Abstract
Several widely used drugs have been associated with inner ear damage, including aminoglycosides prescribed for bacterial infections and cisplatin, a principal component of several chemotherapeutic regimens. Drug-induced ototoxicity remains a significant clinical challenge, often exacerbated by the lack of a systematic framework to map the growing body of scientific literature. To fill this gap, we developed a high-throughput computational pipeline to characterize the chemical and biological landscape of otoactive small-molecules. To do so, an automated Python-based workflow was developed to retrieve and process 7,801 PubMed abstracts. GPT-4 was employed as a text extraction tool, identifying compound names and their explicitly reported roles directly from the provided abstracts. Extracted compounds were enriched with structural data from PubChem and cross-referenced with experimental drug–protein interactions from BindingDB to construct a comprehensive drug-target network. Our LLM-driven approach identified 1,758 otoactive compounds, of which 619 are otoprotective and 1,092 ototoxic. Network analysis revealed the transporters ABCC3, ABCC4, and albumin as central hubs of drug interaction for the ototoxic compounds. All results are organized and publicly accessible through a dedicated web application at https://ototoxdb.streamlit.app. This comprehensive registry serves as a primary entry point for researchers screening otoactive molecules and provides a systematic framework for the automated surveillance of ototoxic drugs, supporting the future development of otoprotective strategies.
1 Introduction
Ototoxicity refers to the cellular damage or functional impairment of the inner ear resulting from exposure to drugs or pharmacological treatments (; ). This condition often leads to irreversible hearing loss, typically starting at high frequencies, as well as tinnitus, vertigo, and postural instability. Despite its severity, ototoxic effects frequently remain undetected, complicating early treatment and clinical monitoring (; ). In children, the consequences are particularly profound, as undiagnosed hearing impairment can lead to significant developmental and communicative delays (; Watts, 2019).
Drug-induced ototoxicity is increasingly recognized as a convergent pathological process, wherein chemically diverse agents trigger overlapping molecular pathways to damage the inner ear (Watts, 2019). Typically, these compounds enter sensory hair cells through mechanotransduction channels or specialized transporters, such as CTR1 and OCT2 (; ). Once internalized, these agents disrupt redox homeostasis and catalyze excessive ROS production. This cascade induces widespread oxidative damage to cellular lipids and proteins, leading to mitochondrial dysfunction and the eventual activation of pro-apoptotic signaling (; ).
The modern recognition of ototoxicity emerged in the mid-twentieth century, primarily through the clinical use of aminoglycoside antibiotics such as streptomycin and gentamicin. These compounds, characterized by a polycationic structural scaffold, can selectively damage sensory components of the inner ear, affecting both cochlear and vestibular systems (; ; ). Similar patterns were later identified with antineoplastic agents such as cisplatin. Although structurally distinct from antibiotics, cisplatin shares a common mechanistic outcome with aminoglycosides (; ; ).
Although aminoglycosides and cisplatin represent classic examples of ototoxicity, this framework extends beyond antibiotics and chemotherapeutic agents. Commonly used drugs, including analgesics and nonsteroidal anti-inflammatory agents such as naproxen, as well as muscle relaxants and other over-the-counter medications, have been associated with ototoxic effects, particularly under conditions of prolonged administration or insufficient clinical monitoring (; ; ). The widespread availability of these compounds highlights the urgent need for a systematic catalog to support healthcare providers in informed clinical decision-making (Watts, 2019).
However, compiling an exhaustive list of ototoxicity reports remains a significant challenge due to the rapid growth of the scientific literature (). Information is currently fragmented across diverse sources and described using inconsistent terminology, which limits the ability to discern relationships between the molecules (). Consequently, computational approaches, specifically text-mining and Natural Language Processing (NLP), have emerged as essential tools for data processing (). Recent advancements in Large Language Models (LLMs) have demonstrated their capacity for recognizing relevant entities and relationships within unstructured text, allowing for the organization of complex biological data into structured formats (; ). By homogenizing literature-mined data, researchers can cross-reference findings with external repositories such as PubChem () for chemical properties, and BindingDB () or the Human Metabolome Database (Wishart et al., 2021) for protein targets and affinities. This integration provides an opportunity to transition the study of ototoxicity from isolated reports towards a unified, mechanistic framework.
In this work, we present an LLM-based approach for the automated identification and compilation of ototoxic and otoprotective molecules from scientific literature. Through systematic cross-referencing and data integration, we constructed a curated database of otoactive compounds, characterizing and annotating the compounds with their chemical features and biological targets. This approach provides a scalable foundation for the systematic study of drug-induced auditory toxicity and offers a tool for clinical surveillance and the discovery of otoprotective strategies.
2 Materials and methods
2.1 Literature search
A systematic literature search was conducted to identify publications related to ototoxic compounds. The search was performed using the Bio. Entrez package from Biopython () with a set of predefined queries targeting terms associated with drug-induced hearing and vestibular dysfunction. The following search terms were used: “Ototoxicity”, “Drug-induced hearing loss”, “Vestibulotoxicity”, “Cochleotoxicity”, “Drug-induced tinnitus”, “Drug-induced vertigo”, “Drug-induced dizziness”, “Inner ear toxicity”, “Ototoxic side effects”, “Drug-induced cochlear damage”, “Drug-induced vestibular dysfunction”, “Hearing loss AND side effect”, “Tinnitus AND side effect”, “Vertigo AND side effect”, “Dizziness AND side effect” and “sensorineural hearing loss AND side effect”.
We retrieved all available records up to May 2026, excluding articles without a title or abstract. The search was restricted to Journal Articles, Comparative Studies, Case Reports, and Clinical Trials (Phases I–III and Randomized Controlled Trials), only text written in English was considered. The search yielded 13,186 unique articles, which were downloaded in JSON format. For each entry, the title, abstract, publication date, DOI, and PubMed ID were extracted to enable cross-referencing.
2.2 GPT-assisted data extraction
To automatically identify the chemical compounds mentioned in the literature, a large language model (LLM)-based pipeline was implemented. For each article, a structured prompt was generated using the LangChain Python framework ().
Prompts were submitted to the gpt-4.1-nano-2025–04–14 model through the OpenAI API, using a zero-temperature configuration to reduce hallucinations. The OpenAI Python package was employed to handle API communication. The system prompt instructed the model to behave as a scientist with experience in pharmacology and chemical compounds, while the user prompt provided explicit instructions to identify ototoxic and otoprotective compounds and return them as a comma-separated list. The ototoxic mechanism mentioned in the article was also captured. The user prompt was structured as follows:
“You are an expert pharmacologist specializing in otolaryngology. Your task is to extract drug information from the following research article.
Definitions: Ototoxic Agent: A drug or molecule that causes damage to the inner ear (cochleotoxicity or vestibulotoxicity), hearing loss, or dizziness. - Otoprotective Agent: A compound that prevents or mitigates such damage.
Instructions:
Identify all drugs, molecules, or experimental compounds mentioned.
Determine their role: “Ototoxic” or “Otoprotective”.
Identify the administration route and dose (if mentioned).”
2.3 Compound standardization and chemical metadata integration
To ensure consistency in compound nomenclature and to enrich the extracted data with standardized chemical descriptors, all compound names obtained from the GPT-assisted extraction step were queried against the PubChem database using the PubChemPy Python package (). The corresponding PubChem Compound Identifiers (CIDs) were retrieved, as well as the IUPAC name, InChIKey, and all known synonyms. The first listed synonym was used as the representative compound name. If multiple matches were returned, the first valid CID was selected for downstream processing. Records lacking either a valid PubChem Compound Identifier or a compound name were excluded from further analysis. Following data processing, compound-specific information was identified in 7,801 of the 13,186 initially retrieved articles.
2.4 Benchmarking of retrieved data and classification
To further validate the performance of the workflow, we benchmarked our database against the FDA Adverse Event Reporting System (FAERS). Adverse event reports from the first quarter of 2026 (Q1 2026) were downloaded directly from the FAERS public dashboard. To construct a reference set of compounds with suspected ototoxic potential, reports were filtered to retain only those containing at least one of the following preferred terms in the reaction field: hearing loss, tinnitus, vertigo, or dizziness. The suspected drug substance field was then extracted and standardized to retrieve the associated compound names, yielding a reference set of 364 unique compounds. The accuracy of the automated identification and classification process was subsequently confirmed via manual validation of a random sample of abstracts.
2.5 Ototoxicity score calculation
Since some compounds have been reported as either ototoxic or otoprotective in different sources, each record was assigned a number that reflects the ototoxicity behavior reported in the article. An ototoxic value of 1 was assigned if the compound was reported as ototoxic, whereas compounds described as otoprotective or as treatments against drug-induced ototoxicity were assigned a value of −1.
The weighted score for each compound was calculated as shown in Equation 1.Where the score for a specific compound, denoted as , is calculated by taking the sum of ototoxicity values for compound across all its references and dividing it by the global sum of the absolute values of ototoxicity for every compound and reference in the entire dataset . The result is then multiplied by a scaling factor of 100. A positive score indicates that the majority of articles reporting on the compound classified it as ototoxic, while negative values indicate that most reports considered the compound as otoprotective. Compounds with a resulting score of zero were labeled as undetermined, as they were considered to reflect conflicting evidence.
The final dataset contained unique PubChem IDs, standardized compound names, and computed ototoxicity scores, providing a quantitative basis for downstream analyses and visualization of otoactive molecules. After filtering and scoring, a total of 1047 unique PubChem compounds were identified, representing the primary dataset for subsequent analysis.
2.6 Chemical space analysis of otoactive compounds
To organize the library based on structural similarity, canonical SMILES were retrieved from PubChem and processed using the RDKit Python package (https://github.com/rdkit/rdkit.git). This curation pipeline involved stripping salts, neutralizing charges, and generating standardized canonical SMILES. Structural features were then encoded as Molecular ACCess System (MACCS) Fingerprint (). To perform dimension reduction of the fingerprint matrix, a two-stage dimensionality reduction strategy was implemented, initial feature extraction via Principal Component Analysis (PCA), followed by t-distributed Stochastic Neighbor Embedding (t-SNE) using the first 10 principal components, according to (Yoshimori and Bajorath, 2020; ; ). Unsupervised clustering of the resulting t-SNE coordinates was performed using the K-means algorithm (), which partitioned the chemical space into 10 distinct clusters. The number of principal components and the optimal number of clusters were determined using the elbow method. To characterize the structural identity of each group, the cluster centroids were identified, and the molecules nearest to these coordinates were selected as representatives.
2.7 Compound-protein network construction
To facilitate the study of the molecular mechanisms underlying ototoxic compounds, the BindingDB database was used (). For each compound, the corresponding protein targets were retrieved. Only interactions involving Homo sapiens’ proteins were retained. Using this information, a network was constructed with compounds and proteins as nodes, and the reported interactions as edges.
Functional characterization of the ototoxic target profile was conducted via enrichment analysis using the gprofiler2 R package (). The Gene Ontology (GO) database () was queried to identify significantly overrepresented terms across three domains: Biological Process, Molecular Function, and Cellular Component. To refine the results and minimize redundancy, GO term aggregation and semantic reduction were performed using REVIGO Supek et al. (2011).
Network construction, topological analysis, and visualizations were executed using the igraph R package and Cytoscape ().
3 Results
3.1 Otoactive compounds database construction
The systematic literature search and filtering workflow (Figure 1A) yielded 13,582 unique articles meeting all inclusion criteria. By excluding records without a title and abstract and restricting publication types to peer-reviewed articles and clinical trials, we ensured a standardized dataset for structured analysis. Following LLM-based processing, 4,636 articles contained specific drug or chemical compound information that could be cross-referenced with the PubChem database. This dataset contains articles from 1956 up to 2026.
FIGURE 1
To quantify the reported otoactivity of the identified substances, a scoring system was developed. In this model, ototoxic compounds were assigned positive values, and otoprotective compounds were assigned negative values, with the score magnitude corresponding to the number of supporting articles identified. This scoring system showed consistency even after the top 10 most-cited compounds were removed from the calculation (Supplementary Figure S1). This analysis identified 1092 ototoxic and 619 otoprotective agents (Figure 1B). Forty-seven compounds yielded a net score of zero, indicating conflicting evidence; these were subsequently classified as undetermined. Notably, cisplatin and gentamicin are the highest-scoring ototoxic agents, while dexamethasone and amifostine were the most prominent otoprotective compounds (Figure 1C). The top 10 most cited compounds in each category are summarized in Table 1, with a comprehensive list of all 1,758 substances, including PubChem CIDs and citation metrics, provided in Supplementary Table S1.
TABLE 1
| PubChem ID | Name | Score | Clasification |
|---|---|---|---|
| 2141 | Amifostine | −1.549,604 | Otoprotective |
| 12,035 | N-Acetyl-L-cysteine | −1.5,198,386 | Otoprotective |
| 5743 | Dexamethasone | −1.2,911,903 | Otoprotective |
| 24,477 | Sodium thiosulfate | −1.0490,921 | Otoprotective |
| 84,815 | D-Methionine | −0.5,110,962 | Otoprotective |
| 8378 | Neomycin | 3.3,893,746 | Ototoxic |
| 37,768 | Amikacin | 3.7,659,718 | Ototoxic |
| 6,032 | Kanamycin | 3.99462 | Ototoxic |
| 3,467 | Gentamicin | 12.7,639,543 | Ototoxic |
| 5,460,033 | Cisplatin | 23.4,835,239 | Ototoxic |
Top 5 otoprotective and ototoxic compounds.
Where available, specific mechanisms of action, dose, and route of administration were extracted from the source literature and incorporated into the database. Mechanistic annotations reveal compound-specific patterns; for instance, cisplatin records consistently report hair cell damage in the inner ear mediated by oxidative stress (Figure 1D). Complete mechanistic data for all compounds are provided in Supplementary Table S2. Pharmacological details, including dose and route of administration, where reported in the source abstracts, are accessible through the OtotoxDB web portal (https://ototoxdb.streamlit.app).
To validate the sensitivity of our automated extraction and classification pipeline, we compared the identified compounds against a previously published list by the Food and Drug Administration. Each compound in the FAERS-derived reference set was subsequently queried against our ototoxicity database. Of the 364 compounds, our workflow returned classifications for 155, of which 139 were correctly identified as ototoxic (Figure 1E). Compounds absent from our database were considered unclassified and excluded from the precision calculation. This benchmarking exercise yielded a precision of 90%, demonstrating the capability of our workflow to correctly identify and classify molecules.
To assess performance, 350 abstracts were manually curated. The pipeline correctly identified and classified 443 out of 456 compound mentions, resulting in an accuracy of 97.1% (Supplementary Table S3).
3.2 Chemical space of otoactive compounds
To further elucidate the structural relationships among the identified compounds, we employed Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE) to group molecules by structural similarity according to their chemical fingerprints. The compounds were organized into 10 distinct clusters (Supplementary Figures S2 and S3), suggesting a high degree of diversity in the chemical space of the otoactive library (Figure 2A). A representative structure for each cluster is displayed in Figure 2B.
FIGURE 2
Clustering analysis revealed that otoactive properties were non-uniformly distributed across the identified clusters (Figures 2C,D). While most clusters exhibited chemical heterogeneity, cluster 7 was enriched in ototoxic agents, with more than 75% of the cluster composition. To characterize the chemical ontology of the otoactive library, we cross-referenced our dataset with the Human Metabolome Database (HMDB). This analysis revealed distinct chemical signatures that differentiate otoprotective and ototoxic agents (Figure 2E). Cluster 7 was predominantly composed of organooxygen compounds, specifically carbohydrates and carbohydrate conjugates such as amikacin, gentamicin and tobramycin. On the other hand, clusters 6 and 9 had the highest density of otoprotective compounds (more than 50% of the cluster). These clusters were primarily populated by steroids, prenol lipids, flavonoids, and diazines, including bioactive molecules such as ellagic acid, luteolin, genistein, epigallocatechin gallate (EGCG), and naringenin. The complete list of compounds belonging to each cluster can be found in Supplementary Table S4.
3.3 Identification of ototoxic targets related to hearing and ear development
The analysis began by examining co-occurrences of otoprotective and ototoxic drugs, defined as pairs mentioned within the same article. Cisplatin exhibited the highest frequency of co-occurrences, particularly in combination with sodium thiosulfate, dexamethasone, and amifostine. Among otoprotective agents, N-acetyl-L-cysteine exhibited the greatest number of interactions, showing protective effects against cisplatin and carboplatin, as well as aminoglycosides such as gentamicin, neomycin, and vancomycin (Figure 3A).
FIGURE 3
Following the identification of the ototoxic compounds, their targets were retrieved from BindingDB to construct a drug-protein interaction network. In this network, nodes represent drugs and proteins, while edges signify reported interactions. The transporters ABCC3 and ABCC4 emerged as the primary hubs, displaying the highest degree and betweenness centrality with over 100 drug connections each. These were followed by KCNH2 and CYP3A4, which each interacted with more than 25 drugs. Other interesting nodes are SLC22A1, a small molecule transporter present in the liver, albumin (ALB), and the carbonic anhydrases CA2 and CA1 (Figure 3B).
To identify proteins that predominantly interact with ototoxic compounds, we calculated an ototoxic/otoprotective ratio, defined as the number of ototoxic interactions divided by the total number of interactions (ototoxic + otoprotective). A higher value indicates a greater proportion of ototoxic drug interactions for a specific protein. In this regard, ALB is a remarkable node; it maintains high degree centrality and an ototoxic ratio of 1, indicating it binds exclusively to ototoxic compounds in our dataset. Other nodes displaying exclusive ototoxic interactions include TLR4, FFAR2 and 3, TMPRSS2, and RGS16.
To elucidate the biological pathways involved, a Gene Ontology (GO) enrichment analysis was performed on the targets of the ototoxic compounds. The results were dominated by terms related to phosphorylation and phosphorus metabolic processes, likely driven by the numerous targets of kinase inhibitors such as imatinib and lapatinib. Additionally, metabolic and macromolecule modification processes were significantly represented, along with an interesting enrichment in terms related to the response to stimulus, signaling and cell comunication (Figure 3C).
To identify proteins potentially linked to auditory impairment, we extracted all Gene Ontology (GO) terms associated with hearing, ear function, and development (Figure 4). This analysis revealed several critical proteins and their associated ototoxic agents. Several identified proteins, such as MYO3A and MYO3B, are essential myosins for the sensory perception of sound and the development of the cochlea. These proteins were identified as targets for several kinase inhibitors flagged by our workflow as ototoxic, including lapatinib, imatinib, vandetanib, and canertinib. Furthermore, these drugs interact with proteins such as EPHB1, MAPK3, EPHA4, and AKT, which are central to the morphogenesis of the outer ear, cranial nerves, and the brain.
FIGURE 4
Our analysis also highlighted non-steroidal anti-inflammatory drugs (NSAIDs), such as ketorolac, ibuprofen (58,560–75–1), and naproxen, as ototoxic agents targeting CASP13, a protein involved in the sensory perception of sound. Additionally, the brain morphogenesis proteins SOS1 and SLC6A4 were identified as targets of some ototoxic compounds. Notably, SLC6A4 (a serotonin transporter) is targeted by a cluster of ototoxic antidepressants, including sertraline, fluoxetine, venlafaxine, and escitalopram.
Finally, we identified a group of proteins involved in the detection of mechanical stimuli, a process fundamental to sensory perception. Amitriptyline was found to target several proteins in this category, including NTRK1, SCN9A, and SCN1A. Chloroquine was linked to BACE1, another protein critical for mechanical stimulus detection. Collectively, these results suggest that the ototoxicity of diverse compounds may be driven by the disruption of proteins involved in sound perception.
4 Discussion
In this study, we implemented a Large Language Model (LLM)-driven pipeline to systematically extract otoxicity-related data from the scientific literature. Using the GPT-4 model, we identified 1,758 unique compounds with documented otoactivity. A critical feature of our workflow is the contextual classification of substances as either ototoxic or otoprotective. This distinction is essential because compounds appearing in search results for auditory toxicity often serve one of two roles: they may be a damaging agent or a therapeutic candidate being tested for otoprotection (; Watts, 2019). By leveraging the natural language processing capabilities of LLMs, we were able to interpret the linguistic context surrounding each drug mention, ensuring highly accurate categorization that traditional text mining approaches often fail to achieve (). This approach enables a rapid, high-fidelity overview of the field, a necessity in an era where the exponential growth of pharmacological literature has rendered manual curation increasingly unsustainable ().
Beyond classification, the diversity of drug names for a single compound presents a significant challenge for automated text mining (). To address this, our workflow incorporates a robust data harmonization step designed to map various synonyms and trade names to a unique identifier. By standardizing our dataset using PubChem CIDs, we established a foundation for data integration. This facilitated cross-referencing of our findings with external repositories, such as the Human Metabolome Database (HMDB) for chemical classification and BindingDB for retrieval of drug-target interaction data (Wishart et al., 2021; ).
The implementation of the ototoxic score facilitates the identification of compounds frequently cited in the literature as either ototoxic or otoprotective. This metric serves as an approximation for the confidence and veracity of the information retrieved by the Large Language Model (LLM). As anticipated, drugs with widely known ototoxic side effects, such as cisplatin and aminoglycosides (e.g., gentamicin, neomycin, and kanamycin), yielded significantly higher scores compared to compounds with less documented effects, such as terazosin or ketorolac (; ; ; ). This approach led to a substantial expansion of the otoactive landscape compared to previous bibliometric analyses ; . This discrepancy may be attributed to two primary factors. First, while Rizk and colleagues relied on clinical trial reports, our methodology extracted data from a broader corpus of article abstracts, capturing a wider range of experimental and observational evidence. Second, the five-year temporal gap between the two studies allowed our model to incorporate more recent literature, thereby capturing the most current advancements in the field and providing a more comprehensive overview of otoactive chemical space ().
Structural similarity analysis of the otoactive library reveals distinct clustering patterns that align with the pharmacological profiles. The ototoxic group is significantly enriched with aminoglycoside and quinolone antibiotics, both of which are classical archetypes of drug-induced hearing loss (; ). Conversely, the high density of polyphenols and flavonoids within the otoprotective group suggests that antioxidant and radical-scavenging properties represent the most prevalent chemical strategy for otoprotection. This structural enrichment is likely driven by the capacity of these compounds to counteract the reactive oxygen species (ROS) typically generated by ototoxic agents within the hair cells (; ; Tan and Vlajkovic, 2023). By identifying these distinct chemical signatures, our results validate the fidelity of the LLM-driven extraction process. Our library not only captures a massive volume of data but also accurately reflects the current mechanistic consensus on otoactive substances.
Regarding the protein interaction profile, the emergence of the ATP-binding cassette (ABC) transporters ABCC3 and ABCC4, alongside albumin (ALB), as primary hubs suggests that ototoxicity may be fundamentally driven by pharmacokinetics rather than specific target affinity (; ). This supports the hypothesis that ototoxicity might be a result of localized accumulation. Previous works have reported a relationship between ototoxicity and nephrotoxicity, where drug accumulation inside the cells leads to apoptosis (; ). For instance, drugs that are inefficiently cleared from hair cells due to transporter limitations may linger in the inner ear and initiate cellular damage (). This is exemplified by furosemide, where low albumin levels are a recognized risk factor for ototoxicity. Normally, albumin binding restricts furosemide’s systemic availability and prevents its toxic accumulation in the inner ear (Whitworth et al., 1993; ).
Beyond the broad influence of drug distribution and accumulation, our analysis identified a specific subset of compounds that directly target proteins essential to auditory physiology. Of particular clinical significance are those targeting myosins (MYO3A/B), motor proteins that play a critical role in the mechanotransduction of sound stimuli. Disruptions in myosin function have been linked to hereditary hearing loss (; ), suggesting that pharmacological interference with these targets may represent an additional mechanism underlying ototoxicity. Notably, these proteins are implicated in cochlear and inner ear development, which carries particular relevance in pediatric populations, as children are especially susceptible to ototoxic medications and early hearing loss can profoundly affect language acquisition, cognitive development, and academic outcomes (; ; ).
A notable limitation of the present study is the absence of key pharmacological information, such as dose, route of administration, and frequency or duration of exposure, for the majority of records in the database. This constraint stems directly from the literature-mining methodology employed, which relies on abstract-level text as its data source, given that full-text access is unavailable for a large proportion of the published literature. Consequently, the ototoxic or otoprotective classifications presented in this dataset reflect a global assessment of reported pharmacological effects and should be interpreted with caution; users are encouraged to consult the primary sources referenced in the OtotoxDB web portal for compound-specific exposure parameters prior to drawing conclusions regarding ototoxic risk. Finally, given the rapid pace at which the literature is expanding, we encourage authors to include such parameters in future research report abstracts, as systematic reporting would greatly facilitate automated mining pipelines such as the one presented here.
5 Conclusion
We have developed a comprehensive database of ototoxic compounds through a semi-automated workflow powered by a Large Language Model (LLM). This resource serves as a high-value tool for clinicians and healthcare professionals, enabling the proactive identification of drugs with potential ototoxic side effects. The database is also intended to serve as a starting point for chemoinformatic analyses, such as structure-activity property studies. By leveraging this database, practitioners can prioritize patients for otoprotective interventions or implement rigorous follow-up protocols to detect early symptoms of hearing loss. Furthermore, the modular nature of this workflow allows for periodic updates to incorporate emerging literature, ensuring the database remains a current and reliable reference for auditory health.
Statements
Data availability statement
All scripts used for data extraction have been deposited in GitHub and are publicly available at the following repository https://github.com/aylindmm/Ototoxic_DB.git.
Author contributions
AdM-M: Writing – original draft, Conceptualization, Investigation, Writing – review and editing. JA-C: Validation, Data curation, Writing – review and editing. JB: Data curation, Investigation, Writing – review and editing, Formal Analysis. AH: Data curation, Writing – review and editing, Investigation, Formal Analysis. JM-F: Writing – review and editing, Methodology, Supervision. HN: Supervision, Conceptualization, Funding acquisition, Investigation, Writing – review and editing. GP-H: Conceptualization, Methodology, Supervision, Writing – review and editing, Funding acquisition, Investigation.
Funding
The author(s) declared that financial support was received for this work and/or its publication. Del Moral-Morales was supported by the Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI) through postdoctoral funding CVU 894530. Additionally, Del Moral-Morales, Pérez-Hérnandez, and Hugo Nájera received support from the Secretaría de Educación, Ciencia, Tecnología e Innovación (SECTEI) under Grant Number SECTEI/040/2024.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author JM-F declared that they were an editorial board member of Frontiers at the time of submission. This had no impact on the peer review process and the final decision.
Generative AI statement
The author(s) declared that generative AI was used in the creation of this manuscript. Only for style and grammar correction.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fddsv.2026.1834905/full#supplementary-material
References
1
AiL. (2024). Langchain: building applications with large language models. Available online at: https://www.langchain.com/. Version 0.2.0 (Accessed July 2024).
2
AkbariN.LamookiF. M.AminM. R.DisnadS. E.YousefinejadV.GoharniaN. (2025). An update on ototoxicity: from a genetic perspective. J. Toxicol. Sci.50, 245–261. 10.2131/jts.50.245
3
AleksanderS. A.BalhoffJ. P.CarbonS.CherryJ. M.EbertD.FeuermannM.et al (2025). The gene ontology knowledgebase in 2026. Nucleic Acids Res.54, gkaf1292–D1792. 10.1093/nar/gkaf1292
4
BishtM.BistS. (2011). Ototoxicity: the hidden menace. Indian J. Otolaryngology Head and Neck Surg.63, 255–259. 10.1007/s12070-011-0151-8
5
CockP. J. A.AntaoT.ChangJ. T.ChapmanB. A.CoxC. J.DalkeA.et al (2009). Biopython: freely available python tools for computational molecular biology and bioinformatics. Bioinformatics25, 1422–1423. 10.1093/bioinformatics/btp163
6
DagdelenJ.DunnA.LeeS.WalkerN.RosenA. S.CederG.et al (2024). Structured information extraction from scientific text with large language models. Nat. Commun.15, 1418. 10.1038/s41467-024-45563-x
7
DurantJ. L.LelandB. A.HenryD. R.NourseJ. G. (2002). Reoptimization of mdl keys for use in drug discovery. J. Chemical Information Computer Sciences42, 1273–1280. 10.1021/ci010132r
8
ElvasL. B.AlmeidaA.FerreiraJ. C. (2025). Natural language processing in medical text processing: a scoping literature review. Int. J. Med. Inf.204, 106049. 10.1016/j.ijmedinf.2025.106049
9
FetoniA. R.EramoS. L. M.RolesiR.TroianiD.PaludettiG. (2012). Antioxidant treatment with coenzyme q-ter in prevention of gentamycin ototoxicity in an animal model. Acta Otorhinolaryngol. Ital.32, 103–110.
10
FleihanT.NaderM. E.DickmanJ. D. (2024). Cisplatin vestibulotoxicity: a current review. Front. Surg.11, 1437468. 10.3389/fsurg.2024.1437468
11
ForgeA.SchachtJ. (2000). Aminoglycoside antibiotics. Audiol. Neurootol5, 3–22. 10.1159/000013861
12
FriedmanT. B.BelyantsevaI. A.FrolenkovG. I. (2020). “Myosins and hearing,” in Myosins: A Superfamily of Molecular Motors, 317–330.
13
GanesanP.SchmiedgeJ.ManchaiahV.SwapnaS.DhandayuthamS.KothandaramanP. P. (2018). Ototoxicity: a challenge in diagnosis and treatment. J. Audiol. Otol.22, 59–68. 10.7874/jao.2017.00360
14
GarinisA. C.KemphA.TharpeA. M.WeitkampJ.-H.McEvoyC.SteygerP. S. (2018). Monitoring neonates for ototoxicity. Int. Journal Audiology57, S54–S61. 10.1080/14992027.2017.1339130
15
GuptaS.MahmoodA.ShettyP.AdeboyeA.RamprasadR. (2024). Data extraction from polymer literature using large language models. Commun. Mat.5, 269. 10.1038/s43246-024-00708-9
16
Helt-CameronJ.AllenP. J. (2009). Cisplatin ototoxicity in children: implications for primary care providers. Pediatr. Nursing35, 121–127.
17
HintonG.RoweisS. (2002). “Stochastic neighbor embedding,” in Proceedings of the 16th International Conference on Neural Information Processing Systems (Cambridge, MA, USA: MIT Press), 857–864.
18
HumesH. D. (1999). Insights into ototoxicity. analogies to nephrotoxicity. Ann. N. Y. Acad. Sci.884, 15–18. 10.1111/j.1749-6632.1999.tb00278.x
19
IkedaK.MorizonoT. (1989). Effect of albumin-bound furosemide on the endocochlear potential of the chinchilla: alleviation of furosemide-induced ototoxicity. Archives Otolaryngology–Head and Neck Surg.115, 500–502. 10.1001/archotol.1989.01860280098025
20
KimS.ChenJ.ChengT.GindulyteA.HeJ.HeS.et al (2024). Pubchem 2025 update. Nucleic Acids Res.53, D1516–D1525. 10.1093/nar/gkae1059
21
KolbergL.RaudvereU.KuzminI.AdlerP.ViloJ.PetersonH. (2023). G: profiler—interoperable web service for functional enrichment analysis and gene identifier mapping (2023 update). Nucleic Acids Research51, W207–W212. 10.1093/nar/gkad347
22
KrallingerM.ErhardtR. A.-A.ValenciaA. (2005). Text-mining approaches in molecular biology and biomedicine. Drug Discovery Today10, 439–445. 10.1016/S1359-6446(05)03376-3
23
LesterG. M.WilsonW. J.TimmerB. H. B.LadwaR. M. (2024). Audiological ototoxicity monitoring guidelines: a review of current evidence and appraisal of quality using the AGREE II tool. Int. J. Audiol.63, 747–752. 10.1080/14992027.2023.2278018
24
LiuT.HwangL.BurleyS.NitscheC.SouthanC.WaltersW.et al (2024). Bindingdb in 2024: a fair knowledgebase of protein-small molecule binding data. Nucleic Acids Res.53, D1633–D1644. 10.1093/nar/gkae1075
25
MaatenL. v. d.HintonG. (2008). Visualizing data using t-sne. J. Machine Learning Research9, 2579–2605.
26
MacQueenJ. (1965). “Some methods for classification and analysis of multivariate observations [c],” in Proc. of Berkeley Symposium on Mathematical Statistics and Probability, 281–297.
27
MiyoshiT.BelyantsevaI. A.SajeevadathanM.FriedmanT. B. (2024). Pathophysiology of human hearing loss associated with variants in myosins. Front. Physiology15, 1374901. 10.3389/fphys.2024.1374901
28
NagaiJ.TakanoM. (2014). Entry of aminoglycosides into renal tubular epithelial cells via endocytosis-dependent and endocytosis-independent pathways. Biochem. Pharmacology90, 331–337. 10.1016/j.bcp.2014.05.018
29
OhH.ParkM. K. (2025). Assessment and management of chemotherapy-induced ototoxicity in children. J. Audiology and Otology29, 79–85. 10.7874/jao.2025.00073
30
PakJ. H.KimY.YiJ.ChungJ. W. (2020). Antioxidant therapy against oxidative damage of the inner ear: protection and preconditioning. Antioxidants9, 1076. 10.3390/antiox9111076
31
PakenJ.GovenderC. D.PillayM.SewramV. (2020). Perspectives and practices of ototoxicity monitoring. S Afr. J. Commun. Disord.67, e1–e10. 10.4102/sajcd.v67i1.685
32
RizkH. G.LeeJ. A.LiuY. F.EndriukaitisL.IsaacJ. L.BullingtonW. M. (2020). Drug-induced ototoxicity: a comprehensive review and reference guide. Pharmacotherapy40, 1265–1275. 10.1002/phar.2478
33
RybakL. P.RamkumarV. (2007). Ototoxicity. Kidney Int.72, 931–935. 10.1038/sj.ki.5002434
34
RybakL. P.WhitworthC. A. (2005). Ototoxicity: therapeutic opportunities. Drug Discovery Today10, 1313–1321. 10.1016/S1359-6446(05)03552-X
35
SaltA. N. (2005). Pharmacokinetics of drug entry into cochlear fluids. Volta. Rev.105, 277–298.
36
ShannonP.MarkielA.OzierO.BaligaN. S.WangJ. T.RamageD.et al (2003). Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Res.13, 2498–2504. 10.1101/gr.1239303
37
SharpM. E. (2017). Toward a comprehensive drug ontology: extraction of drug-indication relations from diverse information sources. J. Biomedical Semantics8, 2. 10.1186/s13326-016-0110-0
38
SteygerP. S. (2021). Mechanisms of ototoxicity and otoprotection. Otolaryngologic Clinics N. Am.54, 1101.
39
SteygerP. S.CunninghamL. L.EsquivelC. R.WattsK. L.ZuoJ. (2018). Editorial: cellular mechanisms of ototoxicity. Front. Cell Neurosci.12, 75. 10.3389/fncel.2018.00075
40
SupekF.BošnjakM.ŠkuncaN.ŠmucT. (2011). Revigo summarizes and visualizes long lists of gene ontology terms. PloS One6, e21800. 10.1371/journal.pone.0021800
41
TanW. J. T.VlajkovicS. M. (2023). Molecular characteristics of cisplatin-induced ototoxicity and therapeutic interventions. Int. J. Mol. Sci.24, 16545. 10.3390/ijms242216545
42
WattsK. L. (2019). Ototoxicity: visualized in concept maps. Semin. Hear40, 177–187. 10.1055/s-0039-1684046
43
WhitworthC.MorrisC.ScottV.RybakL. P. (1993). Dose-response relationships for furosemide ototoxicity in rat. Hear. Research71, 202–207. 10.1016/0378-5955(93)90035-y
44
WishartD. S.GuoA.OlerE.WangF.AnjumA.PetersH.et al (2021). Hmdb 5.0: the human metabolome database for 2022. Nucleic Acids Res.50, D622–D631. 10.1093/nar/gkab1062
45
YoshimoriA.BajorathJ. (2020). The sar matrix method and an artificially intelligent variant for the identification and structural organization of analog series, sar analysis, and compound design. Mol. Inf.39, 2000045. 10.1002/minf.202000045
Summary
Keywords
database, drug side effects, drug-target network, literature mining, LLM, otoprotection, ototoxicity
Citation
del Moral-Morales A, Arguello‐Camarillo J, Castañón Bello JY, Chavez-Romero A, Medina-Franco JL, Nájera H and Pérez-Hernández G (2026) Mapping the otoactive landscape: an LLM-aided extraction of compounds and their targets. Front. Drug Discov. 6:1834905. doi: 10.3389/fddsv.2026.1834905
Received
20 March 2026
Revised
03 June 2026
Accepted
04 June 2026
Published
07 July 2026
Volume
6 - 2026
Edited by
Arif Nur Muhammad Ansori, Universitas Airlangga, Indonesia
Updates
Copyright
© 2026 del Moral-Morales, Arguello‐Camarillo, Castañón Bello, Chavez-Romero, Medina-Franco, Nájera and Pérez-Hernández.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Gerardo Pérez-Hernández, gperezh@cua.uam.mx; Aylin del Moral-Morales, aylindmm@gmail.com
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.