governance requirements that substantially overlap with good clinical practice (GCP) and good pharmacovigilance practice (GVP) obligations—creating an opportunity for integrated compliance frameworks [96].

 ICH Guidelines and International Harmonization

The International Council for Harmonisation (ICH) provides the primary framework for international regulatory harmonization in pharmaceutical development. ICH E6(R3), the Good Clinical Practice guideline finalized in 2023, explicitly addresses the use of digital technologies and real‎ ‎ world data in clinical trials, requiring systematic data governance, audit trails, and algorithmic validation for AI‎ ‎ assisted trial processes [97]. ICH E17 provides guidance on multi‎ ‎ regional clinical trial designs that incorporate AI‎ ‎ based subgroup analyses across regulatory jurisdictions. ICH Q12 on product lifecycle management establishes a framework for post‎ ‎ approval change management that accommodates AI model updates without requiring full regulatory submissions for pre‎ ‎ specified, validated changes [98].

Comparative International Regulatory Landscape

Beyond the FDA, EMA, and ICH, several national regulatory agencies have developed AI‎ ‎ specific frameworks. The PMDA (Japan) has established a regulatory sandbox for AI‎ ‎ aided drug development under the Sakigake designation scheme and published AI principles emphasizing uncertainty quantification and model documentation [99]. Health Canada has issued guidance on AI as a medical device and is piloting AI‎ ‎ assisted review processes within its regulatory operations. 

The Central Drugs Standard Control Organisation (CDSCO) of India is developing digital health guidelines that encompass AI pharmacovigilance applications within the National Pharmacovigilance Programme. Table 3 provides a comparative summary of major international regulatory frameworks.

Table 3. Comparative International Regulatory Frameworks for AI/ML in Drug Development and Pharmacovigilance