To establish a reliable baseline for model performance, we manually compiled a dataset of 70 peer-reviewed articles from PubMed describing toxicological assessments of various plants, encompassing both known toxic and non-toxic herbs. The dataset was balanced, with approximately half of the articles reporting toxicity concerns (Supplementary Table 2). This balanced dataset was used to evaluate the ability of different Natural Language Processing (NLP) models to distinguish between toxic and safe herbal extracts based on scientific text (Figure 1).

We selected three transformer-based NLP models—GPT-3.5 Turbo, GPT-4, and SciBERT—to represent a spectrum of architectures with varying training corpora, parameter scales, and domain specializations:

GPT-3.5 Turbo (OpenAI) was chosen as a representative large language model (LLM) optimized for speed and efficiency. Although it is a general-purpose model, it has demonstrated solid performance on natural language classification tasks and serves as a useful baseline for comparison against more advanced or domain-specific models.

Each model was prompted and evaluated on the same dataset using standardized input prompts to predict whether an abstract described a toxic or non-toxic plant extract. Model performance was assessed across four metrics: accuracy, precision, recall, and F-score.

Among the tested models, GPT-4 achieved the best overall performance, showing superior classification accuracy and balanced precision–recall trade-offs (Figure 1A). In contrast, GPT-3.5 Turbo demonstrated strong recall but lower precision, tending to over predict toxicity, while SciBERT exhibited good recognition of domain-specific terminology but lower generalization across diverse abstracts.

Statistical comparisons between model outputs were conducted using bootstrap confidence intervals for each performance metric. GPT-4 showed significantly higher accuracy and F-score compared to SciBERT (Supplementary Table 3), whereas other pairwise differences were not statistically significant.

Based on its consistent superiority across all evaluation criteria, GPT-4 was selected as the primary NLP model for the subsequent literature-based safety assessment of Withania somnifera. Its capacity to accurately interpret scientific context and distinguish subtle toxicological signals ensured a robust and reliable foundation for the meta-analysis of Ashwagandha’s safety profile (Figure 1B).

For training the model, we utilized a dataset of molecules known to have potential toxic effects, as reported in the SideR database.¹ This dataset provided a comprehensive resource of chemically diverse compounds with documented toxicity profiles, enabling the model to learn which molecular features are indicative of toxicity.

In addition to the SideR data, we trained the model on a broader dataset of 300,000 molecules, which included information on their interactions with 1,667 different molecular targets. These targets span a wide range of biological pathways and are relevant to the key toxicological endpoints we aimed to predict, such as mutagenicity, hepatotoxicity, and carcinogenicity.

The QSAR model used for predicting toxicity was developed internally for this study using the RandomForestClassifier implementation from the scikit-learn (sklearn) package (Supplementary Table 1). Random forests are ensemble learning algorithms that build multiple decision trees during training and combine their outputs to improve predictive accuracy and reduce the risk of overfitting. In this context, each decision tree learns to classify molecular structures according to their predicted toxicological outcomes, and the aggregated “majority vote” of the forest provides a robust final prediction. This approach is particularly well suited for QSAR tasks, as it can handle high-dimensional molecular descriptor data, capture nonlinear relationships between features and biological activity, and provide measures of feature importance that help interpret the drivers of predicted toxicity.

Using this extensive training dataset, the QSAR model predicted the biological activities of each molecule found in Ashwagandha based on their structural features. By analyzing these predicted activities, we were able to infer potential toxic effects, identifying compounds that might pose risks due to their biological target interactions.

The model achieved a high accuracy in predicting the toxicity of molecules based on their structure, demonstrating its effectiveness in identifying toxicological risks, for both predicting liver toxicity and reproductive toxicity (fetal, hormonal and thyroid related toxicity) (Figures 2A–C). This high level of accuracy was crucial for ensuring that the toxic predictions were reliable and actionable in the context of Ashwagandha’s safety evaluation.

We trained the model on a dataset comprising over 600 data points from more than 50 different plants. Each molecule was labeled according to the plant part in which it is predominantly found, such as the root, leaf, stem, or flower. The molecular structures of these compounds, encoded in SMILES format, were used as input for the model.

This diverse dataset enabled the model to learn structural and chemical patterns associated with the molecular composition of different plant organs. As a result, it could predict with high confidence the most probable localization of a given compound within the plant—particularly distinguishing molecules characteristic of the root from those predominantly found in aerial parts. This predictive capacity improved the precision of the subsequent QSAR analysis by allowing toxicity assessment to be specifically attributed to root- or leaf-derived molecules. A similar approach as the toxicity’s QSAR model was used for the prediction of plant parts, based on random forest (Supplementary Table 1).

The model achieved an accuracy of 0.84 in classifying molecules based on their part of origin and ROC AUC of 0.98 (Figure 3). The high ROC AUC value indicates excellent sensitivity and specificity, which translates into a very low number of false positives—i.e., compounds incorrectly classified as root-derived. This low error rate is critical for downstream analyses, as it reduces the likelihood of misattributing molecular origin and strengthens the reliability of the biological interpretations. This strong performance gave us confidence in the model’s ability to distinguish root-specific compounds from those found in other plant parts.

By applying this model to Ashwagandha, we were able to refine our safety analysis by identifying the specific molecular profiles of the root versus the rest of the plant. This enabled a more detailed and accurate comparison, especially in cases where the leaves might be used despite documented safety concerns.

To facilitate the comparison of the toxicity profiles between the different plant parts of Ashwagandha and a dataset of edible plants and herbal supplements, we developed a toxicity score. This score was calculated based on the predicted toxicity of each molecule in the specific plant part or in the whole plant depending on the scope of analysis. Specifically, the toxicity score was computed as the sum of the toxicity confidence values for all molecules present in the plant, divided by the total number of molecules. This approach provided an aggregate measure of toxicity, allowing for a direct comparison between Ashwagandha and other plants in terms of their overall toxicity risk. The score helped standardize the evaluation, enabling us to quantify the safety profile of Ashwagandha in the context of the broader dataset.

In the MeNow database, there are 2,766 plants classified as edible. To ensure a fair comparison with Ashwagandha, we selected edible plants that had a similar number of active molecules to those found in Ashwagandha. According to the MeNow database, Ashwagandha contains 338 unique molecules. For our comparison, we chose edible plants that have between 100 and 500 active secondary metabolites appearing in the MeNow database, resulting in a selection of 676 plants that matched this criterion.

By narrowing the comparison to plants with a similar molecular complexity, we ensured that the differences in safety profiles were not simply a reflection of the number of molecules present, but rather of their specific biological activities and potential toxicities.

In addition to edible plants, we compared Ashwagandha’s safety profile to that of other well-known herbal supplements. To compile a relevant dataset, we referenced a published list of herbal supplements for which detailed molecular composition data is available in the MeNow database (8). From this list, we identified 228 herbal supplements to include in our comparison dataset.

This comparative approach enabled us to evaluate Ashwagandha’s safety within the broader context of herbal medicine. By comparing its molecular and safety profiles with those of other edible plants and widely used supplements, we could assess Ashwagandha’s relative safety and identify any unique risks linked to its specific molecular composition.

We collected a total of 1,402 publications mentioning Ashwagandha from PubMed, covering the period from 1958 to August 2024. Out of these, six publications were withdrawn by the authors and were excluded from the subsequent text analysis. The remaining 1,396 publications formed the basis of our analysis. The dataset comprised 1,196 journal articles, including both comparative and evaluation studies; 183 reviews that summarized the findings of prior research on WS; 18 clinical trials and case reports, providing firsthand data on the effects of WS in human subjects; Five short news and letters offering brief insights or announcements related to WS. To maintain a high level of scientific rigor, we limited the scope of our analysis to articles indexed in PubMed, ensuring that the data came from sources with recognized scientific accuracy. The NLP analysis was performed using a large language model (LLM)–based system (GPT-4), employing the same prompts that were used during model selection. For cases where the model predicted potential toxicity, an additional set of prompts was applied to classify the type of toxicity (Supplementary Table 1).

According to the NLP analysis, 97% of the articles did not mention any toxic concerns for Ashwagandha (Figure 4). Of the 37 articles that did address potential toxicity, 32 articles dealt with cytotoxicity, specifically targeting cancerous cells; Nine articles mentioned potential liver toxicity, with five of these also discussing cytotoxicity; One article referred to a possible adverse effect on thyroid activity; No articles mentioned concerns for fetal toxicity and endocrine disruption. These findings suggest that while toxicity concerns for Ashwagandha are rare, the most frequently mentioned toxic effects relate to cytotoxicity in cancer treatment contexts, with some reports of liver toxicity and very limited mention of thyroid concerns (Supplementary Table 4).

As part of the NLP analysis, 32 publications specifically reported cytotoxic effects of Ashwagandha (Supplementary Table 4). Importantly, the majority of these studies focused on cytotoxicity in the context of cancer research, highlighting the potential of Ashwagandha and its bioactive compounds as anti-cancer agents rather than indicating general toxicity.

Several studies detailed selective cytotoxic activity against various cancer cell lines. For example, Halder et al. and Iguchi et al. demonstrated cytotoxic effects against melanoma and gingival carcinoma cells, respectively (1, 10), while Yadav et al. reported growth inhibition across five cancer cell lines (11). Pretorius et al. observed that Ashwagandha extracts were non-toxic at low concentrations but could exert cytotoxic effects at higher doses, particularly against MRC-5 cells (12). Other studies, including Liu et al. and Choudhary et al., highlighted cytotoxicity against lung cancer cells (13, 14), and Senthil et al. reported anticancer activity of leaf extracts against human gastric adenocarcinoma cells (15). Similarly, Siddique et al. described strong cytotoxic activity against liver and breast cancer cell lines (16), and Al-Fatimi et al. reported notable anti-breast cancer effects of a fungus growing on Ashwagandha root (17).

The cytotoxic effects are largely attributed to withanolides, particularly Withaferin A, a key bioactive compound. Studies by Devi et al., Xia et al., and Misra et al. demonstrated Withaferin A’s capacity to induce apoptosis and enhance cancer therapy (18–20). Mechanistic investigations, including work by Jilani et al. and Lv & Wang, revealed disruption of cancer cell division and inhibition of G2/M checkpoint proteins, further supporting selective cytotoxicity toward malignant cells (21–23). Comprehensive reviews (Kumar et al., Albahri et al.) provide an overarching summary of Ashwagandha’s anti-cancer potential (24, 25).

In summary, cytotoxicity observed in these studies is largely targeted to cancerous cells and does not indicate safety concerns for general consumption of Ashwagandha. Rather, these findings underscore its therapeutic potential in oncology.

Our NLP analysis identified nine publications reporting potential liver toxicity associated with Ashwagandha (Supplementary Table 4). While Ashwagandha is generally considered safe, these studies highlight rare, context-specific cases of hepatotoxicity. Many reports involved confounding factors such as pre-existing liver conditions, concomitant use of other supplements, or high dosages, limiting the generalizability of these findings.

For example, case series from India indicate that liver injury following Ashwagandha use is uncommon and often occurs in individuals with underlying health conditions or concurrent supplement intake (4). Another study noted that improper preparation or excessive consumption could contribute to hepatocellular stress, but routine use remains low-risk (26). Systematic reviews of herb-induced liver injury (HILI) similarly classify Ashwagandha-related hepatotoxicity as rare and context-dependent (27).

Clinical monitoring studies observed only mild, reversible liver enzyme elevations in a small number of participants taking Ashwagandha root extracts, which normalized upon discontinuation (28). Animal studies indicate that liver damage occurs only at extremely high doses, far above typical human consumption (29). Individual case reports, including instances in healthy adults, suggest that hepatotoxicity is uncommon and generally reversible (30–32).

Overall, the literature indicates that while isolated liver toxicity cases exist, Ashwagandha’s hepatotoxic risk for the general population at recommended dosages is minimal. These findings reinforce the importance of proper dosing and consideration of pre-existing conditions in supplement use.

Out of nearly 1,400 PubMed-indexed articles analyzed in our NLP study, only one article raised concerns about thyroid toxicity linked to Ashwagandha. This study, Thyroid Dysfunction Induced by Ashwagandha: A Case Review, detailed a single case of thyroid dysfunction in an individual who developed abnormal thyroid activity after taking Ashwagandha supplements (33). The symptoms subsided once the individual stopped taking the supplement, suggesting a potential link between Ashwagandha and thyroid function in this particular case.

However, it is important to note that this study represents just one isolated case report out of nearly 1,400 articles. Given that none of the other analyzed publications mentioned thyroid toxicity, the evidence for a widespread concern is virtually negligible. The case highlighted in the study involved a unique scenario, and the authors themselves acknowledged the need for more research to determine whether the findings are an anomaly or indicative of a broader risk.

Considering the vast body of research that does not report any thyroid-related issues, the likelihood of Ashwagandha causing thyroid toxicity appears to be extremely low, particularly for individuals without pre-existing thyroid conditions.

Since it is challenging to draw definitive conclusions solely from isolated reports, we expanded our analysis by comparing the literature on Ashwagandha’s toxicity with other edible plants and herbal supplements. For this purpose, we prompted the NLP models on a dataset of 331 edible plants and 171 herbal supplements. This dataset was based on the larger pool presented in the materials and methods section, containing 676 edible plants and 227 herbal supplements, but was further restricted to botanicals with sufficient literature to support a robust analysis.

For each plant, the NLP models analyzed articles for various toxicity parameters, including general toxicity (including cytotoxicity), cytotoxicity, liver toxicity, endocrine disruption, thyroid toxicity, and fetal toxicity. For each toxicity type, we computed an average score for each plant. The results showed that Ashwagandha was below the average of both edible plants and herbal supplements across all categories of toxicity, except for cytotoxicity (Figure 5).

In terms of cytotoxicity, Ashwagandha’s results fell between the averages for edible plants and herbal supplements, indicating that its cytotoxicity is not unusually high but rather consistent with its known use in anti-cancer contexts. Importantly, none of these differences were statistically significant.

Additionally, a similar study presented in the article Herbal Supplements and Hepatotoxicity reviewed various herbal supplements for Herbal-Induced Liver Injury (HILI) (34). The review was conducted by expert reviewers, who categorized supplements as hepatotoxic based on the thoroughness of case reports. These reports had to include specific details such as serum biochemistry, liver biopsy findings, and well-documented clinical courses of injury. In their analysis, Ashwagandha was classified as negative for HILI, meaning no cases met the criteria for hepatotoxicity.

We compared our liver toxicity scores to the classifications made in the MDPI article, specifically contrasting herbs they classified as HILI-positive versus HILI-negative. The herbs classified as HILI-positive had significantly higher liver toxicity scores in our analysis (p-value = 8.38e-05), further validating our approach (Figure 6). In contrast, Ashwagandha’s liver toxicity score was consistent with herbs classified as HILI-negative, supporting the conclusion that Ashwagandha is not a significant liver toxicity risk (Figure 5).

The results of our NLP analysis indicate that Ashwagandha’s toxicity profile is comparable to other edible plants and herbal supplements, with no statistically significant differences across various toxicity parameters. Ashwagandha consistently ranked below average for general toxicity, liver toxicity, endocrine disruption, thyroid toxicity, and fetal toxicity. While its cytotoxicity was slightly higher, this aligns with its known anti-cancer properties and does not raise concerns for general safety. Furthermore, our comparison with HILI-positive herbs reinforced that Ashwagandha poses minimal risk of liver toxicity, as confirmed by its negative classification for HILI in the literature.

As outlined in the methods, we predicted the presence of molecules in different parts of Ashwagandha using their SMILES representations. Of the 338 molecules identified in Ashwagandha, our model predicted that 79 are likely to be found in the root with a confidence threshold of 0.7 (Supplementary Table 5).

To further explore the relationships between these molecules, we used the Tanimoto similarity index and Morgan fingerprints to construct a phylogenetic tree, which allowed us to cluster the molecules into six distinct groups based on their structural similarities. By Tanimoto similarity, the distance between clusters represents the average distance between the molecules on the clusters. Each of these groups, marked as A-F in Figure 7, was then classified according to their molecular class using the Coconut database, as shown in Table 1. Clusters D and F correspond to withanolides and their derivatives. Withanolides are a group of naturally occurring steroidal lactones predominantly found in the Withania genus, in particular in Withania somnifera. They are well known for their diverse bioactivities, including anti-inflammatory, anticancer, and neuroprotective properties.

We found that the largest clusters corresponded to withanolides and their derivatives, which are well-known for their bioactive properties, particularly in stress response and anti-cancer activities. Additionally, we identified steroid lactones, flavonols, and curcuminoids in smaller clusters, all of which contribute to the diverse bioactivity associated with Ashwagandha’s therapeutic uses. These classifications help to provide a clearer understanding of the molecular diversity in Ashwagandha’s root, highlighting its potential as a source of various bioactive compounds with different structural properties.

Using the approach outlined earlier, we scanned all 338 identified molecules from Ashwagandha to assess their potential for liver toxicity. Of these, only two molecules, L-Ala (PubChem CID 602) and Tropine (PubChem CID 8424), were predicted to exhibit liver toxicity, and even these predictions were made with medium confidence (0.7 and 0.72 respectively). L-Ala (L-Alanine) is a naturally occurring amino acid, typically found in proteins and known to play a role in energy metabolism. While generally considered safe, some studies suggest that in excess or under certain metabolic conditions, it could contribute to liver strain (35). Tropine, a tropane alkaloid, is found in several plant species and is known for its anticholinergic properties. Some alkaloids of this class have been associated with toxicity, particularly when consumed in large quantities, potentially leading to adverse effects on the liver (36). Notably, neither of these molecules was predicted to be present in the root, based on the threshold we selected (0.7 and above) (Supplementary Table 5).

In contrast, 143 molecules from Ashwagandha were predicted to be non-toxic with very high confidence, as their liver toxicity confidence was below 0.2. This group included several key bioactive compounds such as Withanolides, including: Sitoindoside IX (PubChem CID 78173488), Withanoside II (PubChem CID 163025782), Withaferin A (PubChem CID 580064), and flavonoids, including: Rutin (PubChem CID 5293655), Cosmetin (PubChem CID 5385553), Rhamnetin (PubChem CID 5281691) that are depicted in Figure 8.

These findings indicate that the safety profiles of the molecules differ significantly between the root and the whole plant (Figures 9A,B). As mentioned, for the whole plant, only two molecules (L-Ala and Tropin) are predicted to be liver toxic. An additional 90 molecules (27% of the total molecules) are predicted to exhibit low liver toxicity, with confidence levels between 0.5 and 0.7. By contrast, for the root, all of the molecules are predicted to be safe for the liver, with toxicity confidence levels below 0.5. Even more striking, 96% of the molecules in the root were predicted to be extra safe, with liver toxicity confidence scores below 0.2. Overall, very few molecules in Ashwagandha are predicted to pose a liver toxicity risk, and none of these are predicted to be present in the root. These findings suggest that the root of Ashwagandha is particularly safe for liver health, further supporting its traditional use in herbal medicine.

Next, we analyzed whether the observed differences between the predicted liver toxicity of the root and non root molecules is statistically significant. For this, a non-parametric Wilcoxon test was performed. Root molecules appeared to be significantly safer for the liver than their non root counterparts with a p-value of 2.24e-12 (Figure 10).

To further validate Ashwagandha’s safety, we compared its molecular safety profile to a set of 676 edible plants and 227 herbal supplements. For this analysis, we computed a toxicity score that accounted for the proportion of molecules with predicted liver toxicity (liver toxicity confidence above 0.5) and weighted this based on the confidence of the predictions.

When compared to the broader set of edible plants and supplements, Ashwagandha emerged as one of the plants with a strong safety profile, performing better than the average in our dataset. This was true when considering all molecules from Ashwagandha, but the root, in particular, displayed an even stronger safety profile (Figures 11A,B). These findings align with the results obtained from our previous NLP analysis, further reinforcing Ashwagandha’s reputation as a safe herbal supplement.

Our approach was validated through the successful identification of known toxic molecules in certain herbs, with our model’s predictions supported by literature evidence. For instance, in Comfrey, one of the molecules flagged by our model as toxic was a protoberberine alkaloid (PubChem CID 440229) that was also highlighted in the literature (37). The toxicological effects of Comfrey have been well-documented, particularly in studies on genotoxicity and carcinogenicity. Our model identified three toxic molecules in Ephedra, including two benzenes and synephrine. Synephrine, in particular, has been linked to toxic effects, especially when used in combination with other substances such as ephedrine and caffeine (38). This finding is in line with concerns about the hepatotoxicity of weight-loss products containing Ephedra. Known for its hepatotoxic potential, Kava was another herb where our model successfully identified toxic molecules. Interestingly, the same toxic molecule found in Comfrey was also present in Kava. According to Olsen et al., the toxicity in Kava is also believed to be related to alkaloids (39).

This comparative analysis underscores Ashwagandha’s favorable safety profile, particularly when focusing on the root, and supports the effectiveness of our model in identifying toxic molecules in other herbs. By successfully predicting known toxic components in plants such as Comfrey, Ephedra, and Kava, our approach demonstrates its reliability and reinforces Ashwagandha’s standing as a safe herbal supplement.

In addition to assessing liver toxicity, we carried out a similar analysis to evaluate the reproductive toxicity of molecules in Ashwagandha. By reproductive toxicity, we refer to potential activity in hormonal disruption, thyroid toxicity, and fetal toxicity. Our analysis revealed a significant difference in the safety profiles between the root and the aerial parts of the plant (Supplementary Table 5).

Regarding the safety of the root, none of the 79 molecules predicted to be present in the root were identified as having reproductive toxicity. In fact, nearly all (78 out of 79) had a toxicity score below 0.2, meaning their safety confidence was above 0.8, indicating that the root of Ashwagandha is extremely safe in terms of reproductive toxicity.

In contrast, two molecules in the aerial parts of Ashwagandha were predicted to exhibit reproductive toxicity (Figure 12)—Tropin (PubChem CID 8424), with a toxicity score of 0.91, is a tropane alkaloid known for its anticholinergic properties and potential toxicity. Tropane alkaloids have been linked to toxicity concerns, especially when consumed in higher doses, with effects ranging from hormonal disruption to nervous system impacts (40). Theophyllin (PubChem CID 2153), with a toxicity score of 0.85, is a xanthine derivative commonly found in tea leaves. While theophyllin is used therapeutically to treat respiratory diseases, it has been associated with side effects, including potential reproductive toxicity at high concentrations (41).

Additionally, 47 molecules in the aerial parts were identified with a low toxicity potential (toxicity confidence between 0.5 and 0.7). Despite this, the aerial parts also contained 152 molecules that were non-toxic with high confidence, including notable bioactives such as the Withanolides Withacoagin (PubChem CID 14236708), Withastramonolide (PubChem CID 73800706), and 12-Deoxywithastramonolide (PubChem CID 53398767); and the Organooxygen Compounds Heriguard (PubChem CID 348159) and Withanamide H (PubChem CID 72951313) depicted in Figure 12 and Supplementary Table 5.

To quantify the differences in safety profiles between the root and aerial parts, we conducted a Wilcoxon test. The test confirmed that the differences in toxicity predictions between the root and aerial molecules were statistically significant (p-value = 2.31e-15), further underscoring the safety of the root compared to the rest of the plant (Figure 13).

To provide further context for Ashwagandha’s reproductive toxicity profile, we compared its safety to our extended dataset of edible plants and herbal supplements. This analysis offered valuable insights into how Ashwagandha ranks among commonly consumed botanicals in terms of reproductive safety, considering hormonal disruption, thyroid toxicity, and fetal toxicity.

When evaluating the whole plant, Ashwagandha’s reproductive safety profile was found to be comparable to the average safety profiles of both edible plants and herbal supplements. This means that while the whole plant does not stand out as particularly concerning in terms of reproductive toxicity, it falls within the expected range of safety found across a variety of other plants and supplements used in traditional and modern herbal medicine (Figures 14A,B).

However, when focusing specifically on the root of Ashwagandha, the results were markedly different. The root was found to be among the safest in the entire dataset, outperforming the vast majority of the edible plants and herbal supplements in terms of reproductive safety. This is particularly significant as the root contains no molecules predicted to pose reproductive toxicity risks, reinforcing its historical and traditional use in Ayurvedic medicine for its safety and efficacy (Figures 14A,B).

The safety score for Ashwagandha’s root reflected high confidence in the absence of reproductive toxicity, positioning it as an exceptionally safe option among herbal remedies. This stark difference between the safety profile of the root and that of the whole plant further underscores the importance of focusing on the specific parts of plants when assessing safety and toxicity, as the root’s lack of toxic molecules differentiates it from the aerial parts, which did contain some molecules with potential reproductive toxicity.

Our findings echo the results of earlier analyses and further validate Ashwagandha’s long-standing reputation for being a safe and reliable herbal supplement, particularly when derived from the root. In conclusion, Ashwagandha’s root exhibits a superior reproductive safety profile, making it one of the safest options available among both edible plants and widely used herbal supplements.