Recent epidemiological studies have shown a strong correlation between consuming Ultra-Processed Foods (UPFs) and an increased risk of developing various chronic diseases (Jardim et al., 2021). The use of artificial dyes in the industry remains a topic of considerable controversy and debate among both scientists and consumers (Rambler et al., 2022; Amchova et al., 2015; Debras et al., 2022). While food additives at acceptable levels are deemed safe for human consumption, their widespread presence and conflicting data suggest the potential for toxicological risks to both humans and marine organisms (Tkaczyk et al., 2020; Li et al., 2024). The ADI, expressed in mg/kg/day, represents the maximum amount of a substance a person can ingest daily throughout their lifetime without causing harm. Estimating the adequate ADI level is challenging, especially since many consumed products contain the same additives. It is widely acknowledged that children are the demographic most at risk of exceeding the recommended daily intake of food additives. Numerous food additives are found in products that are particularly popular among children, such as sweets, flavoured drinks, ice cream, sausages, and fast food. Furthermore, there is a lack of conclusive evidence regarding the impact of food additives, environmental contaminants, cosmetic ingredients, and pharmaceuticals during the prenatal period and early childhood. Due to numerous reports of adverse effects on consumers, the list of synthetic colours is being systematically updated (Amchova et al., 2015; Østergaard and Knudsen, 1998; Soni et al., 2024). On the other hand, data on the exposure of aquatic organisms to colourings is lacking in the world literature. Children are at greater risk of developing chronic diseases resulting from early exposure to environmental substances. However, with access to advanced toxicity prediction methods and new test models (zebrafish), safety studies of synthetic dyes are warranted.

QY belongs to the group of quinophthalone dyes. Commercially available as a mixture of monosulphonic, disulphonic, and trisulphonic acid derivatives, QY is used in confectionery, isotonic, and carbonated drinks. It can cause allergies and hyperactivity in children, as well as potentially be mutagenic and genotoxic (Chequer et al., 2015; Chequer et al., 2017; McCann et al., 2007; Scientifc Opinion on the re-evaluation of Quinoline Yellow, 2009). On the other hand, a study conducted in rats where QY was administered subcutaneously did not demonstrate evidence of carcinogenicity. The available in vivo oldest evidence does not indicate a genotoxic effect of quinoline yellow. Several methods have been used in in vitro and in vivo testing, including bacterial mutation assays, the L5178Y mouse lymphoma gene mutation assay, and the NMRI mouse micronucleus assay (Scientifc Opinion on the re-evaluation of Quinoline Yellow, 2009). In 2009, the European Food Safety Authority (EFSA, 2025; Scientifc Opinion on the re-evaluation of Quinoline Yellow, 2009) panel reduced the Acceptable Daily Intake (ADI) of QY from 10 mg/kg/b.w. to 0.5 mg/kg/b.w. A study examining the impact of the combination of food colours known as ‚Southampton’ (Tartrazine, Quinoline Yellow, Sunset Yellow, Ponceau 4R, Allura Red AC, Carmoisine) and Sodium Benzoate revealed an augmentation in behavioural activity among children (McCann et al., 2007). Nevertheless, there is a certain degree of ambiguity in interpreting the outcomes of these subsequent tests when evaluating the safety of quinoline yellow. This discrepancy can be attributed to the fact that the studies above used a test substance containing a high proportion of the monosulphonate component, ranging from 85% to 91%. In contrast, the specifications for QY intended for food use indicate that disulphonate is the primary component (over 80%), with monosulphonate present only in minimal amounts (15%). These compositional differences could significantly affect the safety assessment, highlighting the need for further research. Recent studies by Macioszek and Kononowicz (Macioszek et al., 2004) suggest that QY may have clastogenic and/or mutagenic properties, potentially causing DNA damage. An additional concern is the widespread use of QY in pharmaceuticals and cosmetic products. Although regulatory guidelines set permissible levels of synthetic dyes in individual products, they often fail to account for cumulative exposure, including that originating from environmental sources.

The Danio rerio has been demonstrated as an invaluable tool for the expedient screening of chemical toxicity, as it identifies numerous phenotypic abnormalities (McColl et al., 2011). Furthermore, Danio rerio retains key vertebrate traits such as high fecundity, rapid development, and translucent juveniles, all of which contribute to the speed and ease of experimentation compared to mammalian models. Despite specific anatomical and physiological differences characteristic of aquatic species, Danio rerio possesses most organs analogous in structure and function to those of humans (MacRae and Peterson, 2015).

Nevertheless, the primary challenge remains in the labour-intensive morphometric analysis of the acquired microscopic images, which limits the attainment of maximum throughput. The application of machine learning tools holds the potential to facilitate the development of models capable of distinguishing between the morphological characteristics of diverse organisms, including Danio rerio (Lin et al., 2018). Initial studies concentrated on identifying specific embryonic phenotypes, such as mortality and developmental stages. However, processing large datasets and distinguishing between numerous sublethal morphological abnormalities continue to pose significant challenges. Recent literature on Danio rerio has primarily focused on specific organs (Dong et al., 2023). Furthermore, researchers are interested in quantitative measurements, such as body length and curvature. Consequently, it is crucial to develop tools that can simultaneously capture the most prevalent phenotypes. The application of machine learning to replace manual measurements would undoubtedly enhance the applicability and efficiency of the zebrafish model. Convolutional Neural Networks (CNNs) are the example of such model commonly employed for image recognition due to their ability to efficiently capture spatial hierarchies in images through convolutional layers that detect edges, textures, and patterns. The CNN architecture diminishes the necessity for manual feature extraction, enabling the network to learn pertinent features directly from raw images. Moreover, CNNs’ capacity to handle extensive volumes of high-dimensional data renders them highly effective for intricate image classification tasks (Gu et al., 2018; Li et al., 2022).

Animal tests’ high costs and time-intensive nature pose significant challenges to evaluating chemicals on domestic and international markets. In silico techniques, such as computer modelling and analysis, offer an alternative by predicting the toxicity of chemicals based on correlations between molecular properties and biological activity. The primary advantage of in silico methods lies in their utility during early research stages. Virtual screening of molecular libraries facilitates the rapid identification of promising structures from thousands of potential candidates. For instance, these methods can eliminate molecules with specific structural alerts (toxicophores) that may indicate a particular toxic effect. In silico techniques are distinguished by their efficiency, speed, low cost, and precision, making them particularly valuable prior to synthesising chemical compounds, including potential medicinal substances (Segall and Barber, 2014). In recent years, numerous models for toxicity prediction have been developed to support risk assessment, along with open-access websites that leverage machine learning and structural alerts, all of which are freely available (Cheng et al., 2012). Quantitative Structure-Activity Relationship (QSAR) and Quantitative Structure–Toxicity Relationship (QSTR) models, which predict biological and/or physicochemical properties from the structural parameters of a chemical compound, constitute a well-established approach to chemical data analysis. They are particularly valuable for predicting various toxicity indices, including mutagenicity, carcinogenicity, and acute toxicity (Kar and Leszczynski, 2019).

This study investigated the potentially toxic effects of QY in a model of Danio rerio (zebrafish). The scope of the work included conducting an experiment using different concentrations of QY, assessing survival, embryo morphology, and possible developmental defects. In silico analysis of QY toxicity was also conducted using the ADMET Predictor software. Additionally, a machine learning protocol was established to analyze toxic endpoints in zebrafish embryos treated with QY.

In silico predictions based on the structures of QY indicated that the compound may cause endocrine disruption, chromosomal aberrations, reproductive toxicity and mutagenic effects in the Ames test, as well as skin and respiratory sensitisation in animal models. Furthermore, no hepatotoxic effects were observed, although elevated γ-glutamyl transferase levels were noted in the presence of QY. In a study conducted by Damayanti et al. (2015), the researchers employed in silico methods to predict the acute toxicity (LD50), mutagenicity, carcinogenicity, reproductive toxicity, chronic toxicity (NOEL), metabolite toxicity for synthetic additives that act as antioxidants. In a separate study, butylhydroxytoluene (BHT) was employed as a case study to examine the efficacy of computational techniques, including reverse screening and molecular docking, in identifying protein-ligand interactions of artificial additives based on their toxicological effects (Tortosa et al., 2020). In silico analyses can be employed to ascertain the characteristics of synthetic additives and to examine their functional mechanisms or potential adverse effects. It is worthy of note that this may be relevant in cases where experimental studies are clearly lacking, as is the case with BHT. In the search for xenoestrogens in synthetic additives, an integrated in silico and in vitro approach has been employed, as outlined in (Amadasi et al., 2009). In recent years, there has been a notable increase in the number of studies examining the toxic effects of food additives, with the brindle danio serving as a model organism in many of these investigations. The toxic effects of food additives in the zebrafish model have been found to be primarily manifested as dose-dependent developmental toxicity. The findings of the Sarmah et al. (2020) analysis demonstrate the developmental toxicity of the synthetic antioxidant BHT. Other studies employing the zebrafish model have indicated that exposure to sodium dehydroacetate (DHA-S), an approved preservative commonly added to processed foods, may represent a potential cardiovascular risk factor. In a comprehensive evaluation of the adverse effects of propylparaben and methylparaben on the early developmental stages of Bereketoglu and Pradhan (2019) observed a range of abnormalities, including spine and pigmentation defects, pericardial edema and reproductive toxicity. Furthermore, the altered expression of the androgen receptor (AR) and estrogen receptor 2 alpha (ESR2a) indicated anti-androgenic and estrogenic effects of parabens in zebrafish.

Synthetic dyes are frequently employed as colourings, particularly in confectionery products. Consequently, it can be hypothesised that they may have more adverse health effects in children than in adults. In view of the toxicity of azo dyes, EU countries regularly undertake reviews and revisions of their ADI values. The developmental toxicity of four azo dyes, including Tartrazine, Sunset Yellow, Amaranth Red and Allura Red, has been evaluated using zebrafish embryos. At concentration levels of 5–50 mM, it has been demonstrated that azo dyes can impede the process of leaving the chorion and induce developmental abnormalities, including cardiac edema, a slowed heart rate, a swollen yolk follicle, a curvature of the spine and a tail deformity. The embryos demonstrated complete lethality at 100 mM of the analysed azo dyes. Conversely, impediments to the embryos’ departure from the chorion and aberrant developmental outcomes were observed at concentrations exceeding the ADI (Jiang et al., 2020). The dose-dependent toxic effects of the caramel colouring agent E150d on the embryos of zebrafish included disruption of chorion exit, survival, phenotype, heartbeat, and swimming ability, as well as damage to skeletal muscle and the pericardial cavity. These effects were observed at varying doses of the food dye sulphite ammonia caramel (Capriello et al., 2021; Hou et al., 2023).

The findings of the present study indicate that QY induces lethal alterations in embryos of the zebrafish (Danio rerio). The two endpoints observed at the 96-h exposure period were coagulation and the absence of a heartbeat. Embryos exposed to a QY solution at concentrations of 0.1 mg⋅mL⁻¹ and above exhibited a higher mortality rate. At a concentration of 0.5 mg⋅mL⁻¹ mg, the survival rate was 46.47% for embryos. The LC50 parameter value was found to be 0.64 mg⋅mL⁻¹. In a study conducted by Duy-Thanh et al. (2022), the embryotoxic and teratogenic effects of QY were analysed using a zebrafish model. The mortality data yielded divergent results from those of the present study. The survival rate remained above 90% up to a concentration of 0.5 mg⋅mL⁻¹. The authors of the publication report that the median lethal concentration level was 6.89 mg⋅mL⁻¹. Joshi and Katti (2018) conducted embryotoxicity studies using the zebrafish model to assess the toxicity of Tartrazine, a compound belonging to the azo-structured chemical class. An LC50 value of 15.7 mg⋅mL⁻¹ was determined. Moreover, the aforementioned authors conducted a similar study, this time focusing on orange yellow. The LC50 value was approximately 19.3 mg⋅mL⁻¹. The substance under examination in this study displays a greater toxicity potential than other azo dyes. Among the embryotoxic effects observed, developmental defects such as yolk sac swelling and reduced eye size were noted. These alterations were observed at concentrations of 0.5 mg⋅mL⁻¹ and 0.75 mg⋅mL⁻¹, respectively. The test dye was observed to contribute to the development of cardiac emphysema and blood stasis, while maintaining circulation and cardiac activity. In the aforementioned study, standard substances devoid of impurities were subjected to analysis. The occurrence of yolk sac edema was observed in over 50% of the larvae at concentrations of 0.5 mg⋅mL⁻¹. Conversely, isolated instances of under-eye were documented at concentrations as low as 0.02 mg⋅mL⁻¹, with this phenotype occurring in nearly 100% of larvae at concentrations of 0.5 mg⋅mL⁻¹. The swelling of the yolk sac may be associated with impaired absorption of the nutrients it contains (Ramos‐Souza et al., 2023; Kashyap et al., 2007; (Harding et al., 2021). Our study identified a form of blood stasis that has not previously been described in the scientific literature. A recent transcriptomics study has demonstrated that exposure to a non-sulphonated form of QY results in a reduction in the expression of metabolic genes in zebrafish embryos. Of particular note is the disruption of the retinoic acid signalling pathway, which suggests a potential impairment in eye development. The observation that a widely used food additive may interfere with nutrient metabolism, even at sub-ADI exposure levels, warrants a critical re-evaluation of its current safety thresholds (0.5 and 3 mg/kg b.w., respectively, set by EFSA and JECFA) (Refined exposure assessment for quinoline yellow, 2015). Further studies are required to elucidate the toxicity of synthetic additives using the model organism Danio rerio. A paucity of definitive experimental data exists in the global scientific literature on the safety of chemical additives, including their interaction effects during the embryonic development of the organism. Recent studies indicate that quinoline yellow can induce protein aggregation (Khan et al., 2019). One in vitro study demonstrated that the dye may modulate the expression of 21 genes involved in DNA repair, raising concerns about its toxicological implications (Chequer et al., 2015). Additionally, quinoline yellow acts as a potent agonist of the aryl hydrocarbon receptor (AHR), induces CYP1A1 expression, and inhibits estrogen receptor signaling through AHR-dependent pathways, suggesting potential endocrine-disrupting effects (Tarnow et al., 2020)

The application of machine learning to replace manual measurements would undoubtedly enhance the applicability and efficiency of zebrafish model. Machine learning approaches are becoming increasingly widespread and are now present in most areas of research. CNN employ the concept of deep learning, wherein the augmentation of the number of hidden neuron layers within the network leads to a concomitant enhancement of overall image recognition accuracy. To further augment image recognition capabilities with CNNs, techniques such as transfer learning and attention mechanisms can be employed. In transfer learning, a pretrained recognition model (for instance, an ImageNet CNN trained on a comprehensive data set of images) is selected and augmented with additional layers or modified to adopt a distinct model architecture. This approach is particularly beneficial in scenarios where the availability of domain-specific training data is constrained, such as in the classification of biomedical images. Kim et al. (2022) reviewed 121 publications on the application of transfer learning in medical image recognition. The study outlines various transfer learning approaches, where a pretrained CNN can be extended with additional convolutional layers or used as a feature extractor, with the extracted features then applied to train a separate machine learning model, such as a Support Vector Machine (SVM). Additionally, the pretrained model may or may not be fine-tuned with new data. As indicated by Kim et al. (2022), the pretrained ResNet50 CNN model achieved superior classification results compared to two other models, AlexNet and VGG, when trained on the ImageNet dataset. As presented by our experiments, the base ResNet50 model also provides the best classification results.

Additionally, to augment classification accuracy, we incorporated the attention mechanism into CNNs. The attention mechanism in machine learning emulates the human capacity to concentrate on specific portions of information to enhance comprehension. Tsotsos et al. (1995) and Niu et al. (2021) identify two types of human attention. The first type of attention is often induced by a stronger stimulus that draws human focus, while the second involves intentional concentration on a task. In general, machine learning implements the latter type of attention mechanism. Previous research has indicated that applying an attention mechanism can improve the quality of an image recognition task (Rodriguez et al., 2020). Specifically, in this work we decided to adapt the Convolutional Block Attention Mechanism (CBAM). This attention mechanism comprises two modules: the Channel Attention Module and the Spatial Attention Module. As outlined by the authors of (Woo et al., 2018), CBAM has been demonstrated to enhance image classification outcomes. In our research, we applied CBAM to improve the detection of abnormal blood stasis in embryos. Blood stasis is frequently distinguished from other pathological states by the presence of minute blood clots in the images. However, our experiments indicated that a neural model lacking an attention mechanism encountered difficulties in accurately identifying these blood clots. In summary, we posit that the developed model can effectively support the identification of pathological states in Danio rerio embryos induced by specific concentrations of QY. The results suggest that these concentrations of QY may indeed have deleterious effects.

The in silico study presents evidence suggesting that QY may pose potential health risks, including developmental toxicity, allergic activity, and possible endocrine disruption. These findings indicate that the current safety assessments may be insufficient and warrant a more rigorous evaluation of QY, potentially accompanied by stricter regulatory measures. The synthetic colour QY, has been observed to induce lethal changes in Danio rerio embryos, at concentrations ranging from 0.5 mg⋅mL⁻¹ to 2 mg⋅mL⁻¹. The lethality of Danio rerio is evidenced by the formation of cardiac dysfunction and coagulation. The LC50 of QY was determined to be 0.64 mg⋅mL⁻¹. The administration of QY at concentrations above 0.5 mg⋅mL⁻¹ has been observed to elicit a dose-dependent pericardial edema, swollen and necrosed yolk sac, blood stasis and reduced eye size. Attention-based CNNs, combined with transfer learning, can be employed for the classification of developmental toxicity in zebrafish embryos. The model demonstrated consistent performance in identifying healthy samples but exhibited challenges in distinguishing between disease states. Notably, the applied AI model also suggests that the employed doses of QY induce pathological states in embryos. The findings of these studies on the embryotoxic potential of QY underscore the necessity for further experiments to gain a comprehensive understanding of the health effects of this dye, particularly within the context of embryonic development. Consequently, further research is required to provide consistent data, identify suitable alternatives to chemical colours, and assess their safety.