Journal of Functional Foods: Roberto Martín-Hernández, Guillermo Reglero, Alberto Dávalos
Journal of Functional Foods: Roberto Martín-Hernández, Guillermo Reglero, Alberto Dávalos
Journal of Functional Foods: Roberto Martín-Hernández, Guillermo Reglero, Alberto Dávalos
A R T I C L E I N F O A B S T R A C T
Keywords: Regular consumption of certain foods has shown beneficial effects on cardiometabolic health. However, it is not
Nutrigenomics clear by which molecular mechanisms they may exert their beneficial effects. Many genomic experiments
Microarrays available in public databases have generated gene expression data following the treatment of human cells with
Clustering different food nutrients. Exploration of such data offers great possibilities for gaining insights into the molecular
Gene signature
effects of nutrients at cellular level. In this work, we explored the genomic responses triggered by food bioactive
Anticancer
compounds with well-known healthy properties. We show that human cell lines treated with different food
compounds tend to cluster in a cell type dependent manner based on gene expression, with an influence of the
physiological attributes of cells. Finally, we identify a genomic signature of 18 genes implicated in cell cycle,
which may characterize a protective effect of certain food compounds against cancer. Our data provides evi-
dence that nutrigenomic studies found in public databases can be used to discover novel signatures of gene
expression and identify common mechanism of actions of food bioactive compounds.
1. Introduction extracted from foods used in human diet, have shown great potential as
anti-proliferative agents on cultured cancer human cells (Gonzalez-
Nutritional genomics, also known as nutrigenomics, is a relatively Vallinas et al., 2013; Ramirez de Molina et al., 2015). Indeed, a wide
new science which explores the effects of nutrients on the genome, range of drugs for treating diseases such as diabetes and cancer are
proteome and metabolome. Whereas the idea of modulating human derived from natural products. Interestingly, some food compounds
health by food intake is a millennial concept, there are great expecta- have proved their ability to interact with the epigenome, thus mod-
tions on the tremendous potential this science may have to change the ifying microRNA expression (Gil-Zamorano et al., 2014). However, the
future of dietary guidelines in order to improve health and hence to molecular mechanisms by which food bioactive compounds exert their
build up a precision nutrition era (DeBusk, Fogarty, Ordovas, & beneficial effects are still not well understood.
Kornman, 2005). Omics technologies are widely adopted to study the expression of
A functional food has been defined as “any modified food or food thousands of genes and proteins at a time. These technologies generate
ingredient that may provide a health benefit beyond that of the tradi- a vast amount of gene expression data that accumulates in public re-
tional nutrients it contains” (Snetselaar, 1994). During the last 20 years positories such as the NCBI Gene Expression Omnibus (GEO) (Barrett
there have been substantial efforts to identify bioactive compounds in et al., 2013). Whereas these data remains unclassified by phenotype or
food which might be associated with beneficial biological activities. For experimental condition, the user interface allows easily querying and
example compounds such as long-chain polyunsaturated fatty acids (n- mining the database for experiments.
3 PUFAS), which consumption has been associated with a reduced risk Other databases such as the Broad Institute's Connectivity Map
of cardiovascular disease, are known to act as ligands for cellular re- (CMap) (Lamb et al., 2006) collect highly specific expression data from
ceptors to trigger a signaling cascade that inhibits the expression of cell lines treated with drugs and other chemicals. Such type of tran-
proinflammatory genes (Ferguson, 2009). Also, many natural products, scriptomic data has previously been utilized to establish functional
Abbreviations: GEO, NCBI Gene Expression Omnibus; GES, Gene Expression Signature; TCT, Tocotrienols; LB, Lactobacillus; AMF, Amorfrutin; I3C, Indole-3-carbinol; RSM, Rosemary;
CA, Carnosic Acid; WFNA, Withaferin A; SFN, Sulforaphane; RVT, Resveratrol; TRVT, Transresveratrol; FC, Fold change
⁎
Corresponding author at: Bioinformatics Unit, IMDEA Food Institute, CEI UAM + CSIC, Ctra. De Canto Blanco 8, E-28049 Madrid, Spain.
E-mail address: [email protected] (R. Martín-Hernández).
https://doi.org/10.1016/j.jff.2018.01.021
Received 30 October 2017; Received in revised form 19 January 2018; Accepted 20 January 2018
1756-4646/ © 2018 Elsevier Ltd. All rights reserved.
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
connections between drugs, genes and diseases using computational 2.4. Statistical analysis
approaches. From a transcriptomic point of view, mathematical models
have already been applied on gene expressions data for the identifica- Moderated t-test statistics were applied to microarray features once
tion of pathway responsiveness to drugs (Pratanwanich & Lio, 2014). a linear model was fitted. Statistical significance of the overrepresented
Another approach allows the generation of a list of drugs triggering a GO biological processes in our target gene list was obtained with chi-
similar gene expression pattern at cellular level (Lee et al., 2012) and square test. False discovery rate (FDR) method was employed to adjust
thus possibly sharing a common mechanism of action. From a disease the obtained p-values.
point of view, other approaches consider that a particular gene ex-
pression signature (GES) related to a disease might be reverted using a 3. Results
drug which triggers an opposite GES (Setoain et al., 2015), showing
promising opportunities within the drug repositioning field (Jia et al., 3.1. Data collection
2016). Artificial intelligence, and specifically deep learning algorithms,
has been applied on large transcriptional response data sets with the Experimental gene expression data corresponding to nutrigenomics
aim of classifying various drugs to therapeutic categories solely based experiments was identified from GEO database by launching specific
on their transcriptional profiles (Aliper et al., 2016). However, to the queries. Results were filtered in order to obtain gene expression data
best of our knowledge there is no evidence about such approaches from Homo sapiens as organism, and expression profiling by array as
applied to the emerging field of nutrigenomic studies, seeking to in- study type. Few of these studies, corresponding to human nutritional
vestigate the effect of food and nutrients on gene expression. interventions with large cohorts, were filtered out since we were strictly
We extracted from GEO repository all the available experiments interested on experiments performed on cultured cells. We initially
related to nutrigenomics in human cells to survey the gene expression identified 71 potential GEO studies (Table S1) to be included in our
patterns. The correlation of gene expression patterns can show potential analysis. Of those, 34 studies were filtered out due to different criteria
connections between bioactive compounds, indicating that they may such as studies corresponding to human interventions, lack of replicates
share a common mechanism of action, and allowing the discovery of in the experimental designs, expression data obtained with rare or
new potential therapeutic molecules (Lamb et al., 2006). Here we custom arrays, and expression data corresponding to micro RNA’s. We
present a comprehensive data mining analysis of a set of nutrigenomics ended up with a set of 37 GEO studies.
experiments extracted from GEO database. The assessment of human
cell's gene expression cultured in vitro after treatment with bioactive 3.2. Gene expression analysis workflow
compounds obtained from food should lead to a better characterization
of the molecular mechanisms that confer a beneficial effect to certain Experiments included in each study were carefully assessed before
food products. analysis in accordance with their experimental design, by manually
assigning control and perturbation samples. That is to say, for each
experiment their appropriate control was obtained within the same
2. Materials and methods study. Subsequently, a common computational analysis workflow ap-
plying linear models was used to assess differential expression in each
2.1. Data collection and analysis experiment. Finally, microarray features were annotated with Gene
Symbol and Entrez gene identifiers to allow cross-platform data in-
Studies corresponding to nutrigenomics were identified from GEO tegration. Thus, we obtained a database which includes gene differ-
database. Specific queries were launched containing words such as ential expression data from 81 comparisons among different com-
“nutrient”, “nutrition”, “natural product”, “extract” and “phytochem- pounds, treatments and cell types (Fig. 1) that arise from the 37 GEO
ical”. For data corresponding to Affymetrix platforms, raw data was studies.
downloaded and normalized locally with the RMA algorithm using
specific Bioconductor packages. For data generated by other platforms, 3.3. Cluster analysis
the normalized matrix was directly downloaded for analysis. Gene
differential expression was assessed using LIMMA package from The clustering has been performed using log2 fold change (FC) ex-
Bioconductor. pression values obtained following the gene expression analysis work-
flow. After removing missing values and aggregating expression values
for duplicate gene ID's, we proceeded to integrate gene expression data
2.2. Hierarchical clustering from all the microarray platforms used in our database. Our database
included expression data obtained from 19 distinct microarray plat-
A hierarchical clustering algorithm was applied using gene’s log2 forms. A first limitation is that only overlapping genes represented in all
fold change (FC) from each analyzed experiment as input values. A platforms could be used in our analysis. Therefore, we used the corre-
distance matrix was computed among all the experiments within the sponding Entrez gene ID of the features screened in each platform for
database, using the Euclidean distance as a metric. The agglomeration data integration. Indeed, gene symbols can be hard to match across
method of the clustering process was set to complete. Heatmaps.2 li- platforms because of the continuous updates of gene names, as well as
brary was used for dendrogram and heatmap generation. All the sta- the many to one relationship issues where different gene symbols might
tistical computations were performed using R software. To evaluate the correspond to the same gene.
batch effects presence, normalized gene average expression data for We ended up with a log2 FC expression matrix of 15,591 genes
each experiment was used as data input for hierarchical clustering among 81 variables (experiment comparisons). Such an expression
analysis. matrix included NA values corresponding to the genes that were absent
in a microarray platform. With the aim of grouping experiments which
trigger similar gene expression profiles across the studied cell lines and
2.3. Functional enrichment treatments, we performed a hierarchical clustering on the nu-
trigenomics gene expression matrix obtained from our database
Genecodis3 software was used for functional enrichment using de- (Fig. 2).
fault parameters and selection of GO Biological Process as target an- We observed in the cluster dendrogram that the most remarkable
notations. property is that, as previously observed (Lamb et al., 2006), cell lines
381
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
from common tissues tend to be clustered together. This is coherent and 3.4. Treatments with potential anticancer natural compounds cluster
it is probably due to the expression of groups of genes which are cell together
specific. For example, we identified separate clusters grouping HeLa
cells after treatment with different Tocotrienols (TCT) (alpha, delta and One of the identified and clearly independent cluster (blue), groups
gamma), Myotubes treated with different fatty acids (eicosapentaenoic, together different types of cancer cell lines (MCF7, HT29 and subtypes
linoleic and oleic acids), adipocytes CHUB-S7 cells treated with dif- of MDA cells) exposed to different bioactive compounds with previously
ferent concentrations of folate and vitamin B, and placental cells ex- reported anticancer properties, such as Indole-3-carbinol (I3C) (Quirit
posed to fish oil. Interestingly, clustering based on gene expression et al., 2017), a rosemary (RSM) extract, carnosic acid (CA) (Valdés,
appears to be sensitive to the cell transitional state, as shown in the case Sullini, Ibáñez, Cifuentes, & García-Cañas, 2015), withaferin A (WFNA)
of Keratinocytes treated with retinoic acid. Indeed, differentiating and (Szarc vel Szic et al., 2014), sulforaphane (SFP) (Agyeman et al., 2012)
proliferating Keratinocytes are found in separate clusters, in- and resveratrol (RVT) (Varoni, Lo Faro, Sharifi-Rad, & Iriti, 2016). We
dependently of the retinoic acid type, either retinoic acid or 3,4-dide- hypothesize that these compounds might be exerting their potential
hydroretinoic acid, used for treatments. anticancer activity by a common mode of action on those cancer cells. 2
When paying attention to the clustering analysis from a compound experiments corresponding to treatments with a RSM extract, at con-
point of view, interesting relationships are identified. There is a rela- centrations of 60 and 100 μg/mL respectively on SW620 cells, were not
tively big cluster (green) which groups different cell types after treat- grouped within the anticancer cluster. This fact might be explained
ments with molecules like TCT (compounds naturally occurring in ve- because the compound concentration used for this treatment was too
getable oils), lysophosphatidic acid (a natural bioactive lipid with high and triggered a pronounced cell death.
growth factor-like functions), hydroxycholesterol, two amorfrutin However, we should also consider that all of these clustering ob-
(AMF) extracts and treatments with different strains of Lactobacillus servations may be influenced by previously reported non-biological
(LB). Paradoxically, four different AMF extracts cluster in two relatively factors such as experiments generated in different laboratories/techni-
distant clusters. Those treatments were performed with the same con- cians, and experimental protocols (Luo et al., 2010), the so called batch-
centration of 30 µM which evidences two distinct biological activities effects. Such a bias appears whenever merging expression data from
between AMF extracts 2–3 and AMF extracts 1–4. different microarrays experiments for an integrative analysis, prior to a
Thus, we can infer that these compounds may share common bio- gene differential expression analysis, which should yield more robust
logical properties in a cell type independent manner. Indeed, all of statistics on differential gene expression results in most cases, for in-
these compounds have previously demonstrated promising antioxidant stance when studying gene expression levels in a specific disease or
and anti-inflammatory properties (Ahsan, Ahad, Iqbal, & Siddiqui, following medical treatment. In our case, each experiment has been
2014; Contos, Ishii, & Chun, 2000; Fuhr, Rousseau, Plauth, Schroeder, analyzed independently and the resulting gene differential expression
& Sauer, 2015; Richard, Kefi, Barbe, Bausero, & Visioli, 2008). With the data (log2 FC expression values) has been combined in a large matrix in
aim of getting insights into the effect of such compounds on cells from a order to perform a data mining exercise. Because in this work we are
functional point of view, we searched for common features among the not performing an integrative analysis, our results cannot be affected by
top 3000 statistically significant differentially expressed genes in each such bias. To further test that batch effect does not influence our ana-
experiment (green cluster, Fig. 2). However, we were not successful at lysis, we performed a clustering analysis of a matrix containing nor-
identifying any significant overlap among the top 3000 statistically malized expression values from the same experiments (Fig. S1), which
significant differentially expressed genes. would mimic an integrative analysis prior to the gene differential ex-
pression processing. In this new dendrogram we can clearly appreciate
the batch-effects bias. Indeed, all the included experiments strictly
382
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
Fig. 2. Cluster dendrogram of the hierarchical clustering result. Log2 fold change values obtained after differential expression of each experiment were used as input data. The two
clusters highlighted in blue and green color correspond to clusters grouping experiments with potential cancer protective and anti-oxidant/anti-inflammatory compounds respectively.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
cluster by study (experiments generated in the same lab), revealing the WFNA, SFP and RVT at 150 nM. As previously stated these 4 natural
presence of batch-effects, and as a consequence no biological connec- compounds have previously shown potential antiproliferative effects on
tions are detected anymore among the studied food compounds. cancer cells. Experiments corresponding to treatments with I3C were
Therefore, we can assume that our clustering analysis is not affected by not included in the comparison since they were found in a separated
the so called batch-effects. branch of the same cluster. The RSM treatment performed on HT29
We decided to compare the most statistically significant differen- cells experiment was not included in this comparison since CA is mainly
tially expressed genes of each experiment in order to get insights into a found in RSM. In addition, RSM and CA experiments show a high
common mode of action for these potential anticancer compounds. correlation since they are found in two close branches within the same
Therefore, in order to identify common and stable differential expressed cluster. As well as for RVT treatments, since both experiments are found
genes, we chose one experiment from the most populated branch of this in the same cluster, we decided to include the treatment with 150 nM as
cluster as a representative for each different anticancer compound: CA, a representative for RVT treatments.
383
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
3.5. Identification of a cancer protective gene expression signature underlying the main hypothesis of the original experiment.
384
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
Fig. 4. Heatmap of nutrigenomics experiments based on the identified cancer protective gene signature. The represented log2 fold change values were normalized at row level.
2017; Zhang et al., 2016). functional foods with potential cancer protective properties, tailored for
The present investigation shows the tremendous potential of pub- human nutrition with a focus on both disease prevention and clinical
licly available experimental data for making new discoveries by ap- nutrition purposes, thus setting up a framework for the use of functional
plying in silico approaches. To our knowledge, this work is the first food on human health.
applying such data analysis approaches within the nutrigenomics field.
Combining the results of multiple studies is a powerful tool that allows Acknowledgements
establishing solid correlations and relationships among variables, in our
case gene expression, studied by different labs and platforms labs. In This research was supported in part by the Spanish Agencia Estatal
addition it also serves as a method for checking the reproducibility of de Investigación and the European FEDER Funds (AGL2016-78922-R)
results. We believe that in silico approaches are gaining weight in the to AD and RM-H. AD lab is supported in part by the Fundación Ramón
present post-genomics era as more experimental data becomes publicly Areces to the project “Modulation of exosomes transporting miRNAs
available. However the combination of such approaches with wet lab and lncRNAs for intercellular communication as therapeutic tools in
strategies, as gene expression confirmation by RT-PCR, will be a rule of dyslipidemia treatment”.
thumb and would be hardly avoided. In our case, the original papers
studying the effects of RVT, SFN, WFNA and CA focused on confirming Authors Contributions Statement
the expression level of specific genes, which were found to be im-
plicated in cholesterol metabolism and chromatin remodeling (Szarc vel RM-H and AD designed the study. RM-H performed the study and
Szic et al., 2014; Valdés et al., 2015). analyzed data. RM-H and AD wrote the main manuscript text. All au-
Natural products have been an important source of lead compounds thors reviewed the manuscript.
for drug discover, because of their vast chemical diversity, good drug-
like properties and potential interactions with multiple cellular target
Competing financial interests
proteins. For instance, the U.S. National Center for Complementary and
Integrative Health (NCCIH) provides a list of a wide variety of natural
The authors declare no competing financial interest.
products libraries (https://nccih.nih.gov/grants/naturalproducts/
libraries) for screening purposes. The molecular signature identified
Appendix A. Supplementary material
in the data analysis presented in this work might provide potential
biomarkers of cancer disease progression, and could therefore be used
Supplementary data associated with this article can be found, in the
for prognosis. Importantly such a GES could be used as a benchmark for
online version, at http://dx.doi.org/10.1016/j.jff.2018.01.021.
385
R. Martín-Hernández et al. Journal of Functional Foods 42 (2018) 380–386
References (2006). The Connectivity Map: Using gene-expression signatures to connect small
molecules, genes, and disease. Science, 313(5795), 1929–1935.
Lee, J. H., Kim, D. G., Bae, T. J., Rho, K., Kim, J. T., Lee, J. J., ... Kim, S. (2012). CDA:
Agyeman, A. S., Chaerkady, R., Shaw, P. G., Davidson, N. E., Visvanathan, K., Pandey, A., Combinatorial drug discovery using transcriptional response modules. PLoS ONE,
& Kensler, T. W. (2012). Transcriptomic and proteomic profiling of KEAP1 disrupted 7(8), e42573.
and sulforaphane-treated human breast epithelial cells reveals common expression Liu, S. L., Lin, H. X., Qiu, F., Zhang, W. J., Niu, C. H., Wen, W., ... Guo, L. (2017).
profiles. Breast Cancer Research and Treatment, 132(1), 175–187. Overexpression of Kinesin family member 20A correlates with disease progression
Ahsan, H., Ahad, A., Iqbal, J., & Siddiqui, W. A. (2014). Pharmacological potential of and poor prognosis in human nasopharyngeal cancer: A retrospective analysis of 105
tocotrienols: A review. Nutrition & Metabolism (London), 11(1), 52. patients. PLoS ONE, 12(1), e0169280.
Aliper, A., Plis, S., Artemov, A., Ulloa, A., Mamoshina, P., & Zhavoronkov, A. (2016). Luo, J., Schumacher, M., Scherer, A., Sanoudou, D., Megherbi, D., Davison, T., ... Zhang,
Deep learning applications for predicting pharmacological properties of drugs and J. (2010). A comparison of batch effect removal methods for enhancement of pre-
drug repurposing using transcriptomic data. Molecular Pharmaceutics, 13(7), diction performance using MAQC-II microarray gene expression data.
2524–2530. Pharmacogenomics Journal, 10(4), 278–291.
Barrett, T., Wilhite, S. E., Ledoux, P., Evangelista, C., Kim, I. F., Tomashevsky, M., ... Pratanwanich, N., & Lio, P. (2014). Pathway-based Bayesian inference of drug-disease
Soboleva, A. (2013). NCBI GEO: Archive for functional genomics data sets–update. interactions. Molecular BioSystems, 10(6), 1538–1548.
Nucleic Acids Research, 41(Database issue), D991–995. Quirit, J. G., Lavrenov, S. N., Poindexter, K., Xu, J., Kyauk, C., Durkin, K. A., ... Firestone,
Contos, J. J., Ishii, I., & Chun, J. (2000). Lysophosphatidic acid receptors. Molecular G. L. (2017). Indole-3-carbinol (I3C) analogues are potent small molecule inhibitors
Pharmacology, 58(6), 1188–1196. of NEDD4-1 ubiquitin ligase activity that disrupt proliferation of human melanoma
DeBusk, R. M., Fogarty, C. P., Ordovas, J. M., & Kornman, K. S. (2005). Nutritional cells. Biochemical Pharmacology, 127, 13–27.
genomics in practice: Where do we begin? Journal of the American Dietetic Association, Ramirez de Molina, A., Vargas, T., Molina, S., Sanchez, J., Martinez-Romero, J., Gonzalez-
105(4), 589–598. Vallinas, M., ... Reglero, G. (2015). The ellagic acid derivative 4,4′-di-O-methylellagic
Ferguson, L. R. (2009). Nutrigenomics approaches to functional foods. Journal of the acid efficiently inhibits colon cancer cell growth through a mechanism involving
American Dietetic Association, 109(3), 452–458. WNT16. Journal of Pharmacology and Experimental Therapeutics, 353(2), 433–444.
Forsyth, C. B., Farhadi, A., Jakate, S. M., Tang, Y., Shaikh, M., & Keshavarzian, A. (2009). Richard, D., Kefi, K., Barbe, U., Bausero, P., & Visioli, F. (2008). Polyunsaturated fatty
Lactobacillus GG treatment ameliorates alcohol-induced intestinal oxidative stress, acids as antioxidants. Pharmacological Research, 57(6), 451–455.
gut leakiness, and liver injury in a rat model of alcoholic steatohepatitis. Alcohol, Setoain, J., Franch, M., Martinez, M., Tabas-Madrid, D., Sorzano, C. O., Bakker, A., ...
43(2), 163–172. Pascual-Montano, A. (2015). NFFinder: An online bioinformatics tool for searching
Fuhr, L., Rousseau, M., Plauth, A., Schroeder, F. C., & Sauer, S. (2015). Amorfrutins are similar transcriptomics experiments in the context of drug repositioning. Nucleic
natural PPARgamma agonists with potent anti-inflammatory properties. Journal of Acids Research, 43(W1), W193–199.
Natural Products, 78(5), 1160–1164. Snetselaar, L. (1994). Opportunities in the nutrition and food sciences: Research chal-
Gil-Zamorano, J., Martin, R., Daimiel, L., Richardson, K., Giordano, E., Nicod, N., ... lenges and the next generation of investigators. Journal of the American Dietetic
Davalos, A. (2014). Docosahexaenoic acid modulates the enterocyte Caco-2 cell ex- Association, 94(6), 598.
pression of microRNAs involved in lipid metabolism. Journal of Nutrition, 144(5), Szarc vel Szic, K., Op de Beeck, K., Ratman, D., Wouters, A., Beck, I. M., Declerck, K., ...
575–585. Vanden Berghe, W. (2014). Pharmacological levels of Withaferin A (Withania som-
Gonzalez-Vallinas, M., Molina, S., Vicente, G., de la Cueva, A., Vargas, T., Santoyo, S., ... nifera) trigger clinically relevant anticancer effects specific to triple negative breast
Ramirez de Molina, A. (2013). Antitumor effect of 5-fluorouracil is enhanced by cancer cells. PLoS ONE, 9(2), e87850.
rosemary extract in both drug sensitive and resistant colon cancer cells. Valdés, A., Sullini, G., Ibáñez, E., Cifuentes, A., & García-Cañas, V. (2015). Rosemary
Pharmacological Research, 72, 61–68. polyphenols induce unfolded protein response and changes in cholesterol metabolism
Gonzalez-Vallinas, M., Molina, S., Vicente, G., Zarza, V., Martin-Hernandez, R., Garcia- in colon cancer cells. Journal of Functional Foods, 15(Supplement C), 429–439.
Risco, M. R., ... Ramirez de Molina, A. (2014). Expression of microRNA-15b and the Varoni, E. M., Lo Faro, A. F., Sharifi-Rad, J., & Iriti, M. (2016). Anticancer molecular
glycosyltransferase GCNT3 correlates with antitumor efficacy of Rosemary diterpenes mechanisms of resveratrol. Frontiers in Nutrition, 3, 8.
in colon and pancreatic cancer. PLoS ONE, 9(6), e98556. Zhang, W., He, W., Shi, Y., Gu, H., Li, M., Liu, Z., ... Zhang, Y. (2016). High expression of
Jia, Z., Liu, Y., Guan, N., Bo, X., Luo, Z., & Barnes, M. R. (2016). Cogena, a novel tool for KIF20A is associated with poor overall survival and tumor progression in early-stage
co-expressed gene-set enrichment analysis, applied to drug repositioning and drug cervical squamous cell carcinoma. PLoS ONE, 11(12), e0167449.
mode of action discovery. BMC Genomics, 17, 414.
Lamb, J., Crawford, E. D., Peck, D., Modell, J. W., Blat, I. C., Wrobel, M. J., ... Golub, T. R.
386