A deep learning based ensemble approach for protein allergen classification

Introduction

The prevalence of allergic reactions is a serious public health problem that may cause diseases like high fever, rhinitis, asthma, dermatitis etc., and affects a sizeable fraction of the world’s population (Pomés et al., 2018). Currently, the methods that can be used to cure allergies are not completely understood, and the only strategy that is known for preventing allergies is to abstain from substances that contain allergens (Sena-Torralba et al., 2020). However, the inaccurate detection of allergens may result in excessive dietary restrictions, which can then lead to nutritional issues (Shin et al., 2022). Hence it is clear that an efficient mechanism, which can amalgamate multiple parameters together for accurate allergen detection, is necessary for allergy control. The computational methods combined with bioinformatics have the ability to analyze multiple characteristics together for better allergen detection accuracy (Jeevanandam et al., 2022).

Protein allergens are a key contributor to the development of allergic reactions; hence, locating and describing these allergens is essential to the research and development of efficient diagnostic and therapeutic methods (Dimitrov et al., 2014a). Bioinformatics techniques may be used to conduct an analysis of protein allergen sequences in order to locate potentially allergenic areas because they contain carefully curated information on the sequences and structures of protein allergens (Yu et al., 2023). However, traditional techniques of allergen prediction take a lot of time and frequently depend on methods of experimentation that require a lot of manual resources (Zhou et al., 2019).

Machine learning and deep learning algorithms have only very recently emerged as potentially useful techniques for allergen prediction and classification (Westerhout et al., 2019; Bhardwaj et al., 2023). These algorithms are meant to analyse enormous volumes of data in order to recognise patterns that human analysts, depending on their level of experience, can miss. These intelligent algorithms have been utilised in a number of studies for the purpose of allergen prediction and categorization.

Individuals who are afflicted by allergies may benefit from a better diagnosis and treatment if researchers and clinicians include intelligent mechanisms in their work. So, in this article, we have described various essential factors/properties of protein sequences that can be used for the analysis and detection of different types of allergens.

The researchers have worked on various machine learning techniques for allergy detection and control. However, this article proposes a novel combination of ensemble learning, machine learning, and deep learning approaches for enhancing overall performance. Therefore, in this research, a deep learning based ensemble of three models: extra trees, DBN, and catboost has been proposed for the detection of protein allergens. The performance has been analyzed using different parameters like accuracy, F1-score, False Alarm Rate, Specificity, and Matthews correlation coefficient (MCC) etc.

The contribution of this article is summarised as:

• The article describes in detail the role and responsibilities of various entities related to allergy detection and control and also highlights the importance of protein allergen detection mechanisms.

• The features of the studied dataset, along with the relevance of each feature for allergen detection, have also been discussed in the article.

• A deep learning-based ensemble approach has been proposed, and its performance comparison with various state-of-the-art computational mechanisms for protein allergen detection has been presented.

The rest of the article is organized as: “Allergy Detection and Related Concepts” describes various concepts and entities that are essential for protein allergen classification; The recent state-of-the-art literature techniques pertaining to the field of allergy detection are discussed in “Related Work”; The analysis and working of the proposed mechanism is described in “Proposed Mechanism for Allergen Detection”; “Results and Discussion” presents the performance evaluation of various machine learning, deep learning and proposed mechanism for the detection of allergenic and non-allergenic protein sequences; Finally the article is concluded in “Conclusion”.

Allergy detection and related concepts

This section describes in detail various entities and concepts related to allergies and mechanisms that can affect the detection of different types of protein allergen sequences.

Related work

Westerhout et al. (2019) talks about how hard it is to figure out how allergenic new proteins might be, and how we need to find new, safe sources of protein to make food in the future. The current rules for genetically edited proteins are based on a weight-of-evidence method, which looks at how similar the sequence is to known allergens, how resistant the protein is to being broken down by pepsin, and how it is linked to sugars. But other physical and biochemical features of proteins are not being looked at right now. In this study, the Random Forest method was used to make an in silico model that predicts the allergenic potential of a protein based on its physicochemical and biological features. The ProtParam tools and the PSIPred Protein Sequence Analysis programme were used to figure out 29 factors from the protein sequence.

Fernandez et al. (2021) talked about how important it is to make food allergy risk assessment systems that rank allergens based on how likely they are to cause an allergic reaction. This score will help the risk assessment process get more accurate knowledge about how allergens affect health. The study says that this method will improve on the present, too simple way of classifying proteins as allergens or not based on whether or not they are in an allergen database. Creating tailored biology tools based on better algorithms will make risk assessment methods more efficient and give the public more reliable information. The review comes to the conclusion that an international agreement on a more robust approach to allergen-sequence database curation is needed to improve the quality of allergenicity risk assessment of foods made with biotechnology and new foods, which is urgently needed in an age of climate change and the move towards more sustainable food systems.

The physiochemical and structural characteristics of allergen proteins originating from plants and animals were examined and compared in Behbahani, Rabiei & Mohabatkar (2020) using in-silico analysis and bioinformatics methods. The study attained an accuracy of 88.24% by analysing the attributes of the allergens and applying the PseAAC concept together with deep learning algorithms for categorization. In terms of their extinction coefficient and secondary structure, plant allergen proteins showed a more randomly coiled shape than animal allergen proteins, according to the investigation. The research shows the promise of bioinformatics-based methods for comparing allergens and elucidating their characteristics.

Omurca et al. (2019) designed an intelligent diagnostic assistant for predicting the type of an allergic disease across Turkey automatically by using well-known machine learning algorithms such as Decision Tree, Logistic Regression, Support Vector Machines (SVM), K nearest neighbour (kNN), and ensemble classifiers. This was accomplished by using Turkey as a case study. An allergic illnesses dataset, which originates from the Kocaeli University Research and Application Hospital, was utilised in the studies that were conducted. As a consequence of this, the highest accuracy rate of 77% was attained with majority vote when recognising 18 distinct allergy diagnoses.

In the context of sustainable food systems and goods produced from biotechnology, the scientific opinion in EFSA Panel on Genetically Modified Organisms (GMO) et al. (2022) discusses the urgent requirement for the advancement of allergenicity evaluation and protein safety. The ideas and recommendations that now underpin allergenicity risk assessment methodologies may not be in line with contemporary scientific developments. Significant information gaps still exist despite the European Food Safety Authority’s (EFSA) and EU-funded research programmes’ attempts to develop the area. The goals of this Scientific Opinion include identifying knowledge gaps in allergenicity prediction, pinpointing specific research requirements for enhancing allergenicity risk assessment, figuring out how recent discoveries and technological advancements can improve the current risk assessment methodologies, and prioritising research funding. The ‘weight-of-evidence’ method to evaluating allergenicity is still applicable, but the specific evidence needed will depend on the kind of biotech food under consideration. Improved gene and protein information standardisation, updated in silico tools, improved in vitro testing integration, and clearer guidance on the overall weight-of-evidence strategy for protein safety are among the major modernization areas. Future goods, such as those made using novel genomic methods and synthetic biology, will require the allergenicity risk assessment to change in order to handle their complexity. A road map that addresses important issues for risk assessors and management is required to clarify allergenicity safety objectives and risk assessment requirements.

The study in Singh et al. (2021) focuses on the most significant discoveries made in the field of food allergy research, which includes both computational biology and bioinformatics, as well as experimental investigations. It examines the present status of research prospects and future perspectives in the field of food allergy and offers an account of the tools and databases used for identifying and analysing food allergens. In addition, it provides an overview of the methods and databases used for identifying and analysing food allergens. The study highlights how important it is to identify the allergens that are present in various food sources in order to prevent unwanted effects and treat allergy illnesses that are caused by the intake of certain foods.

Nedyalkova et al. (2023) gives a detailed research endeavour that sought to create a proficient computational framework for forecasting the allergenic properties of proteins through the utilisation of diverse descriptors and machine learning algorithms. The research employed a dataset comprising proteins that are allergenic and non-allergenic, and conducted feature selection to determine the most pertinent descriptors. The performance of classifiers was assessed through cross-validation and diverse performance metrics. The findings indicate that KNN exhibited superior performance compared to alternative classifiers in distinguishing between allergenic and non-allergenic sequences. Conversely, SVM demonstrated better performance when utilising a limited number of descriptors. The study found that the chosen descriptors, which comprised of amino acid composition, evolutionary, and AAindex-based features, had a notable impact on the classification performance. The study showcased the potential of employing computational methodologies for the prediction of protein allergenicity. It also emphasised the significance of feature selection in enhancing the classification model’s efficacy.

The study in Yu et al. (2023) utilised Quantitative Structure Activity Relationship (QSAR) models to predict the binding ability of protein epitopes to IgE, the antibody responsible for allergic reactions. Four algorithms and selected chemical descriptors were used to establish models for predicting the binding capabilities of epitopes to IgE. The study validated the performance of the models through an Enzyme-Linked Immunosorbent Assay (ELISA). The results showed that the models were able to predict the allergic reactions of food protein epitopes effectively. According to the study, the amino acid sequence 116–130 of $\beta$-lactoglobulin ( $\beta$-LG) was identified as a new IgE-binding epitope through both in vitro experiments and the QSAR model.

Wang et al. (2021) proposed a mechanism for predicting the allergenicity of food proteins using deep learning models, specifically the pre-training BERT model, and novel ensemble learning models represented by LightGBM and XGBoost.

The study in Yang et al. (2020) concentrates on predicting human-virus protein-protein interactions (PPIs) using a random forest classifier based on doc2vec embeddings. This computational framework outperformed combinations of other widely-used machine learning algorithms and sequence encoding schemes, according to the study. The use of feature embedding for protein representation enabled the acquisition of additional context information from protein sequences, thereby enhancing prediction performance. The authors anticipate that their findings can be used to identify potential interactions between human and viral proteins and to direct experimental efforts to identify proteins implicated in human-virus interactions and their associated functional functions. The authors hypothesise that future advancements in deep learning architectures, protein structural information, and host PPI network topology can enhance the prediction of human-virus PPIs.

Sharma et al. (2021) used molecular descriptors and machine learning methods to create prediction models for chemical compound allergenicity. The PaDEL programme was used to compute the molecular descriptors for the 403 allergenic and 1,074 non-allergic substances that were taken from the IEDB database. On a dataset that included 2D, 3D, and FP descriptors, the models were trained and put to the test. The highest performance came from the hybrid descriptor-based Random Forest-based model, which on the validation dataset had an AUC of 0.93 and a maximum accuracy of 83.39%. It was discovered that several chemical fingerprints, such as GraphFP1014 and PubChemFP129, were more prevalent in allergens. The study also identified pharmaceuticals that are known to produce allergy symptoms and projected the probable allergenicity of FDA-approved medications. ChAlPred, a web server created for the project, enables the prediction and creation of compounds having allergenic qualities.

In comparison to previous computational techniques, AllerCatPro 2.0 (Nguyen et al., 2022) is a computational tool that has been designed to predict the allergenicity potential of proteins in food and personal care items. The similarity between input proteins and datasets of trustworthy proteins linked with allergenicity, which have been carefully selected from several sources, is assessed using both amino acid sequences and projected 3D structures. A total of 4,979 protein allergens, 162 mild allergenic proteins, and 165 autoimmune allergens are all included in these databases. Along with new features, AllerCatPro 2.0 offers more thorough results for possible cross-reactivity, protein information (UniProt/NCBI), functionality (Pfam, InterPro, SUPFAM), clinical relevance with regard to IgE prevalence, and allergen information pertaining to the allergen that is most similar. With an 84.7% sensitivity and 68.9% specificity on challenging benchmark datasets, the tool has been evaluated on numerous examples of profilins, autoimmune illnesses, low allergenic proteins, extremely big proteins, and nucleotide input sequences. It has demonstrated improved accuracy over earlier iterations.

Akbar et al. (2017) proposed an ensemble learning approach for the detection of anticancer peptides using genetic algorithm. The author also gives a detailed mathematical representation of the proposed ensemble technique. The authors have also extended the work in Akbar et al. (2021, 2023) for identifying antitubercular peptides and Neuropeptides by combining various machine learning techniques like Support Vector Machine, Random Forest, K-nearest neighbor etc., using ensemble learning.

MacMath, Chen & Khoury (2023) have reviewed the use of AI techniques for the detection, prevention, and cure of allergic reactions and other related tasks. The article also describes various challenges involved in the implementation AI techniques in such fields.

Zhang et al. (2023) presented a mechanism for screening of antihypertensive peptides using deep learning model. The effective results of the proposed deep learning model suggested that similar deep learning models can be used for related tasks involving peptides.

Proposed mechanism for allergen detection

In this section we describe the details of proposed model for the detection of protein allergens. The literature suggests that by combining the outcome of different machine learning models together using ensemble learning techniques can enhance the overall accuracy of prediction. This work uses this concept and hence proposes an ensemble learning model to detect protein allergens in a given dataset. The system model for the proposed mechanism is shown in Fig. 1.

In the initial phase, we processed each protein sequence collected from different sources. The allergen protein sequences were accessed from AllergenOnline (Goodman et al., 2016), and non-allergen protein sequences were collected from National Center for Biotechnology Information (Wheeler et al., 2003), UniProt (Bairoch et al., 2005) and SWISS-PROT (Bairoch & Apweiler, 2000). These protein sequences were then processed for extracting features, that can be used to train different machine learning and deep learning models. In the Fig. 1 we have mentioned few protein sequences in the “Protein Sequence” block. The features were extracted using Peptide package available on CRAN. The physical, chemical and other properties, as described in Table 1, were extracted from each protein sequence.

The performance of any machine learning algorithm depends on the values contained in the dataset, and the prediction results are effective if all the feature values are numeric and lie in a particular range. In order to achieve this, we have replaced any missing value with the mean of the corresponding feature vector and encoded non-numeric values to numeric values using the labeled encoding technique.

The updated dataset was then passed to the feature scaling phase, where we used a standard scalar to transform all the values of feature vectors. If $x$ is the feature vector and $\overline{x}$ and $\sigma$ are the mean and standard deviation of the feature vector, then the transformed feature vector $\hat{x}$ is calculated using standard scalar as shown in Eq. 1.

(1) $$\hat{x}=\frac{(x-\overline{x})}{\sigma }$$

The completion of feature extraction, pre-processing and feature scaling results into a pre-processed dataset containing 128 features and 4,854 entries (among which 2,427 represent allergens and the rest represent non-allergens), which can then be used for training and testing the machine learning models.

We have selected three models for constructing the proposed ensemble learning model for detecting protein allergens. The selection was based on a thorough performance analysis of different machine learning and deep learning models on the dataset.

The proposed ensemble learning model contains a combination of Extra Tree, Deep Belief Network (DBN) and CatBoost models, the details of which are given as:

• Extra Tree: Decision tree algorithms are the basis of the Extra Trees (Extremely Randomised Trees) classifier, another ensemble learning approach. Training numerous decision trees and aggregating their predictions results in a more accurate and robust categorization. In comparison to the standard Random Forest algorithm, the Extra Trees classifier introduces additional randomness, leading to more diverse trees and improved generalisation performance (Polikar, 2006).

  The Extra Trees classifier generates a collection of decision trees, each of which is constructed using a different subset of the training data and a different set of characteristics at each node. Due to the algorithm’s lack of optimisation of the splitting criterion, it is both quicker and more random than conventional decision tree approaches like Random Forest, from which it borrows its name. By adding some randomness to the tree-building process, we may generate a more varied set of trees, which helps lower the model’s variance and boosts its generalisation performance. Rather of relying on just one tree’s predictions, the classifier aggregates the results from all of the trees in the ensemble, typically by majority vote, to provide more reliable and accurate classifications.

• DBN: To quickly and effectively create hierarchical representations of data, a DBN blends unsupervised and supervised learning approaches. DBNs are constructed from a number of layered Restricted Boltzmann Machines (RBMs). They are taught greedily, layer after layer, using unsupervised pre-training to set their starting weights and then supervised learning to fine-tune them. DBNs have been successful in a number of contexts, including those involving image and speech recognition, NLP, and dimensionality reduction (Lopes et al., 2015).

• CatBoost: A gradient boosting technique designed specifically for categorical features is called CatBoost. In addition to providing effective management of big datasets containing categorical variables, it is optimised for high-performance prediction (Bentéjac, Csörgő & Martínez-Muñoz, 2021).

  A gradient boosting technique designed specifically for categorical features is called CatBoost. In addition to providing effective management of big datasets containing categorical variables, it is optimised for high-performance prediction. To lessen target leakage and overfitting, it uses ordered boosting, which rearranges the training examples using a random permutation and then calculates target statistics for each category using the instances that came before them in the permutation.

As discussed earlier in this section, we have selected three models based on the performance analysis of various machine learning and deep learning models, to create an ensemble model. The other machine learning models used in the study include: Logistic Regression (with maximum iterations of 300 on L2 regularization and Limited-memory BFGS optimizer), k-Nearest Neighbour (with $k=10$), Support Vector Machines (with three kernel variants: Linear, Polynomial of degree 3 and Radial Basis Function), Random Forest (with size of base estimator as 100 and criterion for node splitting as Gini index), Ada Boost, Gradient boosting and XG Boost (Zhou, 2021).

The deep learning models include: Multi-layer perceptron and DBN, the architecture of which is shown in Fig. 2. The number of neurons in the input layer of each deep learning model equals the number of available features. The first hidden layer has the same number of neurons as that of the input layer, but the second hidden layer has 50% of neurons in the first hidden layer (i.e., 64 neurons). The third hidden layer has 13 neurons that are responsible for pattern recognition for major feature descriptors which include hydrophobicity, kideraFactors, amino acid property scales, VHSE scales, Z scales, Cruciani molecular descriptors, WHIM molecular descriptors, BLOSUM matrix scores, Boman index, charge, molecular weight, peptide length and protein fingerprint.

Each of the models in the proposed ensemble learning approach receives feature values concurrently and predicts the class label (allergen or not). The prediction results of the base models are then combined together using majority voting, the output of which gives the final predicted class label for the input values. In majority voting, the class label that receives the most votes becomes the ensemble’s prediction for that instance. We have used majority voting because it lowers the likelihood of overfitting, boosts model stability, and increases prediction accuracy (Polikar, 2012).

Results and discussion

The performance analysis of the proposed model and other machine learning and deep learning models on the dataset was implemented using Python 3.10 with scikit-learn, PyTorch and other related packages on 1.6 GHz Intel Core i5 with 12 GB of RAM. The dataset was divided into training and testing set by random selection of feature values. The training and testing datasets contained 3,397 and 1,457 respectively, randomly selected entries.

Result analysis

In the study, we have also used deep learning models (MLP and DBN) for allergen classification. The loss occurred during training of these models with varying epoch is shown in Fig. 3. The figure shows that as the number of epochs increase, the training loss for both the models decrease and attain a minimum values of 5% (approx.). This reduced value of training loss indicates the effective training of deep learning models.

The test results of various machine learning, deep learning and proposed model are presented in Table 3 in the form of True Negative (TN), False Positive (FP), False Negative (FN), True Positive (TP), Training Time and Testing Time (in seconds).

Table 3:

Classification results of various learning models on the test dataset.

Model	TN	FP	FN	TP	Training time (sec.)	Testing time (sec.)
Logistic regression (LR)	553	158	148	598	0.5419	0.00204
k-NN	590	121	85	661	0.0027163	0.176971
Linear SVM (LS)	558	153	148	598	2.41266	0.23742
Polynomial SVM (PS)	609	102	176	570	1.2014	0.3266
SVM with RBF (RS)	588	123	109	637	0.81344	0.31443
Random forest (RF)	624	87	97	649	3.479488	0.07969
Adaboost (AB)	586	125	116	630	6.21499	0.089167
Gradient boosting (GB)	595	116	105	641	14.23084	0.004366
MLP	664	47	180	566	28.3818	0.0010564
DBN	622	100	84	651	48.268	0.00113
Extra tree (ET)	627	84	102	644	8	0.075486
XG boost (XB)	619	92	85	661	9	0.084128
Cat boost (CB)	622	89	83	663	56.2	0.094152
Proposed model (proposed)	634	77	81	665	56.416	0.12854

DOI: 10.7717/peerj-cs.1622/table-3

The lower values of False Positives and False Negatives for DBN, Extra Tree and CatBoost models, as shown in Table 3, resulted into the formation of proposed ensemble model, with an objective to further enhance the overall classification performance. The proposed model has reduced the number of FP and FN, as show in the table. The test results of each model are then used to compute the value of other performance parameters.

The Fig. 4 shows the value of allergen classification accuracy of different models studied on the test dataset. Accuracy offers a gauge of how effectively the model can generate accurate assumptions about fresh and unseen data. A high accuracy rating means the model can generalize effectively and make precise assumptions about new data. The figure shows that the proposed model achieves maximum accuracy when compared with other models. The false alarm rate of each learning model is shown in Fig. 5. The figure shows that the proposed model has least false alarm rate of 10.83% than the other studied models (except MLP).

The F1-score is a statistic that integrates recall and precision into one performance parameter for the model. It offers a fair assessment of the model’s capacity to recognize both positive and negative instances. When the cost of false positives and false negatives is not equal, this is especially crucial. The F1-score of various studied models is shown in Fig. 6. It is evident from the figure that the proposed model has an F1-score of 89%, which is maximum when compared with other studied models.

Specificity quantifies the percentage of true negatives that the model properly detects, i.e., it tells us how well the model is able to detect undesirable circumstances. Figure 7 shows this specificity value for each studied model. Here again the proposed model has maximum specificity value of 89.17% than the other studied models (except MLP). The MLP model has maximum specificity value due to the lower value of false alarm rate. Besides MLP, the proposed model has maximum value of specificity than the other studied models.

The significance of Matthews correlation coefficient (MCC) is that it does not depend on unbalanced datasets and gives a fair assessment of the model’s performance, accounting for both true positives and true negatives. The MCC runs from −1 to +1, with +1 denoting perfect agreement, 0 denoting predictions that are no better than random, and −1 denoting utter disagreement between the predicted and the actual values. The MCC value of each model on the test dataset is shown in Fig. 8. Here again the proposed model has maximum MCC value of 78.3%, which shows the effectiveness of proposed model for allergen classification.

Apart from the already described performance parameters, we have also plotted ROC and PR curves for various models on the test dataset. Since the number of models are more, that is why we have presented the ROC and PR curves in two parts. The ROC curves are shown in Figs. 9 and 10, and the PR curves are shown in Figs. 11 and 12. We have also calculated the Area Under each Curve (AUC) and is mentioned in each figure. Closer the value of AUC to 100%, more efficient is the classification model. After analysing the figures, it is clear that the proposed model has an area under the ROC curve value of 89.2% and area under the PR curve value of 92.1%, which is maximum when compared with the other studied models.

The combined analysis of all the presented results clearly reveals that the proposed model effectively detects allergen and non-allergen protein sequences. The proposed model has maximum value of detection accuracy, F1-score, MCC and area under the ROC and PR curves.

Performance comparison

To further specify the significance of the proposed model, we have compared its performance with AllerTOP (Dimitrov, Flower & Doytchinova, 2013), AlgPred (Saha & Raghava, 2006), AllergenFP (Dimitrov et al., 2014b), AllerTOPv2 (Dimitrov et al., 2014a), AllerCatPro (Maurer-Stroh et al., 2019) and AllerCatPro 2.0 (Nguyen et al., 2022). We have evaluated the performance of the proposed model on different datasets (independent from the one described in section 2.4) used in AllerTOP, AlgPred, AllergenFP, AllerTOPv2, AllerCatPro, and AllerCatPro 2.0 respectively. The performance comparison is shown in Fig. 13, where it is clear that the proposed model outperforms the specified state-of-the-art literature techniques for allergen classification. Figure 13 shows that the proposed model has an average allergen detection accuracy of 89.16% and MCC of 78.3%, which is highest when compared with the other state-of-the-art literature techniques. The comparative study also shows that the model is free from overfitting issues and is stable.

Conclusion

This article presented the analysis of various machine learning and deep learning techniques to propose an ensemble learning based approach for the identification of protein allergens. The article also highlighted various entities and concepts related to allergy control. In this article, we also described various protein sequence properties like physicochemical, molecular, antimicrobial and other characteristics to extract the feature vectors, giving rise to the dataset used for training and testing of intelligent techniques. The performance evaluation of various machine learning, deep learning techniques and proposed mechanism revealed that the proposed mechanism outperforms the other discussed techniques and achieved an allergen detection accuracy of 89.16%. The performance of proposed mechanism was also compared with the state-of-the-art literature techniques, which used the similar dataset as used in this study, and the results prove that the proposed mechanism outperforms the other techniques.

Even though the proposed model outperforms the other related techniques, the overall accuracy can be further enhanced. So, in the future, various approaches like increasing the size of dataset, use of different optimization and regularization techniques, etc., can be used which will further enhance the performance of the proposed mechanism. Furthermore, the diversity of the dataset can be increased in the future by adding data from different geographical regions for increasing the precision of the proposed model.

Supplemental Information