Identifying chemical compounds in foods and assaying their bioactivities significantly contribute to promoting human health. In this work, we propose a machine learning framework to predict 101 classes of health effects of food compounds at a large scale. To tackle skewedness of class distributions commonly encountered in chemobiological computing, we adopt random undersampling boosting (RUSBoost) as the base learner. In this framework, all chemical molecules including food compounds, natural products and drugs are encoded into MACCSKeys similarity spectrums to define the fingerprint similarities of functional subgroups between molecules of interest with predefined template molecules. Five-fold 5-fold cross validation shows that RUSBoost learners encouragingly reduces model biases. Independent tests on external data show that the proposed framework trained on food compounds generalizes well to natural products (0.8406 ~ 0.9040 recall rates for antibacterial, antivirals, pesticide and anticancer effects) and drug molecules (0.789 ~ 0.9690 recall rates for antibacterial, antiviral, antineoplastic and analgesic effects). Furthermore, dozens of novel effects have been validated against recent literature, convincingly demonstrating knowledge transferability between food compounds, plant or microbial natural products and drugs. Especially, evidences show that the proposed framework helps us to repurpose drugs or find lead compounds for anticancer therapies and bacterial drug resistance. Lastly, we attempt to use the proposed framework to unravel beneficial and risky health effects of food flavor compounds, which potentially benefits recipe composing.
Research Article
Predicting health effects of food compounds via ensemble machine learning
https://doi.org/10.21203/rs.3.rs-2991763/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Food ingredient bioactivities
Natural products
MACCS fingerprints
Ensemble learning
Multi-label learning
Transfer learning
Machine learning
Foods play important roles in providing nutritional necessities and maintaining physiological health. As functional ingredients, food bioactive compounds have clinically demonstrated to take many health-promoting effects such as antioxidant [1], anti-inflammatory [2, 3], anticancer [4, 5], cardiovascular health benefits [6] and regulation of metabolic disorders [7]. Nevertheless, foods also occasionally contain some risky compounds that originate from chemical contaminants [8], food additives [9] or plant and microbial natural products [10]. Epidemiological studies show that these risky compounds could result in health hazards, e.g., gastroenteritis and metabolic dysregulations [11]. Identifying chemical compounds in staple or innovative functional foods and rapid assaying of their bioactivities are significant to address the ever-increasing concerns of worldwide food and nutritional security. Machine learning, as the most rapidly-evolving field of artificial intelligence, has been accelerating the processes of foods research [12–14].
In the last decades, an increasing number of omics databases have been developed for foodomics or foodinformatics technology research [15, 16], most of which focus on some specific characteristics of food or food compositions, e.g., additive chemicals in foods (AdditiveChem) [9], food flavor compounds (FlavorDB) [17], food molecular behaviors [18], risky components in foods (FRCD) [19], food-drug interactions (NutriChem) [20], food adulteration (FADB-China) [21], food bioactive peptides (DFBP) [22], etc. Comparatively, FoodDB (https://foodb.ca/) [23] is a comprehensive database that curates multi-faceted information about health effects, metabolic enzymes and pathways, flavoring, nutrition and chemical structures of food constituents. These data provide avenues for us to investigate the underlying molecular mechanisms how foods maintain normal life metabolism and physiological health from different omics disciplines such as nutrigenomics, pharmacogenomics, metabolomics and proteomics. Artificial intelligence based analytical techniques promise to greatly accelerate food science research via deep mining these foodomics data. The commonly-used techniques in traditional machine learning application domains (e.g., image signal processing and natural language processing) have been successfully adapted to food classification. For instances, Kumar et al [24] train a deep convolutional neural network (CNN) framework on apple images to automatically classify various nutrient-deficient apples. Ingredient statement words in natural language from food labels are exploited via multi-layer perceptron (MLP) to predict food categories and nutrients [25], or learnt via k-nearest neighbors (KNN) to predict added-sugar content of food [26]. Food near-infrared spectra (FT-NIR) data are exploited via support vector machine (SVM) to predict food geographical origins [27] or food categories [28]. These methods in some sense could be deemed as shallow applications of advanced data analytical techniques to automatic food classification, which are actually not involved in foodomics.
Artificial intelligence especially machine learning techniques are supposed to demonstrate full advantages in unravelling quantitative structure-activity relationship (QSAR) of food chemical compounds and gaining molecule-level insights into their health and nutritional effects. At present, the studies of foodomics data mining focus on flavor compounds [29, 30], beneficial or adverse health effects of food ingredients [31–34] and bioactive peptides [35, 36], food-drug interactions [37, 38], food compounds identification [39] and metabolic pathways of food compounds [40]. Bi et al [30] first convert GC-MS spectrum data into fingerprint images and then learn the fingerprint representations of compounds via convolutional neural networks (CNN) for food flavor evaluation. For small chemical molecules, proper encoding of multi-faceted molecular properties into numerical bioactive descriptors is critical to machine learning performance. Berton et al [31] propose a collection of deep neural networks to learn the bioactivity signature representations of small molecules like drugs, natural products and food ingredients, which have demonstrated to yield substantial performance improvement. The current foodmics studies on bioactive effects of food compounds largely focus on some particular health benefits, and these methods directly transfer knowledge of drugs to food ingredients. For instance, Gonzalez et al [32] use graph neural networks to learn the vector representations of drugs and then train a multi-layer perceptron (MLP) to predict anticancer bioactivity of food molecules. The work [33] use drugs’ EP2 fingerprints to train a random forest model for predicting pharmacological effects of food compounds. Beside health benefits, the knowledge about drug side-effects is also exploited to predict health risks such as mutagenicity and carcinogenicity of food chemical contaminants [34]. Like predicting food bioactivities, the knowledge about effects of drug-drug interaction is transferred via a deep learning framework to predict the synergistic or antagonist impacts of food compounds on drug efficacies or bioavailability [37]. Transferring knowledge from well-studied drugs to less-understood food molecules from the point of view of quantitative structure-activity relationship (QSAR) surely helps promote food studies, but the difference of scaffold diversities and other properties between synthetic and natural molecules potentially affects the reliability of predictive models. The transferability of knowledge from drugs to food molecules needs to be further substantiated. Actually, the currently available foodomics data alone provide sufficient supports for us to gain new insights into food compounds without dependency on knowledge transfer from drugs.
In this work, we attempt to exploit currently available bioactive food compounds to mine the inherent bioactive patterns for gaining new insights into known chemical molecules or providing prior knowledge for unknown molecules. For generality, we propose an ensemble learning framework embracing 101 binary models to predict most of the major classes of health effects curated in FoodDB (https://foodb.ca/). For each binary model, the health effect of interest is treated as positive class and the other health effects are treated as negative class, so that the class distribution is highly skewed to yield models biased towards the majority class. In this framework, random undersampling boosting (RUSBoost) [41] is used as the base learner to tackle the skewedness of training data. To mine the inherent bioactive patterns of food compounds, each molecule is numerically vectorized into MACCS fingerprint similarity spectrum, which defines its similarities in terms of MACCS structural scaffolds and functional subgroups with all template molecules. All the binary models in this framework are evaluated using 5-fold cross validation on the training data and independent test on external plant natural products and drug molecules. As an application scenario, the proposed framework is used to predict the potential beneficial or adverse health effects of food flavor molecules. Gene ontology (GO) and metabolic pathways enrichment analyses are further conducted to provide support to the predicted health effects.
Data and materials
To our knowledge, FoodDB (https://foodb.ca/) is currently the largest database of food ingredients with 1083 classes of health-associated bioactivities or health effects. Nevertheless, most classes of health effects have very few food compounds associated hereto, wherein the health effects with only one compound account for 50.78%. From computational point of view, two few examples are not sufficiently representative for the minority class they belong to, and the limited information they contain is easily flooded by the majority classes with ample examples. For the reason, the health effects with no more than 10 compounds are discarded and the other health effects remain for computational modeling. After filtering out the compounds whose SMILES do not pass the parse of RDKit [42], we totally obtain an entire set of 2828 food compounds belonging to 101 highly imbalanced classes of health effects, in which the largest antibacterial class contains 149 compounds and most classes are very small to contain only 11 ~ 30 compounds. These compounds are used as training data to build predictive classifiers.
Independent test on external data is an indispensable step to check how well the proposed framework generalizes to unseen examples. At present we do not find enough food compounds with known health effects from other sources, so we collect natural products from literature [43–48] and drug molecules from ChEMBL [49] as independent test data, which is more challenging than independent test on food molecules due to larger structural gaps. Among the natural products molecules, we have collected 1477 anticancer molecules from NPACT [43], 368 antibacterial molecules from BioPhytMol [44], 188 antiviral molecules from [45, 46] and 436 pesticide molecules from [47, 48]. All the drug molecules are collected from ChEMBL [49] including 76 antiviral molecules, 226 antibacterial molecules, 193 antineoplastic molecules and 76 analgesic molecules.
Feature engineering for food molecules
Chemical fingerprints are often used to numerically represent small molecules. MACCSKeys fingerprint [50] is designed to describe the presence or absence of the predefined functional groups, substructure motifs or fragments that have demonstrated successes in revealing structure-activity relationship (SAR). Nevertheless, the 166-bit vector of MACCSKeys fingerprint is too short to fully describe a large training data and thus easily results in model underfitting. For the reason, we choose the 2828 food compounds in the training set as templates and describe each molecule of interest with 2828-dimensional embedding vector, in which each dimension defines the MACCSKeys fingerprint similarity between the molecule and corresponding template molecule, hereinafter termed as MACCSKeys similarity spectrum. In this study, we adopt Tanimoto coefficient (actually Jaccard index) to measure the rate of common substructures between two food compounds.
$$\text{T}\text{a}\text{n}\text{i}(\text{a},\text{b})=\frac{{ N}_{ab}}{{ N}_{a}+{ N}_{b}{-N}_{ab}}$$
1
where \({ N}_{a},{N}_{b},{N}_{ab}\)denote the number of substructures and the number of common substructures of two compounds, respectively.
Formally, a template compound is denoted as\({ T}_{i}, i=\text{1,2},\dots ,N (N=2828)\). For a compound \({d}_{i}\) concerned, it’s MACCSKeys similarity spectrum is defined as the following hot embedding vector.
$${V}_{{d}_{i}}≔[\text{T}\text{a}\text{n}\text{i}\left({d}_{i},{ T}_{1}\right),\text{T}\text{a}\text{n}\text{i}\left({d}_{i},{ T}_{2}\right),\dots ,\text{T}\text{a}\text{n}\text{i}\left({d}_{i},{ T}_{k}\right),\dots ,\text{T}\text{a}\text{n}\text{i}\left({d}_{i},{ T}_{N}\right)]$$
2
whose kth component defines MACCSKeys fingerprint similarity between compound \({d}_{i}\)and template compound\({ T}_{k}\).
where \({Z}_{t}\)is the normalizer of instance weight distributions and Y denotes the label set. As shown in Eq. (3 ~ 5), the larger errors mislabelled instances yield, the higher weights are assigned to these instances to train the next weak learner, and the lower weight is assigned to the vote of the current learner. Although random undersampling alleviates class imbalance to a certain degree, the minority class is still more easily to be erroneously classified because it is much less diverse or representative than the majority class. Fortunately, AdaBoost.M2 assigns larger weights to mislabelled minority class so that the bias towards majority class promises to be reduced. Interested readers are referred to RUSBoost [41] and AdaBoost.M2 [51] for algorithmic details.
Random undersampling boosting as base binary learner
In many cases, a food compound commonly manifests more than one health effect, which is naturally modelled via multi-label learning in machine learning community. However, each compound is assigned to only one unique health effect in FooDB (https://foodb.ca/). For the reason, two alternative strategies are available for FooDB mining, one is multi-class learning and the other is assemble of binary learning. In this work, we choose binary learning to computationally model each food health effect, in which the compounds of the health effect concerned are treated as the positive class and all the remaining compounds in the entire set are treated as the negative class. We choose this strategy based on the two major concerns. As one concern, the strategy adopted in this study is actually binary relevance (BR) of multi-label learning, so that the proposed framework is easily adapted and further evaluated in terms of multi-label learning metrics with the accumulation of knowledge of food health effects. As the other concern, class imbalance is more easily handled in binary learning via data undersampling or oversampling than in multi-class learning. Mining highly imbalanced 101-class data via multi-class learning would yield a highly biased model, in which many small classes are potentially ignored or vanishing completely.
For all the binary models, the minority class (positive class) is much smaller than the major class (negative class), accounting for 0.39%~5.27% of the entire training data. To tackle class imbalance, we adopt random undersampling boosting (RUSBoost) [41] as the base learner. RUSBoost [41] makes a slight modification on AdaBoost.M2 [51] via randomly resampling the majority class rather than the entire training data. For T iterations, each iteration trains a weak learner (e.g., decision tree) termed hypothesis\({ h}_{t}\) on the training data \({ S}_{t}\), which is produced via randomly removing data from the majority class until some desired ratio of minority to majority is achieved. An instance \({({x}_{i},{y}_{i})\in S}_{t}\) is mislabelled if \({ h}_{t}({x}_{i},{y}_{i})=0 \&\& { h}_{t}{(x}_{i},{y\ne y}_{i})=1\); otherwise is correct if \({ h}_{t}({x}_{i},{y}_{i})=1 \&\& { h}_{t}{(x}_{i},{y\ne y}_{i})=0.\)Given weight distribution \({ D}_{t}\left(i\right)\)of mislabelled instance i, the pseudo-loss \({ϵ}_{t}\)of iteration t is defined as \({ϵ}_{t}=\sum _{\left(i,y\right):{y\ne y}_{i}}{ D}_{t}\left(i,y\right)(1-{ h}_{t}\left({x}_{i},{y}_{i}\right)+{ h}_{t}{(x}_{i},y)),\) based on which to further define weight of the t-th weak learner \({ h}_{t}\) (see Eq. (3)) and update the weights of mislabelled instances for training (t+1)-th weak learner (see Eq. (4)). The final decision or hypothesis is made via weighted votes of the T weak learners (see Eq. (5)).$${\alpha }_{t}={ϵ}_{t}/(1-{ϵ}_{t})$$
3
$${D}_{t+1}\left(i,y\right)=\frac{{ D}_{t}\left(i,y\right)}{{Z}_{t}}{{\alpha }_{t}}^{\frac{1}{2}(1-{ h}_{t}\left({x}_{i},{y}_{i}\right)+{ h}_{t}{(x}_{i},y))}$$
4
$$h\left(x\right)=arg\underset{y\in Y}{\text{max}}\sum _{t=1}^{T}\left(log\frac{1}{{\alpha }_{t}}\right){ h}_{t}(x,y)$$
5
All the binary RUSBoost models could be used as independent classifiers to predict some specific health effects of interest, e.g., anticancer and antibacterial bioactivities. These models could also be used as an assemble to predict multiple health effects of one food compound, which is essentially a multi-label learning scenario.
Model evaluation
The proposed framework is evaluated via 5-fold cross validation and independent test. As mentioned before, predicting food health effects is eventually a multi-label learning scenario, but multi-labels are not available in FooDB. For the reason, all the binary models in this framework are evaluated via the classical performance metrics for binary classification like Precision, Recall, F1 score and Matthews correlation coefficient (MCC) as defined in Eq. (6)
where TP, TN, FP and FN denote true positive, true negative, false positive and false negative, respectively. For intuitively perceiving predictive bias between imbalance classes, we also provide precision and recall metrics on the negative class to, which could be easily calculated via transposing positive class and negative class. In addition, area under the ROC curve (AUC) is also used to evaluate these binary models.
Independent test on external data collected from natural products and drugs is conducted to estimate how well the proposed framework generalizes to unseen small molecules solely based on molecular structures, and meanwhile to validate knowledge transferability from foods to natural products and drugs. The performance of independent test is evaluated via recall on molecules of natural products and drugs whose health effects are known.
Visualizing MACCSKeys similarity spectrum of chemical molecules
The 2828 food compounds in the training set are used as templates to generate 2828-dimensional vectors termed as MACCSKeys similarity spectrums, depicting MACCSKeys fingerprint similarities of the molecules of interest with all the templates. The MACCSKeys similarity spectrums of three chemical molecules are illustrated in Fig. 1A ~ C. The first molecule FDB001388 (25-Methyl-24-methylenecholesterol), curated in FoodDB (https://foodb.ca/), is a food constituent of seeds of Indian mustard (Brassica juncea), sunflower (Helianthus annuus) and French beans (Phaseolus vulgaris). The second molecule Tithofolinolide, with identifier NPACT00978 in NPACT [43], is a plant natural product that belongs to terpenoids class with the therapeutic effects of treating colon, haematopoietic and lymphoid tissue cancers. The third molecule Vidarabine, with identifier CHEMBL1090 in ChEMBL [49], is a virus inhibitor that targets human herpesvirus 1 DNA polymerase to cure eye infections. As shown in Fig. 1, the number and amplitudes of spectra peaks decrease in the order of food compounds, natural products and drugs. Natural product Tithofolinolide shows many medium spectra peaks ranging from 0.4 to 0.8 (see Fig. 1B) and drug Vidarabine also shows several high peaks above 0.8 (see Fig. 1C). These spectrums to some degree indicate knowledge transferability between food compounds, natural products and drug molecules.
Performance of 5-fold cross validation
The results of stratified 5-fold cross validation are illustrated in Fig. 2A ~ G. For all the binary models, the positive classes account for very low percentages of the entire training data, ranging from the highest 5.27% to the lowest 0.39% (see Fig. 2A). The numberings of health effects are provided in Supplementary file S1. Nevertheless, most of the RUSBoost learners achieve encouraging performance in terms of accuracy (see Fig. 2B) and AUC scores (see Fig. 2C). The accuracies range from 68.33–94.87% with 60% learners achieving more than 80% accuracy. The ROC AUC scores range from 0.5991 to 0.9905 with 57% learners achieving more than 0.8 AUC score. In Fig. 3A, the ROC curves for 14 health effects (e.g., antibacterial, anti-inflammatory, anticancer) are illustrated as examples. In the scenario of extreme class imbalance, these two metrics still report optimistic performance when the models deterioratively or corruptively behave, for instance, all the minority classes are misclassified to the majority classes. For the reason, the recall rates on both classes are more important to know how much biased the models are. It is shown in Fig. 2D that the RUSBoost learners achieve satisfactory recall rates on the minority class (positive class) with 76.24% models achieving recall above 0.6 and 31.68% models achieving recall above 0.8. On the majority class (negative class), all the models unsurprisingly achieve very high recall rates with 100% models achieving recall above 0.6 and 58.42% models achieving recall above 0.8. Recall metrics show that the RUSBoost learners still show a certain degree of bias, but the bias is much alleviated.
In the scenario of imbalanced-class learning, the misclassification of majority class heavily impacts the precision of the minority class, but slightly vice versa. As shown in Fig. 2E, the precisions on the minority class are very low with 22.77% models achieving precision above 0.05. The results are not surprising, because only a small fraction of misclassified majority class (i.e., false positive) will greatly recue the ratio of true positive to the whole predicted positives. Comparatively, the precisions on the majority classes are very high with 100% models achieving precision above 0.9 (see Fig. 2F). These precisions are similarly distributed as the work [52], which uses deep learning to predict drugs targeting specific enzymes or genes with precisions all lower than 0.2 and many precisions ranging from 0.01 to 0.06. In this framework, 77.23% models achieve precision above 0.015. Similarly, deep learning for imbalanced drug discovery [52] mostly achieves MCC scores ranging from 0.05 to 0.3. In this framework, 87.13% models achieve MCC scores above 0.05 (see Fig. 2G), though the training data is much smaller as compared to 246,591 ~ 304,987 training examples fed to deep learning framework [52].
Performance of independent test
We cannot find other sources of food compounds whose health effects are known, so natural products and drug molecules are used instead as independent test data. The bioactivities of these data are usually not consistent with the health effects adopted in this framework, for instance, anticancer in NPACT [43] is categorized according to tissues, potentially corresponding to 8 health effects in FooDB, e.g., cancer-preventive, antileukemic, anticancer, antitumor, antineoplastic, antiproliferant, antimetastatic and anticarcinogenic. For the sake of semantic difference, any food compound from NPACT [43] is deemed to be correctly recognized if predicted to be positive by any one model of these 8 health effects. For other independent test data, we adopt similar strategy. The recalls on plant natural products from NPACT [43] range from 0.8406 to 0.9040 (see Fig. 3B) and the recalls on drug molecules from ChEMBL [49] range from 0.7895 to 0.9690. These results show that the proposed framework generalizes well to unseen chemical molecules and meanwhile the knowledge is transferable between food compounds, natural products and drug molecules.
Repurposing natural products and drugs with supporting evidences from recent literature
As demonstrated in independent test, the models trained on food compounds are useful for predicting novel activities of natural products and drugs. Based on this point, we use the proposed framework to find new purposes of the natural products in NPACT [43] and drugs in ChEMBL [49]. As shown in Fig. 4, most natural product and drug molecules exhibit multiple bioactivities. For clarity, only the health effects with more than 30% molecules are illustrated. The entire predictions are provided in Supplementary file S2 (NPACT natural products) and Supplementary file S3 (ChEMBL drugs). The anticancer natural products exhibit antiedemic (68.61%), anti-inflammatory (57.72%) and antiviral (46.53% anti-HIV) health effects (see Fig. 4A). Meanwhile, 32.87% molecules also exhibit antibacterial effects (see Fig. 4B). Among the antiviral drug molecules, most molecules exhibit anticancer effects, e.g., anti-carcinomic (94.74%) and anticancer (68.42%) (see Fig. 4C). In addition, 42.11% drugs also exhibit cytotoxic adverse effects (see Fig. 4D). These results to some degree reveal the associations between carcinogenesis, viruses and inflammation as well as drug repurposing. The predicted effects for us to turn these molecules for other therapeutic purposes. Independent test demonstrates to what percentage of labels are correctly recognized. As for novel health effects, we resort to recent literature for supporting evidences. In this section, we only conduct dozens of case studies via manual literature validation. Automatic literature search via Recurrent Neural Network (RNN) and natural language processing (NLP) promise to enlarge the coverage of validation, which could be treated as an independent research topic.
Case studies on anticancer natural product Nobiletin
Nobiletin, a plant metabolite or natural product found in Citrus tankan, is a methoxyflavone used as an antineoplastic agent. When applying Nobiletin to the 101 binary models, we obtain other 22 health effects besides anticariogenic (see Supplementary file S2 for details), e.g., anti-inflammatory, antioxidant, anti-alzheimeran, anti-hepatotoxic, anti-HIV, anti-histaminic, choleretic, etc. To validate these novel effects, we manually retrieve publications via the keywords Nobiletin and specific effects. The retrieved evidences and corresponding article PubMed identifiers are provided in Table 1. As shown in Table 1, two articles (PMID: 25191498, PMID: 31295812) support that nobiletin takes anti-dementia activities to improve the symptoms of neurodegenerative diseases such as Alzheimer’s disease and Parkinson’s disease. Furthermore, the article (PMID: 26874072) demonstrates that Nobiletin takes antioxidant, anti-inflammatory and anti-alzheimeran health effects, which are all predicted by the proposed framework. We also find supporting evidences to show that nobiletin takes many other effects such as antiulcer, anticataract, hepatoprotective, vasodilator, VEGF-Inhibitor and antihypertensive. These evidences convincingly demonstrates that the proposed framework trained on food compounds generalizes well to natural products.
| Predicted health effects | Supporting evidences | References |
|---|---|---|
| Antialzheimeran | Nobiletin is shown to improve cognitive deficits and pathological features of neurodegenerative diseases such as Alzheimer’s disease and Parkinson’s disease. | 1. Anti-dementia Activity of Nobiletin, a Citrus Flavonoid: A Review of Animal Studies (2014) [PMID: 25191498] 2. Potential Benefits of Nobiletin, A Citrus Flavonoid, against Alzheimer's Disease and Parkinson's Disease (2022) [PMID: 31295812] |
| Antioxidant, Antiinflammatory | Nobiletin is reported to have anti-oxidant and anti-inflammatory effects in treating asthma, colitis and Alzheimer's disease. | 1. Nobiletin promotes antioxidant and anti-inflammatory responses and elicits protection against ischemic stroke in vivo (2016) [PMID: 26874072] |
| Vasodilator | Nobiletin is shown to have cardiovascular protection effects (e.g., prevention of cardiac hypertrophy and platelet aggregation) and relaxation of endothelium-denuded rat aortic smooth muscles. | 1. Nobiletin, a citrus flavonoid, activates Vasodilator-stimulated phosphoprotein in human platelets through non-cyclic nucleotide-related mechanisms (2017) [PMID: 27959381] 2. Endothelium-independent vasodilator effects of nobiletin in rat aorta (2019) [PMID: 31088764] |
| Antiulcer | Nobiletin is shown to have anti-ulcerogenic activity that significantly attenuates the ethanol-induced gastric ulcer in mice. | 1. The gastroprotective effect of nobiletin against ethanol-induced acute gastric lesions in mice: impact on oxidative stress and inflammation (2017) [PMID: 28948855] |
| Anticataract | Nobiletin is verified to prevent formation of cataract via inhibiting the proliferation of human lens epithelial cells. | 1. Polymethoxyflavones as agents that prevent formation of cataract: nobiletin congeners show potent growth inhibitory effects in human lens epithelial cells (2013) [PMID: 23199882] |
| Hepatoprotective | It is shown that hepatic proinflammatory TNF-α mRNA expression and liver damage indicators are significantly lower in Nobiletin-supplemented mice; and amorphous solid dispersion of Nobiletin improves Nobiletin’s hepatoprotective properties. | 1. Long-term dietary supplementation with low-dose nobiletin ameliorates hepatic steatosis, insulin resistance, and inflammation without altering fat mass in diet-induced obesity (2017) [PMID: 28116779] 2. Physicochemical and biopharmaceutical characterization of amorphous solid dispersion of nobiletin, a citrus polymethoxylated flavone, with improved hepatoprotective effects (2013) [PMID: 23707470] |
| VEGF-Inhibitor | Nobiletin inhibits the secretion of key angiogenesis mediators and vascular epithelial growth factor (VEGF). | 1. The flavonoid nobiletin inhibits tumor growth and angiogenesis of ovarian cancers via the Akt pathway (2015) [PMID: 25845666] |
| Antihypertensive | Nobiletin reduces high levels of blood pressure, circulating angiotensin II and angiotensin‑converting enzyme activity. | 1. Nobiletin resolves left ventricular and renal changes in 2K-1C hypertensive rats (2022) [PMID: 35662276] 2. Nobiletin alleviates vascular alterations through modulation of Nrf-2/HO-1 and MMP pathways in l-NAME induced hypertensive rats (2019) [PMID: 30864566] |
Case studies on antibacterial implications of anticancer natural products
We further conduct case studies to show that the anticancer natural products could be properly repurposed for antibacterial therapies, which is an alternative solution combat bacterial drug resistance. In Table 2, supporting evidences with PubMed identifiers are provided for 8 natural product compounds. As shown in Table 2, sapintoxins A takes strong effect against Mycobacterium tuberculosis H37Ra (PMID: 24877849). Xanthochymol and Stigmasterol exhibit both antibacterial and antifungual effects. Rapanone is effective against Staphylococcus aureus strains of bacteria that has developed methicillin resistance. These results show that natural chemicals from plants and microbes are natural reservoirs for antimicrobial drugs development and the proposed framework could be used as a tool for discovery of antibacterial lead compounds.
| Natural products | Supporting evidences | References |
|---|---|---|
| Menthol | Menthol oils strongly inhibit plant pathogenic microorganisms, whereas human pathogens are only moderately inhibited. | 1. Antimicrobial screening of Mentha piperita essential oils (2003) [PMID: 12083863] |
| Glycyrrhizic acid ammonium salt | Studies demonstrate that glycyrrhizic acid ammonium salt via nanoparticles suppress phytoplasma infection. | 1. Nanoinhibitory Impacts of Salicylic Acid, Glycyrrhizic Acid Ammonium Salt, and Boric Acid Nanoparticles against Phytoplasma Associated with Faba Bean (2022) [PMID: 35268567] |
| Sapintoxin A | Sapintoxins A and C exhibit strong activities against Mycobacterium tuberculosis H37Ra. | 1. A review of the medicinal uses, phytochemistry and pharmacology of the genus Sapium (2014) [PMID: 24877849] |
| β-Sitosterol | β-sitosterol identified in Parthenium hysterophorus is verified via in-vitro bioassay and analytical chemistry method to have antibacterial activities against aquatic bacterial pathogens. | 1. β-Sitosterol: An Antibacterial Agent in Aquaculture Management of Vibrio Infections (2020) [https://doi.org/10.22207/JPAM.14.4.48] |
| Xanthochymol | Xanthochymol isoforms isolated from the genus Garcinia have antibacterial and anti-fungal activities. | 1. Garcinia L.: a gold mine of future therapeutics (2021) [https://link.springer.com/article/10.1007/s10722-020-01057-5]. |
| Stigmasterol | Stigmasterol is a potent and broad-spectrum antibacterial and antifungal agent as a lead compound of antimicrobial drugs. | 1. Antimicrobial activity of stigmasterol from the stem bark of Neocarya macrophylla (2018) [https://doi.org/10.4102/jomped.v2i1.38] |
| Rapanone | Rapanone is shown to be active against methicillin resistant Staphylococcus aureus (MRSA) strains of bacteria. | 1. Antibacterial activities and structure-activity relationships of a panel of 48 compounds from Kenyan plants against multidrug resistant phenotypes (2016). [PMID: 27386347] |
| Geniposide | Geniposide exhibits antibacterial, anti-inflammatory therapeutic efficacies. | 1. Geniposide reduces Staphylococcus aureus internalization into bovine mammary epithelial cells by inhibiting NF-κB activation (2018) [PMID: 30321590] |
Case studies on anticancer and antibacterial repurposing implications of antiviral drugs
Comparatively, drugs less resemble food compounds than natural products in terms of origin and scaffold diversity, but independent test shows that the proposed framework trained on food compounds also generalizes well to drugs. Actually, many synthetic drug molecules are modified from natural products, thus reducing the structural gaps between drugs and food compounds. The antiviral drugs from ChEMBL [49] are chosen as case studies to validate the reliability of novel predictions by the proposed framework. The entire predictions are provided in Supplementary file S3. Twelve antiviral drugs are found to have anticancer and/or antibacterial activities via manual literature retrieval (see Table 3). As shown in Table 3, most antiviral drugs exhibit anticancer effect. For instance, oseltamivir demonstrates potential treatment of liver cancer (PMID: 34859259), adefovir dipivoxil inhibits tumorigenesis of colon cancer cells (PMID: 30872078) and Maraviroc demonstrates therapeutic potential for colorectal cancer liver metastasis (PMID: 32902795). Didanosine phosphoramidates exhibit both anticancer and antibacterial effects. The antiviral drug demonstrates potential effects against bacteria S. aureus, which is sensitive to and resistant against methicillin (PMID: 34668851). These evidences suggest that the proposed framework cold help us find the antiviral drugs that are potentially repurposed for anticancer therapies.
| Predicted effects of drugs | Supporting evidences | References |
|---|---|---|
| Efavirenz (antibacterial) | Efavirenz significantly inhibits biofilm formation of both methicillin-sensitive S. aureus and methicillin-resistant S. aureus; Efavirenz inhibits the growth of B. subtilis. | 1. The antiviral drug efavirenz reduces biofilm formation and hemolysis by Staphylococcus aureus (2021) [PMID: 34668851] 2. Antibacterial effects of antiretrovirals, potential implications for microbiome studies in HIV (2018) [PMID: 28497768] |
| Oseltamivir (anticancer) | It has been reported that Oseltamivir induces death of liver cancer cells. | 1. Potential of antiviral drug oseltamivir for the treatment of liver cancer (2021) [PMID: 34859259] |
| Didanosine (antibacterial, anticancer) | Didanosine phosphoramidates exhibit marked activities against Gram negative and positive bacteria; Didanosine represents a widely prescribed class of antiviral and anticancer drugs. | 1. Didanosine phosphoramidates: synthesis, docking to viral NA, antibacterial and antiviral activity (2014) [10.1007/s00044-014-1073-2] 2. Bioretrosynthetic construction of a didanosine biosynthetic pathway (2014) [PMID: 24657930] |
| Adefovir dipivoxil (anticancer) | Adefovir dipivoxil inhibits tumorigenesis of colon cancer cells. | 1. Adefovir dipivoxil sensitizes colon cancer cells to vemurafenib by disrupting the KCTD12-CDK1 interaction (2019) [PMID: 30872078] |
| Ritonavir (anticancer) | Ritonavir been shown to have anticancer or synergistic anticancer activities in several cancers. | 1. Overcoming cancer therapeutic bottleneck by drug repurposing (2020) [PMID: 32616710] 2. HIV-1 protease inhibitor, ritonavir: a potent inhibitor of CYP3A4, enhanced the anticancer effects of docetaxel in androgen-independent prostate cancer cells in vitro and in vivo (2004) [PMID: 15492266] |
| Darunavir (anticancer) | Darunavir, as a synthetic nonpeptidic protease inhibitor, has been verified to have anticancer activities. | 1. Molecular designing and in silico evaluation of darunavir derivatives as anticancer agents (2014) [PMID: 24966524] |
| Ribavirin (anticancer) | Ribavirin is an effective anticancer agent, but its efficacy seems to vary with cancers or tissues. | 1. Ribavirin as an anti-cancer therapy: acute myeloid leukemia and beyond? (2010) [PMID: 20629523] |
| Saquinavir (anticancer) | Saquinavir has exhibited effects in several types of cancer, e.g., Kaposi carcinoma, neuroblastoma. | 1. Saquinavir: From HIV to COVID-19 and Cancer Treatment (2022) [PMID: 35883499] |
| Nelfinavir (anticancer) | Nelfinavir takes anti-cancer effects via modulating different cellular conditions, e.g., cell cycle, tumor microenvironment. | 1. The Anti-Cancer Properties of the HIV Protease Inhibitor Nelfinavir (2020) [PMID: 33228205] |
| Maraviroc (anticancer) | Maraviroc significantly inhibits CRC liver metastasis in animal model. | 1. Antineoplastic effects of targeting CCR5 and its therapeutic potential for colorectal cancer liver metastasis (2021) [PMID: 32902795] |
| Abacavir (anticancer) | Abacavir induces senescence and cell death in prostate cancer cells. | 1. The reverse transcription inhibitor abacavir shows anticancer activity in prostate cancer cell lines (2010) [PMID: 21151977] |
| Cidofovir (anticancer) | Cidofovir takes antineoplastic activities against HCMV-infected glioblastoma cells, or even induces death of glioblastoma cells free of HCMV infection. | 1. Cidofovir: a novel antitumor agent for glioblastoma (2013) [PMID: 24170543] |
Unravelling beneficial and risky health effects of food flavors
Food flavors are widely used to enhance food acceptance and popularity, and thus health safety of flavors becomes the first concern to be addressed. From FooDB, we obtain 1476 flavor compounds. We use the proposed framework to unravel the beneficial and risky effects of these compounds. The entire predictions are provided in Supplementary file S4. From the predicted effects, we choose antibacterial and antiviral effects as beneficial and choose cytotoxic and mutagenic effects as risky. It is noted that the proposed framework may classify a compound as mutagenic and antimutagenic contradictorily, e.g., phenethyl phenylacetate (FDB013613). In this case, the compound is deemed to be mutagenesis associated without determining its beneficial or adverse property. The contradictory predictions result from the independency of binary models, none of which is deliberately designed to discriminate mutagenic effect from antimutagenic effect. The interpretation is applicable to other pairs of opposite effects.
Case study on antibacterial and antiviral curcumenol
Computational results show that many food flavor molecules exhibit both antibacterial and antiviral activities. Ten compounds are illustrated as examples in Fig. 5A, out of which curcumenol is chosen for case study. To estimate the confidence level of the two predicted health effects, we calculate MACCSKeys fingerprint similarities between curcumenol and the compounds in FooDB, which have been annotated with antibacterial and antiviral health effects. Eight compounds with top high similarities (> 0.7) are illustrated in Fig. 5B, in which (2) ~ (5) are antiviral compounds and (6) ~ (9) are antibacterial compounds. Common functional subgroups to a certain degree could interpret that these molecules together with curcumenol exhibit antibacterial and antiviral activities. More convincing evidences still need resort to lab experiments. Assays via serial dilution method demonstrate that curcumenol exhibits significant antibacterial activities against Pseudomonas aeruginosa and Streptococcus pyogenes [53]. Meanwhile, curcumenol together with the other ingredients of C. kwangsiensis like curcumin, curcumol and curdione exhibit antiviral activities [54].
Case study on cytotoxic and mutagenic 3-Methyl-1-phenyl-3-pentanol
Food flavorings could beneficially promote health but also potentially do damages to human and animals. In Fig. 6A, ten compounds potentially inducing cytotoxicity and genotoxicity are illustrated. Take 3-Methyl-1-phenyl-3-pentanol for example, six structurally similar compounds with (2) ~ (4) annotated with cytotoxic effects and (5) ~ (7) annotated with mutagenic effect are illustrated in Fig. 6B. The MACCSKeys fingerprint similarities are all above 0.5, and these molecules resemble in terms of scaffolds and functional subgroups. As surveyed in [55], acute toxicities of 3-Methyl-1-phenyl-3-pentanol are observed in animals but human subjects show no irritation reaction. As far, no publications report 3-Methyl-1-phenyl-3-pentanol induced human cell mutagenicity.
QSAR (quantitative structure-activity relationship) model reveals that structurally similar molecules often exhibit similar activities. This model provides an avenue for us to uniformly study three classes of chemical molecules, i.e., food compounds, natural products and drugs. Artificial intelligence-based techniques, especially classic machine learning approaches and emerging deep learning framework, have demonstrated encouraging successes in food research and drug discovery. Classic machine learning methods resort to expert knowledge for molecule representation and thus easily achieve interpretability, whereas deep learning directly learns molecule representation from molecule graph without expert intervention, which to choose depends on application scenarios and data abundance.
In this work, we attempt to classify health effects of food compounds at a large scale, with health effects amounting to 101 classes. We choose random undersampling boosting (RUSBoost) as the base learner because the training data are small and class distribution is highly skewed. Each chemical molecule is vectorized as MACCSKeys similarity spectrum to define the fingerprint similarities of functional subgroups between molecules of interest with predefined template molecules. Each model takes charge of one specific health effect and thus the ensemble of these models could yield multiple predicted effects. Existing methods mostly work for one specific health effect in a binary learning scenario, while this framework works in a multi-label learning scenario. Besides, existing methods train model on drugs to classify food molecules. This framework is directly trained on food compounds, which can be used to classify food, natural product and drug molecules. Knowledge transferability between these three classes of molecules has been validated by independent test.
Independent test demonstrates how well the known health effects are correctly recognized. The novel effects need resort to recent literature for supporting evidences. In this work, dozens of novel effects have been validated with convincing evidences. These evidences show that the proposed framework can be used to repurpose drugs or find lead compounds for new therapeutic purposes. Especially, there exist associations and drug repurposing between carcinogenesis, inflammation and viral infection. Lastly, we use the proposed framework to unravel beneficial and risky health effects of food flavor compounds, some of which are also validated against recent literature. These results are potentially useful for recipe composing.
Competing interests
The author declares that there are no competing interests.
Authors’ contributions
MS conducted the study and wrote the paper.
Acknowledgements
Compliance with ethical standards
Funding
No funding was received for this study.
- Brewer MS (2011) Natural Antioxidants: Sources, Compounds, Mechanisms of Action, and Potential Applications. Compr Rev Food Sci F 10:221–247
- Zhu F, Du B, Xu B (2018) Anti-inflammatory effects of phytochemicals from fruits, vegetables, and food legumes: A review. Crit Rev Food Sci Nutr 58:1260–1270
- Liu Y, Kakanic R, Nair MG (2012) Compounds in functional food fenugreek spice exhibit anti-inflammatory and antioxidant activities. Food Chem 131:1187–1192
- Correia-da-Silva M, Sousa E, Pinto MMM, Kijjoa A (2017) Anticancer and cancer preventive compounds from edible marine organisms. Semin Cancer Biol 46:55–64
- Tasdemir SS, TSanlier N (2020) An insight into the anticancer effects of fermented foods: A review. J Funct Foods 75:104281
- Oliveira G, Volino-Souza M, Conte-Júnior CA, Alvares TS (2021) Food-derived polyphenol compounds and cardiovascular health: A nano-technological perspective. Food Biosci 41:101033
- Liu N, Li X, Zhao P, Zhang X, Qiao Q et al (2021) A review of chemical constituents and health-promoting effects of citrus peels. Food Chem 365:130585
- Akpambang VOE, Purcaro G, Lajide L, Amoo IA, Conte LS, Moret S (2009) Determination of polycyclic aromatic hydrocarbons (PAHs) in commonly consumed Nigerian smoked/grilled fish and meat. Food Additives and Contaminants Part a-Chemistry Analysis Control Exposure & Risk Assessment 26:1096–1103
- Zhang D, Cheng X, Sun D, Ding S, Cai P et al (2020) AdditiveChem: A comprehensive bioinformatics knowledge-base for food additive chemicals. Food Chem 308:125519
- Duarte SC, Pena A, Lino CM (2010) A review on ochratoxin A occurrence and effects of processing of cereal and cereal derived food products. Food Microbiol 27:187–198
- Chassaing B, Koren O, Goodrich JK, Poole AC, Srinivasan S, Ley RE, Gewirtz AT (2015) Dietary emulsifiers impact the mouse gut microbiota promoting colitis and metabolic syndrome. Nature 519:92–96
- Wang X, Bouzembrak Y, Lansink AO, van der Fels-Klerx HJ (2022) Application of machine learning to the monitoring and prediction of food safety: A review. Compr Rev Food Sci Food Saf 21:416–434
- Kudashkina K, Corradini MG, Praveena Thirunathan P, Yada RY, Fraser EDG (2022) Artificial Intelligence technology in food safety: A behavioral approach. Trends Food Sci Tech 123:376–381
- Miyazawa T, Hiratsuka Y, Toda M, Hatakeyama N, Ozawa H et al (2022) Artificial intelligence in food science and nutrition: a narrative review. Nutr Rev 80:2288–2300
- Nayak SN, Aravind B, Malavalli SS, Sukanth BS, Poornima R et al (2021) Omics Technologies to Enhance Plant Based Functional Foods: An Overview. Front Genet 12:742095
- Balkir P, Kemahlioglu K, Yucel U (2021) Foodomics: A new approach in food quality and safety. Trends Food Sci Tech 108:49–57
- Garg N, Sethupathy A, Tuwani R, Rakhi Nk, Dokania S et al (2018) FlavorDB: a database of flavor molecules. Nucleic Acids Res 46(D1):D1210–D1216
- Yang ZF, Xiao R, Luo FJ, Lin QL, Ouyang D et al (2020) Food bioactive small molecule databases: Deep boosting for the study of food molecular behaviors. Innov Food Sci Emerg 66:102499
- Zhang D, Gong L, Ding S, Tian Y, Jia C et al (2020) FRCD: A comprehensive food risk component database with molecular scaffold, chemical diversity, toxicity, and biodegradability analysis. Food Chem 318:126470
- Ni Y, Jensen K, Kouskoumvekaki I, Panagiotou G (2017) NutriChem 2.0: exploring the effect of plant-based foods on human health and drug efficacy. Database (Oxford) 2017:bax044
- Zhang D et al (2020) FADB-China: A molecular-level food adulteration database in China based on molecular fingerprints and similarity algorithms prediction expansion. Food Chem 327:127010
- Qin D et al (2022) DFBP: a comprehensive database of food-derived bioactive peptides for peptidomics research. Bioinformatics 38(12):3275–3280
- Harrington RA, Adhikari V, Rayner M, Scarborough P (2019) Nutrient composition databases in the age of big data: foodDB, a comprehensive, real-time database infrastructure. BMJ Open 9:e026652
- Kumar Y, Dubey AK, Arora RR, Rocha A (2022) Multiclass classification of nutrients deficiency of apple using deep neural network. Neural Comput Appl 34:8411–8422
- Ma P et al (2022) Deep learning accurately predicts food categories and nutrients based on ingredient statements. Food Chem 391:133243
- Davies T, 1 Louie JCY, Ndanuko R, Barbieri S, Perez-Concha O, Wu JH (2022) A Machine Learning Approach to Predict the Added-Sugar Content of Packaged Foods. J NUTR 152:343–349
- Richter B, Rurik M, Gurk S, Kohlbacher O, Fischer M (2019) Food monitoring: Screening of the geographical origin of white asparagus using FT-NIR and machine learning. Food Control 104:318–325
- Tsakanikas P, Karnavas A, Panagou EZ, Nychas GJ (2020) A machine learning workflow for raw food spectroscopic classification in a future industry. Sci Rep 10:11212
- Park D, Kim K, Kim S, Spranger M, Kang J (2021) FlavorGraph: a large-scale food-chemical graph for generating food representations and recommending food pairings. Sci Rep 11:931
- Bi K, Zhang D, Qiu T, Huang Y (2020) GC-MS Fingerprints Profiling Using Machine Learning Models for Food Flavor Prediction. Processes 8:23
- Bertoni M, Duran-Frigola M, Badia-I-Mompel P, Pauls E, Orozco-Ruiz M et al (2021) Bioactivity descriptors for uncharacterized chemical compounds. Nat Commun 12:3932
- Gonzalez G, Gong S, Laponogov I, Bronstein M, Veselkov K (2021) Predicting anticancer hyperfoods with graph convolutional networks. Hum Genomics 15:33
- Westerman KE, Harrington S, Ordovas JM, Parnell LD (2020) PhyteByte: identification of foods containing compounds with specific pharmacological properties. BMC Bioinformatics 21:238
- Frenzel F, Buhrke T, Wenzel I, Andrack J, Hielscher J, Lampen A (2017) Use of in silico models for prioritization of heat-induced food contaminants in mutagenicity and carcinogenicity testing. Arch Toxicol 91:3157–3174
- Zaky AA, Simal-Gandara J, Eun JB, Shim JH, El-Aty AMA (2022) Bioactivities, Applications, Safety, and Health Benefits of Bioactive Peptides From Food and By-Products: A Review. Front Nutr 8:815640
- Imai K, Shimizu K, Honda H (2021) Machine learning screening of bile acid-binding peptides in a peptide database derived from food proteins. Sci Rep 11:16123
- Ryu JY, Kim HU, Lee SY (2018) Deep learning improves prediction of drug-drug and drug-food interactions. Proc Natl Acad Sci U S A 115(18):E4304–E4311
- Bennett-Lenane H, Griffin BT, 'Shea O JP (2022) Machine learning methods for prediction of food effects on bioavailability: A comparison of support vector machines and artificial neural networks. EUR J PHARM SCI 168:106018
- Low DY et al (2021) Data sharing in PredRet for accurate prediction of retention time: Application to plant food bioactive compounds. Food Chem 357:129757
- Jia Y, Zhao R, Chen L (2020) Similarity-Based Machine Learning Model for Predicting the Metabolic Pathways of Compounds. IEEE Access 8:130687–130696
- Seiffert C, Khoshgoftaar TM, Hulse JV, Napolitano A (2010) RUSBoost: A Hybrid Approach to Alleviating Class Imbalance. IEEE Trans Syst Man Cybernetics - Part A: Syst Hum 40:185–197
- Landrum G (2013) RDKit: a software suite for cheminformatics, computational chemistry, and predictive modeling. Academic Press, London
- Mangal M, Sagar P, Singh H, Raghava GPS, Agarwal SM (2013) NPACT: Naturally Occurring Plant-based Anti-cancer Compound-Activity-Target database. Nucleic Acids Res 41(Database issue):D1124–D1129
- Sharma A, Dutta P, Sharma M, Rajput NK, Dodiya B, Georrge JJ, Kholia T, Consortium O, Bhardwaj A (2014) BioPhytMol: a drug discovery community resource on anti-mycobacterial phytomolecules and plant extracts. J Cheminform 6:46
- Owen L, Laird K, Shivkumar M (2022) Antiviral plant-derived natural products to combat RNA viruses: Targets throughout the viral life cycle. Lett Appl Microbiol 75:476–499
- Li Goh VSL, Mok CK, Chu JJH (2020) Antiviral Natural Products for Arbovirus Infections. Molecules 25:2796
- Gerwick BC, Sparks TC (2014) Natural products for pest control: an analysis of their role, value and future. Pest Manag Sci 70:1169–1185
- Cantrell CL, Dayan FE, Duke SO (2012) Natural products as sources for new pesticides. J Nat Prod 75:1231–1242
- Gaulton A, Hersey A, Nowotka M, Bento AP, Chambers J et al (2017) The ChEMBL database in 2017. Nucleic Acids Res 45(D1):D945–D954
- Durant JL, Leland BA, Henry DR, Nourse JG (2002) Reoptimization of MDL Keys for Use in Drug Discovery. J Chem Inf Comput Sci 42:1273–1280
- Freund Y, Schapire R (1996) Experiments with a new boosting algorithm. In Proceedings 13th Int Conf Mach Learn 148–156
- Korkmaz S (2020) Deep Learning-Based Imbalanced Data Classification for Drug Discovery. J Chem Inf Model 60:4180–4190
- Akbar N, Siddiqui R, Iqbal M, Khan NA (2020) Antibacterial Activities of Selected Pure Compounds Isolated from Gut Bacteria of Animals Living in Polluted Environments. Antibiotics 9:190
- Liu M, Xiao H, Zhang Q et al (2019) Antiviral abilities of Curcuma kwangsiensis ingredients against grouper iridoviral infection in vitro and in vivo. Aquac Res 00:1–11
- Scognamiglio J, Jones L, Letizia CS, Api AM (2012) Fragrance material review on 1-phenyl-3-methyl-3-pentanol. Food Chem Toxicol 50(Suppl 2):S168–S172
No competing interests reported.