https://orcid.org/0000-0002-5460-3750
Figures
Abstract
Propolis is a promising natural product that has been extensively researched and studied for its potential health and medical benefits. The lack of requisite high oil-containing propolis and existing variation in the quality and quantity of essential oil within agro-climatic regions pose a problem in the commercialization of essential oil. As a result, the current study was carried out to optimize and estimate the essential oil yield of propolis. The essential oil data of 62 propolis samples from ten agro-climatic areas of Odisha, as well as an investigation of their soil and environmental parameters, were used to construct an artificial neural network (ANN) based prediction model. The influential predictors were determined using Garson’s algorithm. To understand how the variables interact and to determine the optimum value of each variable for the greatest response, the response surface curves were plotted. The results revealed that the most suited model was multilayer-feed-forward neural networks with an R2 value of 0.93. According to the model, altitude was found to have a very strong influence on response, followed by phosphorous & maximum average temperature. This research shows that using an ANN-based prediction model with a response surface methodology technique to estimate oil yield at a new site and maximize propolis oil yield at a specific site by adjusting variable parameters is a viable commercial option. To our knowledge, this is the first report on the development of a model to optimize and estimate the essential oil yield of propolis.
Citation: Nayak G, Sahu A, Bhuyan SK, Akbar A, Bhuyan R, Kar D, et al. (2023) Developing a computational toolbased on an artificial neural network for predicting and optimizing propolis oil, an important natural product for drug discovery. PLoS ONE 18(5): e0283766. https://doi.org/10.1371/journal.pone.0283766
Editor: Sathishkumar V. E., Jeonbuk National University, REPUBLIC OF KOREA
Received: November 3, 2022; Accepted: March 15, 2023; Published: May 8, 2023
Copyright: © 2023 Nayak et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the paper and its Supporting Information files.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Honeybees generated propolis, a natural resinous mixture made up of plant parts, buds, and exudates. Propolis is now a natural remedy available in a variety of topical forms in many health-food stores. It is used in cosmetics and as a popular alternative medicine for treating a variety of ailments. Cold syndrome (upper respiratory tract infections, common cold, and flu-like disorders) as well as wound healing, burn treatment, acne, herpes simplex, genitals, and neurodermatitis, are treated by propolis. Propolis is also used in mouthwashes and toothpaste to treat gingivitis and stomatitis, as well as to prevent caries. It is found in a variety of cosmetics, as well as health, foods, beverages, Capsules, mouthwash, creams, throat lozenges, powder, and numerous refined items from which the wax has been removed are all commercially available. It has antibacterial, antiviral, and antioxidant effects [1,2]. It is frequently utilized in pharmacology, cosmetics, and human & veterinary medicine. Propolis oil from various geographical origins exhibits varying bioactivity, such as antibacterial activities [3,4] antifungal [5], anticancer [6], and antioxidant [7], and also has a therapeutic effect on anxiety [8]. In the international market, the cost of Bee propolis powder is up to 15000$ per kg. There is currently no credible information on total propolis production around the world, and we continue to lack detailed production in India. Some market study has revealed that global propolis output currently stands at several thousand tonnes. Propolis is in high demand in Japan, Korea, and Taiwan. Propolis collection among Indian beekeepers is non-existent. This is primarily due to a lack of knowledge about the quality of indigenous propolis and its business potential. As a result, scientific research projects to investigate the properties of indigenous propolis are urgently needed. India’s vast floral and crop diversity, as well as its diverse climatic conditions, propolis chemical composition is likely to change across the country. According to numerous studies, the main physiologically active chemicals in propolis are caffeic acids, flavonoids, and phenolic esters. Propolis has a complex chemical composition & its biological effects cannot be directly attributed to these components. The qualitative and quantitative composition of propolis, in our opinion, has a significant impact on their biological activity. Its components and biological activity are influenced by a variety of conditions, including geographical location, time of collection, and plant source [9,10].
As a result, it is critical to investigate the level of heterogeneity in propolis content throughout Odisha’s various agro-climatic areas to better understand the attributes of indigenous propolis for each location. Furthermore, because their production is strongly impacted by environmental conditions, basic chemo typing would not be able to identify the features of indigenous propolis. Various statistical methods are applied to determine the relationship between biochemical content and environmental conditions, Common statistical techniques like correlation and multiple linear regressions (MLR) analyses are only used to identify linear relationships and do not accurately deal with non-linear data [11]. Due to superior prediction performance, artificial neural networks (ANN) are now frequently employed to construct and map non-linear relationships between inputs and outputs. The use of ANN modeling simulates how the human brain works [12]. It is chosen because it can directly learn from situations without analyzing the parameters using statistical techniques [13]. An ANN is divided into three primary sections: the input layer, hidden layers, and the output layer of neurons [14]. The neurons in the input layers take in the input data, and then the incoming data is normalized before being passed to the hidden layer [15]. Each neuron in the layer below calculates a linear combination of the data from the input layer’s neurons, and then adds weight values associated with certain nodes to it.The neurons in the hidden layer combine the linear information from the input layer with a transfer function (a particular non-linear function), which results in the output being the projected model [14]. About the environment and edaphic factors, the ANN model has been used to predict the bioactive content of Podophyllum hexandrum’s podophyllotoxin [16], Hypericum perforatum L’s hyperforin, hypericin, and pseudohypericin [11], and Bacopa monnieri’s bacoside A content [17].
In the current work, it will be required to investigate the climatic parameters affecting propolis oil content in various agroclimatic zones in Odisha. So an Artificial Neural Network (ANN) model based on propolis oil content and climatic factors can be developed to predictoptimum yield for the proper regions/sites and to optimize propolis oil yield at a specific region through management of the sensitive and changeable parameters. The flow diagram of this study has been represented in Fig 1.
Materials and methods
Propolis sample collection
From late summer June to October 2021, propolis samples were collected from 31 places in Odisha’s ten agro-climatic regions at various altitudes (0.1–1202 m) (Table 1). Propolis samples were obtained in two duplicates from each site. The distance between duplicates was between 2 to 5 meters. To eliminate dust particles, fresh Bee propolis was collected and cleaned with distilled water. Before calculating the propolis oil content, the cleaned propolis was air-dried at room temperature. Each site’s soil samples were collected in duplicates and taken to the lab for investigation of soil nutrients. From June to October 2021, monthly averages of adequately documented data on environmental parameters such as temperature, humidity, and rainfall were obtained from each site (S1 Table).
Extraction and quantificationof propolis oil
The propolis oil was extracted by combining 10g of cleaned propolis with 200 ml of any refined food quality oil (e.g. coconut oil, sunflower oil, etc.) or 100 g of butter. After that, the components were gently heated in a water bath (no more than 50°C) for around 10 minutes while being constantly stirred. The oil was then purified and stored in the dark in well-sealed containers for future use.
Quantitative analysis of soil
The samples of soil were taken from each sampling site in Odisha’s ten agro-climatic areas. The soil data from different Agro-climatic regions of Odisha were represented in S2 Table. Approximately 200 g of soil was collected and sieved through a 2 mm mesh. For nutrient analysis, the fine soil particles were extracted, and the pH of a 1:2 soil: water suspension was evaluated using the Systronics pH meter after 30 minutes of equilibration with occasional stirring (Model MKVI). The Walkley and Black wet digestion methodwereused to assess the organic carbon content as described in soil chemical analysis, [18] whereas the total nitrogen was determined using alkaline KMnO4 as described in soil chemical analysis [19]. In an 800ml Kjeldahl flask, 20 g of soil sample was added to 100 ml of 0.32 percent KMnO4 solution, followed by 2.5 percent NaOH solution and distilled water. In a 250 mL conical flask containing 20 mL boric acid (2%) and mixed indicator, distillation was continued and collected in a receiver tube. The distillate was titrated in a burette against 0.02 N H2SO4 to a pink color endpoint and the amount of accessible nitrogen was determined.
Brays No-1 techniques were used to estimate total phosphorus in soil samples. The 2 grams of soil were extracted using 40 ml of Bray’s-1 solution (0.025 N HCl and 0.03 NH4F) and mechanically agitated for 5 minutes before filtering through the Whatman filter paper. A 0.5 ml aliquot was transferred to a flask with a capacity of 25 mL. The volume was increased to 25 ml by adding distilled water and 5 mL ammonium molybdate solution. The volume was made up to the mark with diluted SnCl2 (0.5 ml was diluted to 66 ml). A spectrophotometer (Model: Systronics 166) was used to measure phosphorus content at 660 nm. The concentration was determined using a standard graph made from various phosphorus concentrations. 5 g of soil samples were placed in a 100 ml conical flask with 25 ml of 1 N NH4OAc solution, and the potassium content of the soil was measured. The filtrate was then shaken for 5 minutes with a mechanical shaker, and the potassium content was evaluated using a flame photometer (Model: Sistronics128).
Statistical analysis
Data exploration.
All computational work (model development, plot generation, etc.) was performed by using R [20] & Microsoft Excel 2013. The data set consists of 12 features & 62 instances. Out of the features, 11 are predictors. The predictors are soil pH (pH), organic carbon (oc), nitrogen (nitro), the phosphorous (pho), potassium (pot) content the of soil, maximum relative humidity (mxrel), minimum relative humidity (mnrel), average rainfall (avgrf), maximum average temperature (mxt), minimum average temperature (mnt) and altitude (alt). Propolis content is the response. Standard deviations for all features are calculated by using the mlbench [21] library. The formula is provided below.
Eq 1
where x&
are value of each observation & mean of all observations respectively.
Pearson’s correlation coefficient between features and data distribution in each feature was evaluated by using the package psych [22]. The correlation values were provided in the plot in the form of numeric values as well as a correlation ellipse.
Eq 2
where x &y are the values of the two variables;
are the respective means.
A panel plot was generated to find out the data disribution, and multicolinearity among the variables. For this purpose, a psych package was used.
Data splitting.
The dataset is divided into three sets, train, test & validation with 70%, 20% & 10% of data. A train set was used to develop the model by training. A test set was used to evaluate the model. Finally, the model was validated by using a validation set.
InitialModel development.
A comparative modeling approach was applied to identify the best-performing model for the dataset. Several linear algorithms viz. linear regression model (LM), generalized linear regression (GLM), Penalized linear regression model (GLMNET); nonlinear algorithms viz. K-nearest neighbors (KNN), support vector machine (SVM), artificialneural network (ANN) &tree-based models viz. classification & regression tree (CART) &random forest (RF) were developed & evaluated. The resampling method used wascross-validation. The data waspreprocessed withmin-max normalization. Coefficient of determination (R—squared), root mean square error (RMSE), andmean absoluteerror (MAE)were used to evaluatethe models. For this purpose, Caret (classification & regression training) package was used.
Model evaluation & selection.
Out of theabovemodels,the model which had the highest R -squaredvalue & lowest root mean square error (RMSE), mean absolute error (MAE) & coefficient of determination (R-squared) values of model for the train, test & validation data set were calculated using the following equations.
The caret (classification & regression training) package was used to develop the final model. Data was scaled using minimum-maximum normalization. The train set was resampled with cross-validation during training. A grid-tuning approach was applied to find an optimum number of layers & nodes in each layer. A resilient backpropagation algorithm with weight bracketing was used for training. The Logistic function was selected as the activation function. The learning rate was kept at 0.4. The Sum of squared errors (SSE) was used for the calculation of errors.
Eq 6
where y &
are the actual response & predicted response respectively.
Model tuning.
The ANN model was further fine-tuned to improve its prediction performance. The data were preprocessed using the minimum-maximum normalization method. Several hyperparameters are tuned to develop the final model. Those were the number of hidden layers, the number of nodes in each layer, the number of folds in resampling, the activation function & the learning rate. The input data is resampled with cross-validation. Agrid-tuning approach was applied to find an optimum number of layers & nodes in each layer.
The end model was selected based on several metrics viz. symmetricmean absolute percentage error (SMAPE), Nash—Sutcliffe efficiencycoefficient (NSE) along with RMSE, MAE & R squared values. The model was also analyzed with regression & slope-intercept tests as proposed by Rocabruno-Valdes et al. (2019) [23].
SMAPE iscalculated byusing the following formula.
Where predicted & actual are the predicted values by the final model & actual values respectively.
Several authors have used SMAPE to evaluate & improve their developed models by using SMAPE [24,25]. SMAPE was calculated with the help of Microsoft Excel 2013.
Nash—Sutcliffe efficiency coefficient (NSE) was used for the final selection of the model. When the NSE value is 1, the predicted value and actual values are the same. If the NSE value is zero, the predicted values match the mean of theobserved values. When the value is below less than 0, the model is not significant [26].
Effect of predictors.
Partial dependence plots (PDP) were generated to investigate the interaction of predictors with the response. These plots were generated by using the PDP library. PDP plots are used to interpret the output of complex machine-learning models [27]. The use of linear plots is ineffective for explaining the complex relationships between various variables and the response. In this study, single-variable &multiple-variable PDPs are generated. Smoothing is applied by using locally weighted regression (LOESS) in the case of single variable PDPs. It has popularity in the smoothing of scatter plots [28]. LOESS can perform well even if the response is a nonlinear function of the predictor [29]. The relationships of the response variable with two predictors are represented by two-dimensional contour & three-dimensional PDPs.
Sensitivity analysis.
It’s critical to figure out what the most crucial factors are that influence propolis oil yield. As a result, the influential predictors were determined using Garson’s algorithm. The relative importance of a certain predictor is obtained by finding the link strengths between the node [30]. The variable importance was evaluated by using Neural Network Tools [31] library.
Results
Data exploration
The panel plot in Fig 2 shows four different representations of the data set through scatter plots, ellipse plots, histograms & Pearson’s correlation coefficients. The scatterplots with trend lines represent the linear relationship between two variables. The correlation ellipse shows the correlation strength. The higher the stretch, the higher the correlation coefficient. The histogram plots show the data distribution for each variable.Pearson’s correlation coefficients are represented in numeric values. The correlation coefficient values range from-0.48 (between soil phosphorous content to altitude) to +0.73 (between minimum relative humidity & minimum average temperature). The positive& negative signs of the correlation coefficients represent negative & positive correlation respectively. No correlation was observed between soil phosphorous content & maximum average temperature. A negligible correlation was observed between pH & organic carbon content (correlation coefficient = 0.06), pH & soil nitrogen content (correlation coefficient = -0.07), etc. A weak correlation was observed betweenphosphorous & minimum relative humidity (correlation coefficient = 0.1), organic carbon & minimum average temperature (correlation coefficient = 0.34), etc.A strong correlation is observed between minimum relative humidity& minimum average temperature (correlation coefficient = 0.73).
From the panel plot (Fig 2), it is evident that some of the predictors viz. pH, maximum relative humidity (mxrel), minimum relative humidity (mnrel), maximum average temperature (mxt) & minimum average temperature (mnt) are normally distributed. Some of the predictors have a positively skewed distribution. organic carbon (oc), soil nitrogen content (nitro), phosphorous (pho), average rainfall (avgrf), and altitude (alt) show positive skewness. Soil potassium content (pot) showed platykurtik distribution.Standard deviations of all variables are provided in Table 2.
Model selection
Eight different modelswere developed. The MAE,RMSE & R-squaredof the developed modelswere provided in Table 3 & Fig 3. Outof these models, the artificialneural network (ANN)model was found to have a better performancemeasure as compared to other algorithms. Thefinal model had five layers withone input layer, one output layer, andthreehidden layers. The inputand output layers had 11and 1 neurons or nodes respectively.Thefirst hidden layer had 12 nodes;the second &third hidden layershad 5& 3nodes respectively. The model along with layers, nodes & weights was provided in Fig 4.
The final model was selected based on the previously discussed metrics i.e. RMSE, MAE & R- squared (Table 3 & Fig 3).Thefinal model’s performancewas further evaluatedwith SMAPEscores,provided in Table 4 & Fig 5. The NSE values fortrain,testand validationsets are 0.99, 0.83 & 0.92 respectively.
Theregression statistics are presentedin Table 5. The table showstheR-squared, adjusted R-squared values,and standard error of regressionfor train,test, and validation sets.
Theresults of slopeand interceptanalysisfor train, test,and validationsetsare providedin Table 6A, 6B and 6C respectively.
The predictions & the actual responses for the train set were provided in Table 7 & Fig 6A. The RMSE, MAE & R-squared values of the model for the train set were 0.08, 0.06 & 0.99 respectively. The model was evaluated using the test data set once it had been trained. The RMSE, MAE &R squared values were 2.17, 0.85 & 0.85 respectively.Further, the model was validated with a validation set. The model also performed well with RMSE, MAE & R squared values of 1.11, 0.94 & 0.94 respectively.The predictions & the actual responses for the test &validation sets were provided in Tables 8 & 9 and Fig 6B & 6C respectively.To the best of our knowledge, this is the first study to provide insight into propolis oil content prediction.
(a) train; (b) test and (c) validation data.
As measured by statistical criteria such as coefficient of determination (R2) and root mean square error (RMSE) values, the ANN model built in this study demonstrated strong predictive potential for propolis oil content. The stronger the ANN model is, the closer the R2 value is to 1 and the lower the RMSE value. As a result, it’s possible to conclude that the model’s propolis oil content prediction is quite accurate.
Identifying significant predictor
According to the model, altitude was found to have a very strong influence on response, followed by phosphorous & maximum average temperature. Minimum relative humidity was found to have the least effect on propolis content. The relative importance of all variables on importance is provided in Fig 7 Garson’s method of importance calculation removes the predictor which is insignificant,and soil potassium content was not shown. According to the model, altitude was found to have a very strong influence on response, followed by phosphorous & maximum average temperature. Garson’s algorithm was used to determine the influential predictors.
Effect of individual predictors on propolis content
Single variable PDPs are generated for all predictors & provided in Fig 8A–8K. The multicollinearity among the features of the data set is described. Most of the correlation among variables is weak to moderate. Minimum average temperature & minimum average relative humidity have a strong correlation. The variation of propolis content with altitude (the most significant predictor) was provided,with a lower value of altitude, propolis content was found to be higher. From the figure, it is evident that the response value gradually decreases with an increase in altitude. The propolis content reaches a minimum between altitudes of 400m to 800m. After 800m., propolis content increases, but at a lower rate. Similarly, the optimum phosphorous content was found to be between 125 & 200 kg/ha (Fig 8D). Propolis content was found to be higher when both maximum & minimum average temperature values are lower. With the increase in temperature, propolis content decreases gradually (Fig 8I and 8J). Propolis content was lower when the organic carbon content of the soil was between 1.5 to 4 kg/ha. Maximum relative humidity was negatively correlated with propolis content (Fig 8F). Propolis content was maximum when the minimum relative humidity is approximately 60 (Fig 8G). The effect of nitrogen, pH, average rainfall & potassium content of the soil was represented in Fig 8A, 8C, 8H and 8E respectively. The variable importance of input parameters on oil yield (output) is shown in Fig 7.
(a)Partial dependence plot of response with respect to pH; (b) Partial dependence plot of response with respect to organic carbon; (c) Partial dependence plot of response with respect to nitrogen; (d) Partial dependence plot of response in terms of phosphorous; (e) Partial dependence plot of response in terms of potassium; (f) Partial dependence plot of response in terms of maximum relative humidity; (g) Partial dependence plot of response in terms of minimum relative humidity; (h) Partial dependence plot of response in terms of average rainfall; (i) Partial dependence plot of response in terms of maximum average temperature; (j) Partial dependence plot of response in terms of minimum average temperature; (k) Partial dependence plot of response in terms of altitude.
Mutual effect of two predictors on response
Partial dependence plots with two variables show the mutual contribution of the variables to the response. Such plots help identify the optimum range of predictor values for a maximum value of the response. Two variable PDPs Figs 9A–9K and 10A–10K are generated for the top five important variables (altitude, phosphorous content, maximum average temperature, minimum average temperature & soil organic carbon content) to understand the mutual influence of these variables on response. For each pair of predictors, two PDPs are generated; a 2D contour plot & 3D partial dependence plot. In each plot, the color scale on the right-hand side shows the color as a measure of propolis content.
A. Contour plot of altitude, min. rel. humidity & propolis content. B. Contour plot of altitude, min. temperature & propolis content. C. Contour plot of altitude, max. temperature & propolis content. D. Contour plot of altitude, organic carbon & propolis content. E. Contour plot of altitude, phosphorus & propolis content. F. Contour plot of min. temperature, organic carbon & propolis content. G. Contour plot of min. temperature, max. temperature & propolis content. H. Contour plot of max. temperature, organic carbon & propolis content. I. Contour plot of phosphorus, min. temperature & propolis content. J. Contour plot of phosphorus, max. temperature & propolis content. K. Contour plot of phosphorus, organic carbon & propolis content.
A. 3D PDP of altitude, min. rel. humidity & propolis content. B. 3D PDP of altitude, min. temperature & propolis content. C. 3D PDP of altitude, max. temperature & propolis content. D. 3D PDP of altitude, organic carbon & propolis content. E. 3D PDP of altitude, phosphorus & propolis content. F. 3D PDP of min. temperature, organic carbon &propolis content. G. 3D PDP of min. temperature, max. temperature & propolis content. H. 3D PDP of max. temperature, organic carbon & propolis content. I. 3D PDP of phosphorus, min. temperature & propolis content. J.3D PDP of phosphorus, max. temperature & propolis content. K. 3D PDP of phosphorus, organic carbon & propolis content.
From these plots, it is evident that lower value of altitude i.e. 0 to 200m approx. (Figs 9B–9E and 10B–10E) and a soil phosphorous content between 50 to 175kg/ha hasa favourable effect on propolis content (Figs 9E, 9I–9K, 10E and 10I–10K). Similarly, lower values of maximum & minimum average temperature i.e. below 33°C& 17°C were found to increase propolis content. For organic carbon, two ranges exist (as evident from a single variable PDP for organic carbon and two variable PDPs for organic carbon, maximum average temperature, Figs 9F and 10F & organic carbon-altitude, Figs 9D and 10D) which favor higher propolis content. The approximate ranges are between 5 to 9 kg/ha and below 2kg/ha.
Discussion
A data distribution study is necessary before implementing a machine learning model as machine learning models are also influenced by the data distribution of predictors. In this data set, not all variables are normally distributed. In such cases, model selection plays a crucial role. Some models need further data processing. However, tree-based models & artificial neural networks can also perform well when the data distribution is not normal [32,33].
A correlation study among the predictors has significance in model evaluation. When Pearson’s correlation coefficient value is between 0 to 0.1, 0.1 to 0.39, and 0.4 to 0.69, the correlation is negligible, weak & moderate respectively. Correlation is strong when Pearson’s correlation coefficient value is between 0.7 and 0.89 & very strong when the correlation coefficient value is between 0.9 and 1 [34]. A correlation study among the predictors has significance in model evaluation. The machine learning algorithms are affected if a correlation exists among the predictors [35].
The exploratory analysis shows the correlations among the predictors are within a range ofnegligible to strong. Hence, the model’s performance may get influenced due to such multicollinearity among the variables. In such cases, tree based models, and artificial neural networks perform well & are not influenced by multicollinearity among the predictors.
From the comparative study of eight different machine learning models, it is evidentthat the artificial neural network (ANN) model has outperformed all other models with the lowest MAE & RMSE values & highest value of the R-squared (Table 3 and Fig 3).
The use of artificial neural networks (ANN) is suggested as a promising way of predicting propolis oil content. This technology not only gives new possible approaches for bio-compound study into other plants and other environmental conditions, but it also provides new potential approaches for bio-compound research into other plants and other environmental situations. ANN has been proposed by many academics as a predictive method for optimizing operating parameters during the extraction of diverse natural products [36–44].
Akbar et al. (2018) reported on the use of ANNmodeling for the optimization and prediction of essential oil yield in turmeric (Curcuma longa L.). By analyzing the soil and environmental conditions in the eight agro-climatic areas of Odisha and using data on the essential oil of 131 turmeric germplasms, they created a model. The ANN model was trained and tested on each sample with 11 parameters. With an R2 value of 0.88, the results demonstrated that multilayer-feedback neural networks with 12 nodes (MLFN-12) were the most appropriate and logical model to utilizeThis study shows that an ANN-based prediction model is a good method for forecasting oil output at a new location and for optimizing turmeric oil yield at a specific site by adjusting the prediction model’s changeable parameters, and as a result, it has enough commercial relevance.Niazian et al. (2018) used artificial neural networks (ANN) and multiple regression models (MLR) to forecast the oil content of ajowan based on easily quantifiable plant characteristics. Four characteristics (number of rays, number of pedicels, number of flowers per umbellet, and number of umbellets in an umbel) were chosen as input variables in both artificial neural network and multiple linear regression models by simple correlation analysis. Using the SigmoidAxon transfer function and two hidden layers of an artificial neural network, the essential oil concentration of ajowan was accurately predicted with a root mean square error (RMSE) of 0.192%, a mean absolute error (MAE) of 0.112%, and a determination coefficient (R2) of 0.901.ANN outperformed MLR in terms of performance, with an RMSE of 0.262 and an R2 of 0.748. Based on stepwise regression and ANN analyses the most important features for the oil content of ajowan were the number of umbellets in an umbel and thenumber of flowers per umbellet and these qualities can be given as selection criteria for the essential oil content of ajowan.B. monnieri wild accessions were gathered from 81 sites in various eastern Indian areas (Odisha and West Bengal) to create an experimental dataset. According to the ANN results, a single hidden layer with 11 neurons, or the 13-11-1 structure of a multilayer perceptron (MLP) neural network, had the best ability to estimate the amount of bacoside A in a sample. With a coefficient of determination (R 2), a root mean square error (RMSE), and a mean absolute percentage error (MAPE) of 0.90, 0.16, and 7.76%, respectively, the constructed ANN model demonstrated a stronger predictive capacity for the training dataset.Additionally, the findings of the sensitivity analysis revealed that nitrogen concentrations and altitude had the greatest effects on the content of bacoside A. When evaluated at a new location, the ANN model showed a prediction accuracy of 93.60% for the presence of bacoside A. According to the study’s findings, bacoside A content in B.monnieri (L.) at a certain area can be predicted and optimized using an ANN model.
The advantage of ANN over other statistical modeling techniques is that it does not require the inference of a prior data structure and may detect nonlinear correlations and complicated interactions, exposing previously unknown linkages between input parameters [45]. The performance of the ANN model was assessed using root mean square error (RMSE), coefficient of determination (R2), mean absolute error (MAE), symmetric mean absolute percentage error (MAPE) & NSE coefficient. Root means square error (RMSE) was used as a performance measure to select the best model. Though the data set has multicollinearity among variables, the RMSE can be used as a performance measure [46]. The developed model performed well when evaluated with the above-mentioned performance measures. The linearity test & slope intercept test were also added to provide the model’s credibility. A similar study with a high predictive analysis of the ANN model was previously published, with the error values of the ANN model being lower [47–49].
ANNs do, however, have several important disadvantages. The usage of pricey graphics processing units (GPUs) that provide parallel processing is typically required for ANNs. Sharing an ANN after training is challenging. ANNs tend to overfit the training set. Because the size and structure of an ANN are mostly determined by trial and error, overfitting is likely due to this factor among other others.Experience and trial and error are frequently used to determine the size and structure of ANNs. Convergence on a prediction or solution is not guaranteed by ANNs. An ANN can accurately approximate a target function by selecting the appropriate parameters, or "hyperparameters" in ANN parlance. Achievable solutions like this don’t always exist, though. ANNs require a considerable amount of training time.
The correlation strengths among variables are also represented by partial dependence plots. Most of the correlation among variables is weak to moderate. Minimum average temperature & minimum average relative humidity have a strong correlation. In such cases, ANN models provide promising results and are free from any biases due to multicollinearity [50]. So, PDPs generated through artificial neural network models are suitable to study the change in response to predictors.
The coefficients in a generalized linear model are partially equivalent to those in an ANN, the weights that connect neurons. The weights’ cumulative influence on model predictions reflects the relative importance of predictors in their relationships with the outcome variable. In an ANN,there are numerous weights connecting one predictor to the outcome. An ANN’s high number of adjustable weights makes it incredibly flexible in modeling nonlinear phenomena, but it also makes interpretation difficult. The relative importance of a predictor, according to Garson, may be assessed by examining the model weights [51,52]. There are no linkages between each predictor of interest and the outcome. When all weights about a predictor are combined and scaled, a single value ranging from 0 to 1 is generated that shows relative predictor relevance. The Neural Networking Tools (version 1.5.1) package in R can be used to calculate relative relevance [31].
Friedman, 2001 proposed partial dependence plots (PDP) to show how one or two variables affect the model’s predictions. We also propose a new display that, like a scatter plot matrix, shows all pairwise partial dependence plots in a matrix-style layout with a univariate partial dependence plot on the diagonal. With this presentation, the analyst can observe at a glance how important pairs of variables affect the fit. Interpretation is aided once more by careful sequencing of the variables [53].
This is the first study to look at the impact of environmental and soil nutritional parameters on propolis oil content in 10 different Odisha agro-climatic areas. A combination of two or more factors has a greater impact on propolis oil content than a single element, according to the study. The current study found that environmental variables, such as soil macronutrients, can influence propolis oil content. The current study discovered that adjusting height, maximum average temperature, and phosphorus content in the soil can maximize propolis oil content in the ANN model. Differences in Propolis oil content are influenced by the parameters mentioned above, which need to be investigated further. The ANN model was used to predict the amount of propolis oil in a new location.
Conclusions
Artificial Neural Network (ANN) models were constructed in propolis for the prediction and optimization of oil yield in this study. The multilayer feed-forward neural network model was determined to be the most efficient for oil yield optimization, with a coefficient of determination value of 0.93. The oil content of propolis can also be enhanced using the ANN model by adjusting the model’s input parameters (altitude, phosphorus, and maximum average temperature). The ANN model created in this work could potentially be utilized to forecast propolis oil yield at a new location. Experimental findings backed up the forecast. The created ANN model could be useful in the manufacture of high-oil-content propolis and hence has commercial value.
Supporting information
S1 Table. Climatic data for propolis from different Agro-climatic regions of Odisha.
https://doi.org/10.1371/journal.pone.0283766.s001
(DOCX)
S2 Table. Physicochemical properties of soil samples collected from different Agro-climatic regions of Odisha.
https://doi.org/10.1371/journal.pone.0283766.s002
(DOCX)
Acknowledgments
The authors are grateful to Prof (Dr). Sudam Chandra Si, Dean, and Prof (Dr). Manoj Ranjan Nayak, President, Centre of Biotechnology,Siksha O Anusandhan University, for providing all facilities.
References
- 1. Kartal M, Yıldız S, Kaya S, Kurucu S, Topçu G. Antimicrobial activity of propolis samples from two different regions of Anatolia. J. Ethnopharmacol. 2003 May 1;86(1):69–73. pmid:12686444
- 2. Wang BJ, Lien YH, Yu ZR. Supercritical fluid extractive fractionation–study of the antioxidant activities of propolis. Food. Chem. 2004 Jun 1;86(2):237–43.
- 3. Bankova V, Christov R, Popov S, Pureb O, Bocari G. Volatile constituents of propolis. Z. Natur. Forsch. C. 1994 Feb 1;49(1–2):6–10.
- 4. Kujumgiev A, Tsvetkova I, Serkedjieva Y, Bankova V, Christov R, Popov S. Antibacterial, antifungal and antiviral activity of propolis of different geographic origin. J. Ethnopharmacol. 1999 Mar 1;64(3):235–40. pmid:10363838
- 5. Ioshida MD, Young MC, Lago JH. Chemical composition and antifungal activity of essential oil from Brazilian propolis. J. Essent. Oil-Bear. Plants. 2010 Jan 1;13(5):633–7.
- 6. Naik DG, Vaidya HS, Namjoshi TP. Essential oil of Indian propolis: chemical composition and repellency against the honeybee Apisflorea. Chem. Biodivers. 2013 Apr;10(4):649–57. pmid:23576351
- 7. Hames-Kocabas EE, Demirci B, Uzel A, Demirci F. Volatile composition of Anatolian propolis by headspace-solid-phase microextraction (HS-SPME), antimicrobial activity against food contaminants and antioxidant activity. J. Med. Plants. Res. 2013; 7:2140–9.
- 8. Li YJ, Xuan HZ, Shou QY, Zhan ZG, Lu X, Hu FL. Therapeutic effects of propolis essential oil on anxiety of restraint-stressed mice. Hum. Exp. Toxicol. 2012 Feb;31(2):157–65. pmid:21672965
- 9. Bankova V, Popova M, Bogdanov S, Sabatini AG. Chemical composition of European propolis: expected and unexpected results. Z. Natur. Forsch. C. J. Biosci. 2002 Jun 1;57(5–6):530–3. pmid:12132697
- 10. Sforcin JM, Fernandes A Jr, Lopes CA, Bankova V, Funari SR. Seasonal effect on Brazilian propolis antibacterial activity. J. Ethnopharmacol. 2000 Nov 1;73(1–2):243–9. pmid:11025162
- 11. Radusiene J, Stanius Z, Cirak C, Odabas MS. Quantitative effects of temperature and light intensity on the accumulation of bioactive compounds in St. John’s worth. InXXVIII International Horticultural Congress on Science and Horticulture for People (IHC2010): A New Look at Medicinal and 925. 2010 Aug 22;135–140.
- 12. Saffariha M, Jahani A, Jahani R. A comparison of artificial intelligence techniques for predicting hyperforin content in Hypericum perforatum L. in different ecological habitats. Plant. Direct. 2021 Nov;5(11): e363.
- 13. Gopal PM, Bhargavi R. A novel approach for efficient crop yield prediction. Comput. Electron. Agric. 2019 Oct 1;165: 104968.
- 14.
TorkashvandMoradabadi M. Sensitivity Analysis and Reexamination of the Techniques for Evaluating Adult Death Registration (Doctoral dissertation, Ph. D. Thesis, Tehran University, Iran.
- 15. Besalatpour AA, Ayoubi S, Hajabbasi MA, Mosaddeghi M, Schulin R. Estimating wet soil aggregate stability from easily available properties in a highly mountainous watershed. Catena. 2013 Dec 1; 111:72–9.
- 16. Alam MA, Naik PK. Impact of soil nutrients and environmental factors on podophyllotoxin content among 28 Podophyllum hexandrum populations of the northwestern Himalayan region using linear and nonlinear approaches. Commun. Soil. Sci. Plant. Anal. 2009 Sep 1;40(15–16):2485–504.
- 17. Padhiari BM, Ray A, Champati BB, Jena S, Sahoo A, Kuanar A, et al. Artificial neural network (ANN) model for prediction and optimization of bacoside A content in Bacopa monnieri: A statistical approach and experimental validation. Plant. Biosyst. 2022 Mar 3:1–2.
- 18. Jackson ML. Soil chemical analysis, pentice hall of India Pvt. Ltd., New Delhi, India. 1973; 498:151–4.
- 19. Subbaiah BV. A rapid procedure for estimation of available nitrogen in soil. Curr. Sci. 1956; 25:259–60.
- 20.
Team RC. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.org/. 2013 Jun.
- 21.
Leisch F, Dimitriadou E. Machine learning benchmark problems. R package version 2.1–3.
- 22.
Revelle WR. psych: Procedures for personality and psychological research.
- 23. Rocabruno-Valdés CI, González-Rodriguez JG, Díaz-Blanco Y, Juantorena AU, Muñoz-Ledo JA, El-Hamzaoui Y, et al. Corrosion rate prediction for metals in biodiesel using artificial neural networks. Renew. Energ. 2019 Sep 1;140: 592–601.
- 24. Maiseli BJ. Optimum design of chamfer masks using symmetric mean absolute percentage error. Eurasip J. Image Video Process. 2019 Dec;2019(1):1–5.
- 25. Jeya S, Sankari L. Air pollution prediction by deep learning model. In2020 4th International Conference on Intelligent Computing and Control Systems (ICICCS) 2020 May 13 (pp. 736–741). IEEE.
- 26. Papailiou I, Spyropoulos F, Trichakis I, and Karatzas G P. Artificial Neural Networks and Multiple Linear Regression for Filling in Missing Daily Rainfall Data, WATER. 2022; ISSN 2073-4441, 14 (18), 2892.
- 27. Greenwell BM. pdp: An R package for constructing partial dependence plots. R J. 2017 Jun 1;9(1):421.
- 28. Müller HG. Weighted local regression and kernel methods for nonparametric curve fitting. J. Am. Stat. Assoc. 1987 Mar 1;82(397):231–8.
- 29. Cleveland WS, Devlin SJ. Locally weighted regression: an approach to regression analysis by local fitting. J. Am. Stat. Assoc. 1988 Sep 1;83(403):596–610.
- 30.
Garson DG. Interpreting neural network connection weights.
- 31. Beck MW. NeuralNetTools: visualization and analysis tools for neural networks. J. Stat. Softw. 2018;85(11):1. pmid:30505247
- 32. Abbasi B. A neural network applied to estimate process capability of non-normal processes. Expert Syst. Appl. 2009; 36 (2). 3093–3100.
- 33.
S Chowdhury, Y Lin, B Liaw, and L Kerby. "Evaluation of Tree-Based Regression over Multiple Linear Regression for Non-normally Distributed Data in Battery Performance," 2022 International Conference on Intelligent Data Science Technologies and Applications (IDSTA), 2022, 17–25.
- 34. Schober P, Boer C, Schwarte LA. Correlation coefficients: appropriate use and interpretation. Anesth. Analg. 2018 May 1;126(5):1763–8. pmid:29481436
- 35. Nicodemus Kristin K., Malley James D. Predictor correlation impacts machine learning algorithms: implications for genomic studies, Bioinform, 2009; 25 (15). 1884–1890.
- 36. Fullana M, Trabelsi F, Recasens F. Use of neural net computing for statistical and kinetic modelling and simulation of supercritical fluid extractors. Chem. Eng. sci. 2000 Jan 1;55(1):79–95.
- 37. Kamali MJ, Mousavi M. Analytic, neural network, and hybrid modeling of supercritical extraction of α-pinene. J. Supercrit. Fluids. 2008 Dec 1;47(2):168–73.
- 38. Shokri A, Hatami T, Khamforoush M. Near critical carbon dioxide extraction of Anise (Pimpinella Anisum L.) seed: Mathematical and artificial neural network modeling. J. Supercrit. Fluids. 2011 Aug 1;58(1):49–57.
- 39. Eslamimanesh A, Gharagheizi F, Mohammadi AH, Richon D. Artificial neural network modeling of solubility of supercritical carbon dioxide in 24 commonly used ionic liquids. Chem. Eng. sci. 2011 Jul 1;66(13):3039–44.
- 40. Lashkarbolooki M, Vaferi B, Rahimpour MR. Comparison the capability of artificial neural network (ANN) and EOS for prediction of solid solubilities in supercritical carbon dioxide. Fluid. Ph. Equilibria. 2011 Sep 25;308(1–2):35–43.
- 41. Khajeh M, Moghaddam MG, Shakeri M. Application of artificial neural network in predicting the extraction yield of essential oils of Diplotaeniacachrydifolia by supercritical fluid extraction. J. Supercrit. Fluids. 2012 Sep 1; 69:91–6.
- 42. Ghoreishi SM, Heidari E. Extraction of Epigallocatechin-3-gallate from green tea via supercritical fluid technology: Neural network modeling and response surface optimization. J. Supercrit. Fluids. 2013 Feb 1; 74:128–36.
- 43. Lashkarbolooki M, Shafipour ZS, Hezave AZ. Trainable cascade-forward back-propagation network modeling of spearmint oil extraction in a packed bed using SC-CO2. J. Supercrit. Fluids. 2013 Jan 1; 73:108–15.
- 44. Pilkington JL, Preston C, Gomes RL. Comparison of response surface methodology (RSM) and artificial neural networks (ANN) towards efficient extraction of artemisinin from Artemisia annua. Inds. Crop. and. prod. 2014 Jul 1; 58:15–24.
- 45. Alvarez R. Predicting average regional yield and production of wheat in the Argentine Pampas by an artificial neural network approach. Eur. J. Agron. 2009 Feb 1;30(2):70–7.
- 46. Obite CP, Olewuezi NP, Ugwuanyim GU, Bartholomew DC. Multicollinearity effect in regression analysis: A feed forward artificial neural network approach. J. Probab. Stat. 2020 Jan;6(1):22–33.
- 47. Sodeifian G, Sajadian SA, Ardestani NS. Optimization of essential oil extraction from LaunaeaacanthodesBoiss: Utilization of supercritical carbon dioxide and cosolvent. The J. Supercrit. Fluids. 2016 Oct 1; 116:46–56.
- 48. Akbar A, Kuanar A, Joshi RK, Sandeep IS, Mohanty S, Naik PK, et al. Development of prediction model and experimental validation in predicting the curcumin content of turmeric (Curcuma longa L.). Front. Plant. Sci. 2016 Oct 6; 7:1507. pmid:27766103
- 49. Akbar A, Kuanar A, Patnaik J, Mishra A, Nayak S. Application of artificial neural network modeling for optimization and prediction of essential oil yield in turmeric (Curcuma longa L.). Comput. Electron. Agric. 2018 May 1; 148:160–78.
- 50. Veaux RD, Ungar LH. Multicollinearity: A tale of two nonparametric regressions. InSelecting models from data 1994 (pp. 393–402). Springer, New York, NY.
- 51.
Garson DG. Interpreting neural network connection weights.
- 52. Goh AT. Back-propagation neural networks for modeling complex systems. Eng. Appl. Artif. Intell.1995 Jan 1;9(3):143–51.
- 53. Friedman JH. Greedy function approximation: a gradient boosting machine. Ann. Stat. 2001 Oct 1:1189–232.