Pharmacovigilance is a vital component of public health systems, aiming to ensure the safe use of medicinal products. In this study, an artificial intelligence (AI)-based model was developed using TensorFlow to predict the likelihood of adverse drug reactions (ADRs) based on molecular structure and predefined criteria. Data from DrugBank, MedDRA, and SIDER databases were extracted, integrated, and structured in a relational model. A feedforward neural network was trained using chemical and pharmacological descriptors such as SMILES and ATC codes. The model showed consistent performance in estimating ADR risk, highlighting the potential role of AI in supporting early safety assessments. This method may enhance post-marketing surveillance through more timely and data-driven risk identification. Despite certain limitations, AI-assisted modeling represents a valuable addition to pharmacovigilance and patient safety awareness strategies.

pharmacovigilance, artificial intelligence, adverse drug reactions, machine learning, neural networks

Adverse drug reactions (ADRs) represent a significant challenge for public health systems. Meta-analyses and prospective studies have shown that a substantial proportion of hospitalized patients – between 10% and 17%, according to various estimates – experience at least one ADR, with approximately 0.3% of cases resulting in fatal outcomes (Lazarou et al. 1998). Aside from contributing to morbidity and mortality, ADRs are related to approximately 5% of all hospital admissions, placing them among the leading preventable causes of hospitalization (Pirmohamed et al. 2004).

Conventional methods for the assessment of drug safety rely primarily on clinical trials and spontaneous post-marketing reporting. However, the controlled nature of clinical trials, their limited population representativeness, and relatively short duration make them insufficient for identifying rare or delayed-onset ADRs (Miguel et al. 2012). Post-marketing surveillance also suffers from significant underreporting, with many reactions going undocumented (Hazell and Shakir 2006). In response to these limitations, the past decade has seen growing interest in the use of artificial intelligence (AI) and machine learning for predicting drug-related risks before a compound reaches the clinical stage.

Among the most promising approaches are models based on deep neural networks. These algorithms are capable of processing large volumes of heterogeneous biomedical data, including molecular structures, and can detect latent patterns that are often inaccessible through traditional statistical methods (Dey et al. 2018). Models that operate exclusively on structural representations of drug molecules are of particular interest, as they eliminate the need for biological or pharmacokinetic input. One such model is Drug-CNN, which converts SMILES sequences into matrices and inputs them into a convolutional neural network for ADR classification (Mantripragada et al. 2021). Another innovative example is the MoLFormer-XL-CNN hybrid model – a deep architecture combining a pre-trained molecular “language” model (MoLFormer-XL) with a downstream convolutional classifier. This system has already demonstrated success in predicting the risk of specific serious reactions (e.g., QT interval prolongation, teratogenicity, rhabdomyolysis) based solely on the chemical structure of known drugs. The high sensitivity and specificity achieved by these models indicate that deep learning can effectively capture structural features relevant to ADR development (Lin et al. 2024).

The study presents a predictive model using a multilayer neural network to assess the risk of ADRs based solely on the chemical structure of pharmaceuticals. This method aims to enhance traditional pharmacovigilance by enabling early safety evaluations in drug development.

In parallel with structure-based models, the integration of real-world data (RWD) – such as patient registries and electronic health records (EHRs) – has gained attention as a complementary source for ADR signal detection. RWD enhances model robustness by incorporating variables that reflect clinical practice, including comorbidities, polypharmacy interactions, and treatment duration. Emerging machine learning frameworks can process large-scale EHRs and identify patient-specific ADR patterns, which supports personalized pharmacovigilance strategies (Rajkomar et al. 2018; Vilar et al. 2018).

This study aims to develop a predictive model for forecasting ADRs, focusing solely on the chemical structures of pharmaceutical compounds and employing advanced deep learning techniques. To achieve this, we implemented a multilayer feedforward neural network using the Python programming environment along with the TensorFlow library, which provides various types of models suitable for training.

The dataset consisted of 482 distinct drug molecules represented by SMILES notations. These were converted into numerical format using established molecular descriptors (Table 1, Fig. 1).

Figure 1. 

Visual representation of SMILES and the process of molecular deconstruction. Adapted from Wu JN, Wang T, Chen Y, Tang LJ, Wu HL, Yu RQ. t-SMILES: a fragment-based molecular representation framework for de novo ligand design. Nat Commun. 2024 Jun 11;15(1): 4993. https://doi.org/10.1038/s41467-024-49388-6.

ADR annotations were then sourced from reliable databases such as DrugBank and SIDER. The study focused on six clinically relevant ADRs across different physiological systems, including hepatotoxicity, nephrotoxicity, and photosensitivity.

In the second phase of our study, we introduced molecules that had not been included in prior training into the model. This approach aimed to evaluate whether assumptions could be derived regarding their safety profiles based on clinically significant ADRs and possible similarities in chemical structures with well-characterized compounds. The predictive task was then framed as a multi-label binary classification problem. The model assesses the likelihood of each ADR’s presence or absence for a given molecule. The model was trained on a molecular dataset, validated through data partitioning, and externally tested with compounds excluded from training. This external cohort included both well-known medications and novel substances, allowing for a thorough evaluation of the model’s generalization.

The increasing focus on integrating AI into drug development has drawn attention from regulatory authorities such as the U.S. FDA and the European Medicines Agency (EMA). Both agencies have begun issuing guidance and recommendations on the responsible application of AI/ML technologies, particularly in preclinical safety assessments. The FDA’s 2021 Good Machine Learning Practice (GMLP) principles and EMA’s 2023 Reflection Paper emphasize transparency, traceability, and model interpretability. These principles are highly relevant to ADR prediction tools, which must demonstrate reliability, explainability, and clinical utility in order to gain regulatory acceptance (FDA 2021; EMA 2023).

In this context, the results from the current study show that the developed model partially aligns with such regulatory expectations. Overall, it demonstrates acceptable accuracy in predicting ADRs. We could conclude that it successfully identified many expected reactions while producing relatively few false positives. Sensitivity across different ADR categories was adequate, particularly for well-documented and commonly occurring reactions, where the model rarely failed to detect an event. Nonetheless, performance variability was evident, especially in the prediction of rare or underrepresented ADRs, where model confidence was lower and the risk of missed predictions correspondingly higher. Despite these limitations, the neural network showed a consistent behavior pattern in ADR prediction, in line with findings reported in other recent studies employing similar architectures (Mantripragada et al. 2021).

This study demonstrates that deep neural network-based models can serve as effective tools for predicting ADRs using only the chemical structure of pharmaceutical compounds. Training the model on data sourced from established databases such as DrugBank, SIDER, and MedDRA enabled the development of a functional predictive algorithm that accurately identifies known ADRs in well-characterized drugs and provides plausible risk estimates for less-studied or novel compounds.

The results support the central hypothesis of the research that artificial intelligence methods can accelerate and enhance the detection of potential toxicities, serving as a complementary asset to conventional pharmacological approaches. The model delivers quantitative risk assessments that can guide early-stage drug development, aiding in the selection of safer molecular candidates.

Nevertheless, the created model shows some limitations. First and foremost, the training dataset is limited in scope, both in terms of the number of pharmaceutical compounds and the diversity of ADR types included. This constraint may hinder the model’s generalizability and increase the risk of missed predictions for rare reactions or novel molecules (Mohsen et al. 2021).

Secondly, the structural representation of drugs using standard vectorized descriptors, while broadly applicable, fails to capture important molecular characteristics such as stereochemistry and electron density. As a result, two agents with differing toxicity profiles might be interpreted as structurally similar by the model (Dey et al. 2018; Falconer et al. 2018).

Thirdly, the absence of dose-dependent context represents a critical limitation. Many clinically relevant ADRs are dose-dependent and influenced by treatment duration, drug–drug interactions, and individual patient conditions. In its current form, the model evaluates risk as a binary function of structure alone, making it incapable of integrating pharmacokinetic considerations (Farnoush et al. 2024).

Despite these shortcomings, the created model offers practical value in several important areas. Firstly, its application in early drug development stages can save substantial resources by filtering out compounds with unfavorable safety profiles even before in vitro or in vivo testing (Shamim et al. 2024). Secondly, in a regulatory context, such models may serve as supplementary risk assessment tools and support post-marketing surveillance. Early analysis of new compounds using artificial intelligence can aid in prioritizing pharmacovigilance efforts and enable faster responses to emerging safety signals (Hu et al. 2024). Thirdly, in clinical settings, AI-based systems could serve as decision support tools, providing personalized ADR risk assessments based on a drug’s profile and, potentially, individual biomedical factors (Kwak et al. 2005).

Looking ahead, several opportunities for model enhancement are evident. Expanding the dataset to include more compounds, rare ADRs, and diverse populations would improve representativeness and generalization (Lin et al. 2024). Additionally, incorporating more input features such as pharmacokinetics, pharmacodynamics, and clinical parameters could enable more nuanced and context-aware predictions (Liu et al. 2023).

Significant advancements could also come from the use of graph neural networks (GNNs). It operates directly on molecular graphs and shows strong performance in extracting structural relationships (Wang et al. 2024). Transformer-based molecular language models, such as MoLFormer-XL, trained on billions of compounds, have already demonstrated superior results in tasks like toxicity prediction (Lin et al. 2023). Equally important is the interpretability of the model, which is essential for real-world deployment. Techniques such as SHAP values and attention mechanisms can offer insights into which molecular substructures contribute most to a given prediction, thereby increasing trust in the AI system (Dey et al. 2018; Sundararajan et al. 2017).

In summary, the presented model shows applicability across various aspects of drug safety, from molecular design to regulatory monitoring. These approaches are not intended to replace traditional pharmacovigilance methods but rather to complement them through systematic and proactive risk identification. The continued development and integration of such AI models into scientific and clinical workflows represent a key step toward more precise and timely drug safety assessment (Montastruc et al. 2023).

From a practical perspective, the implementation of such AI-driven systems could significantly strengthen pharmacovigilance – both during the evaluation of new substances and in post-marketing surveillance. These models can serve as adjunct tools in safety monitoring strategies, facilitating earlier signal detection and improving agile regulatory responses.

However, it is essential to emphasize that the model’s effectiveness is directly dependent on the scale and quality of the input data. Limitations such as a narrow spectrum of drug agents and the absence of dose-dependent or patient-specific context remain key challenges. Future efforts should focus on expanding the training datasets, integrating multimodal data (including pharmacokinetic and genetic information), and improving model interpretability.

Artificial intelligence is emerging as a promising instrument in the evaluation of drug safety. While it does not replace clinical judgment, laboratory data, or expert evaluation, it can significantly enhance them by enabling systematic, large-scale, and timely ADR prediction—contributing to safer therapies, more informed decisions, and better outcomes for patients.

Conflict of interest

The authors have declared that no competing interests exist.

Ethical statements

The authors declared that no clinical trials were used in the present study.

The authors declared that no experiments on humans or human tissues were performed for the present study.

The authors declared that no informed consent was obtained from the humans, donors or donors’ representatives participating in the study.

The authors declared that no experiments on animals were performed for the present study.

The authors declared that no commercially available immortalised human and animal cell lines were used in the pre­sent study.

Use of AI

No use of AI was reported.

Funding

This study is financed by the European Union-NextGenerationEU, through the National Recovery and Resilience Plan of the Republic of Bulgaria, project BG-RRP-2.004-0004-C01 “Strategic research and innovation program for development of Medical university – Sofia“.

Author contributions

All authors participated in the conduct of the study and preparation of the current manuscript, as follows: Conceptualization: VR, IG; Methodology: VR, IG, MD, VG-K; Writing – original draft preparation: VR, MD, IG; Writing – review and editing: MD, IG, VG-K, SD, AP, VP; Visualization: VR; Supervision: IG, MD, VP; Funding acquisition: MD, VP, IG, VG-K.

Author ORCIDs

Violeta Getova-Kolarova https://orcid.org/0000-0002-7103-3892

Maria Dimitrova https://orcid.org/0000-0002-4868-7775

Valentina Petkova https://orcid.org/0000-0002-6938-1054

Data availability

All of the data that support the findings of this study are available in the main text.