FlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data

FlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article FlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data Fabio Herrera-Rocha, Miguel Fernández-Niño, Jorge Duitama, Mónica P. Cala, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4768238/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 10 Dec, 2024 Read the published version in Journal of Cheminformatics → Version 1 posted 4 You are reading this latest preprint version Abstract Flavor is the main factor driving consumers acceptance of food products. However, tracking the biochemistry of flavor is a formidable challenge due to the complexity of food composition. Current methodologies for linking individual molecules to flavor in foods and beverages are expensive and time-consuming. Predictive models based on machine learning (ML) are emerging as an alternative to speed up this process. Nonetheless, the optimal approach to predict flavor features of molecules remains elusive. In this work we present FlavorMiner, an ML-based multilabel flavor predictor. FlavorMiner seamlessly integrates different combinations of algorithms and mathematical representations, augmented with class balance strategies to address the inherent class of the input dataset. Notably, Random Forest and K-Nearest Neighbors combined with Extended Connectivity Fingerprint and RDKit molecular descriptors consistently outperform other combinations in most cases. Resampling strategies surpass weight balance methods in mitigating bias associated with class imbalance. FlavorMiner exhibits remarkable accuracy, with an average ROC AUC score of 0.88. This algorithm was used to analyze cocoa metabolomics data, unveiling its profound potential to help extract valuable insights from intricate food metabolomics data. FlavorMiner can be used for flavor mining in any food product, drawing from a diverse training dataset that spans over 934 distinct food products. Flavor Chemistry Molecular Machine Learning Molecular Representation Deep Learning Cocoa Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 INTRODUCTION Flavor is defined as the combination between taste and odor, without distinction, of a substance or product[ 1 ]. This feature plays an essential role in defining consumer's acceptance of foods and beverages. The flavor properties of most processed food products are readily manipulable by formulating the proper ingredients to meet the desired sensory quality. Nonetheless, in the case of fermented products, matured foods, fruits, and vegetables, the flavor is purely determined by their chemical composition (including that of the matrix) [ 2 ]. The flavor molecules are produced in complex biological processes or chemical reactions and may be modified during processing, causing a pleasant or unpleasant sensation depending on their interaction with flavor receptors[ 2 , 3 ]. Hence, understanding the impact of these compounds on the flavor profile is essential for engineering and decision-making in food and beverage production. This knowledge is critical for discovering and developing new flavors, standardizing and optimizing processing conditions, and variety selection to produce plant-based foods with a better flavor, among other applications[ 2 , 4 , 5 ]. The FlavorDB database ( https://cosylab.iiitd.edu.in/flavordb/ ) is the most comprehensive collection of molecules with an experimentally validated flavor profile[ 6 , 7 ]. This database also contains information from other flavor databases and external sources such as PubChem. Although this database contains more than 25,500 compounds, only 2,254 metabolites (~ 9%) are linked to the 936 food products[ 6 , 7 ]. In contrast, FooDB ( https://foodb.ca/ ), the largest food metabolomics database, contains more than 24,000 compounds identified in foods[ 8 ]. The number of food-related molecules without flavor profiles will probably continue increasing in the future, boosted by the growing power of high-throughput metabolomics techniques and the complexity of testing the flavor profile of individual molecules. In addition, for most synthetic compounds no flavor profile is known. The current experimental methodologies to assess the flavor profile of individual molecules require the compound to be isolated or synthesized[ 9 ]. Alternatively, the flavor is inferred by correlating metabolomics data with sensory results or tested using trained panelists [ 3 , 10 – 12 ]. This procedure is considerably time-consuming and expensive, especially considering the complex composition of most food products. A recent alternative to streamline this process is the procedure called ultra-fast GC E-nose, which is an automated methodology to discriminate samples based on their odor fingerprint[ 13 , 14 ]. The main drawback of this approach is that it does not assign a flavor profile to single compounds. Similarly, Gas Chromatography-olfactometry (GC-O) combines the power of gas chromatography with the sensitivity of the human nose to assign a flavor profile to individual compounds[ 12 ]. However, only volatile compounds can be tested following this technique, also perception even of an identical compound can be different when ingested orally in a matrix. Similar to several other knowledge domains, different Machine Learning (ML) models have been developed to perform in-silico flavor prediction from molecular structures with available flavor profiles[ 3 ]. This trend has been prompted by advancements in ML algorithms and the availability of large-scale molecular data[ 15 – 17 ]. This approach enables an efficient screening of potential flavors to prioritize compounds for validation using traditional experimental methods, saving time and resources[ 2 , 3 , 18 ]. Nevertheless, ML models require diverse datasets to learn effectively. This variability is even more relevant for flavor prediction because molecules with similar structures can have completely different flavors or divergent molecular structures can have a similar flavor profile[ 3 , 18 ]. Obtaining high-quality flavor data with sufficient coverage of different molecular structures can be challenging. In some cases, data may be limited or biased, leading to potential inaccuracies in predictions. Most studies in this area used binary classifiers concentrated on predicting sweet and bitter flavors, mainly because of the availability of large datasets with labeled sweet and bitter molecules[ 16 , 19 – 22 ]. These developments were also fostered by the need to address consumer preferences and health concerns. With increasing interest in reduced sugar consumption, the food industry is demanding alternative sweeteners or taste enhancers mimicking the desired tastes while reducing sugar content. Similarly, predicting bitterness or off-flavors is crucial for avoiding unpleasant taste experiences. Another reason boosting bitterness prediction is the wide prevalence of this flavor note in natural bioactive compounds, serving then as a potential screening tool[ 22 – 24 ]. Additionally, some predictors for sourness and umami have been published, although with less success because the availability of labeled data is limited for these flavor notes[ 19 , 25 ]. The existing public flavor predictors cannot predict notes like floral, nutty, fruity, or off-flavor[ 22 , 25 , 26 ]. Some attempts have been undertaken in this area, but the performance is still poor and the code and data are not publicly available[ 3 , 27 – 29 ], leading to a lack of available multilabel tools for flavor prediction. These flavor notes are of capital importance in fermented food processing (including coffee, beer, wine, chocolate, bread and others)[ 2 ]. Also, some models based on Generative Artificial Intelligence have been trained to generate new molecules with flavors potentially interesting for the food industry (including the above-mentioned flavors), but they lack any classification capability[ 4 , 17 ]. A major challenge for predicting these flavor notes is the class imbalance, as the number of positive examples is significantly lower than the negatives[ 2 , 25 ]. Even though some class balancing methods are available (e.g., over-sampling and under-sampling)[ 30 , 31 ], it is unclear what approach works better for flavor prediction. The most widespread approaches for flavor prediction are based on Quantitative Structure-Activity Relationship (QSAR) models, which have been extensively used in drug discovery[ 2 , 3 , 18 , 32 ]. These models correlate the chemical structure and properties of molecules with their biological activity, including flavor. The main ML algorithms used have been Random Forest, Support Vector Machines, K-Nearest Neighbors, Deep Neural Networks[ 20 ], and more recently Graph Neural Networks[ 2 , 18 , 19 ]. These models use mathematical features that capture molecular properties, such as atomic composition, connectivity, and physicochemical properties[ 18 ]. Molecular descriptors followed by molecular fingerprints are the most widespread mathematical representation of molecular structure to train flavor predictors[ 2 , 18 , 20 ]. Recently, molecular graph representations were also implemented for flavor prediction[ 4 , 18 , 19 ]. However, to date, no study comparing the performance of these mathematical representations has been reported. In response to the identified gaps, the present study aimed to develop FlavorMiner, a flexible and retrainable flavor ML predictor for seven flavor categories critical for quality control within the food industry (i.e., floral, fruity, sour, sweet, bitter, off-flavor, and nutty). This research comprises an exploration of the performance of different mathematical representations of molecular structures and classification algorithms for flavor prediction. Demonstrating its prowess, FlavorMiner was applied successfully to cocoa metabolomics data, showcasing its ability to unlock valuable insights from food metabolomics data. It can be utilized for flavor analysis in various food products, thanks to a diverse training dataset encompassing over 934 different food products. RESULTS AND DISCUSSION In this work we introduce the FlavorMiner algorithm, which takes as input the Isomeric Smiles of a set of molecules and produces as output their flavor profile (Fig. 1 ). The first step is to query a database of 13,387 molecules with known flavor profiles. Only the set of molecules with no database match pass to the prediction step. Then, the respective mathematical representation of the molecules is generated. In the next step, this mathematical representation is fed to seven independent binary classifiers. The average prediction capability of these predictors is 0.88 (ROC AUC Score). Each classifier predicts one of the seven target flavor categories (bitter, floral, fruity, off-flavor, nutty, sour, and sweet). The results are provided in a table, including the predicted flavor profile for each compound and the source of the flavor profile (database match or prediction). The probability values are also provided, indicating the confidence level of each prediction. Finally, a radar chart showing the recurrence of the molecules with each target flavor is also generated. 2.1. Development of ML models for flavor prediction, including management of class imbalance biases To train the classifiers incorporated in FlavorMiner, a flavor molecule dataset was assembled, containing 13,387 compounds with experimentally validated flavor profiles. The positive examples (those with a specific flavor) represent on average 20% of the dataset, while the negative examples (those without a specific flavor) represent 80% (Supplementary Fig. 1). This is a sign of class imbalance, which is an important challenge in the development of ML models, as it can lead to bias towards the majority class[ 33 , 34 ]. Due to this class imbalance, all algorithms trained on the original data, except for those trained on sweet molecules, had poor recall (Fig. 2), which measures the ability to correctly identify positive examples[ 30 , 31 ]. On the other hand, the specificity, which measures the ability to correctly identify negative examples[ 30 , 31 ], was significantly high. This bias towards the majority class was consistently observed regardless of the target flavor, algorithm, or mathematical representation. Figure 2. Classification Metrics for Algorithms Trained with Original Descriptor Data on Test Set. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. (a) Random Forest trained with molecular descriptors. (b) Random Forest trained with extended connectivity fingerprint. (c) K-Nearest Neighbors trained with molecular descriptors. (d) K-Nearest Neighbors trained with extended connectivity fingerprint. (e) Convolutional Graph Neural Network trained with molecular graph. Most algorithms trained on the original data showed a specificity higher than 0.9 during the test. Nonetheless, these models had a recall lower than 0.5. This performance evidence a bias towards the majority class higher than 40% for most algorithms. The Convolutional Graph Neural Network trained with the original molecular graph had the lowest recall for most target flavors (close to zero) (Fig. 2c). This is likely because this algorithm is more complex (larger number of parameters) and hence it requires more data to be trained effectively[ 35 , 36 ]. On the other hand, the sweet category has a bias of less than 10% with Random Forest and K-Nearest Neighbors, trained either with RDKit descriptors or ECFP, which can be explained because it had the smallest class imbalance. The number of sweet positives is only 2% lower than the negative examples. Conversely, the sour category has a class imbalance of 97% and showed the highest bias towards the majority class (> 85%). This is a common issue in ML models dealing with imbalanced data[ 37 ]. Additionally, overfitting was observed in models using Random Forest and K-Nearest Neighbors after the first training iteration with the original data. This is likely due to the limited number of positive data, which can lead the model to exceedingly consider specific features from the negative training set that constrain the ability to adequately generalize and predict when presented with previously unseen data[ 30 , 38 , 39 ]. The difference between the train and test specificity was under 10% in most cases, but the recall showed a considerable drop between 20–90% from train to test for most algorithms trained with the original data (Supplementary Fig. 2). This is a clear sign of overfitting[ 30 ]. Only some models obtained with the Convolutional Graph Neural Network showed no sign of overfitting. This is only because the recall was near to zero both during training and testing. Similarly, there was a proportional relationship between class imbalance and overfitting. For example, the models for the sweet flavor (the class with no imbalance) showed the lowest overfitting, while the sour flavor (the class with a high imbalance and special problems associated with its perception, v.i.) had the highest overfitting percentage. SMOTE and Cluster Centroid sampling techniques were implemented to address the class imbalance. These strategies significantly reduced bias and overfitting. SMOTE, an oversampling technique previously used in flavor predictors[ 19 , 20 , 28 ], was applied to the minority class to increase the number of positive examples. This resulted in a bias of less than 20% for most algorithms (Supplementary Fig. 3). The overfitting level was also reduced to less than 30% for most algorithms (Supplementary Fig. 4). Under-sampling with Cluster Centroid [ 40 ] was also applied to reduce the number of negative examples (Supplementary Fig. 5). This resulted in an overfitting reduction to less than 30% for K-Nearest Neighbors models and less than 15% for Random Forest models (Supplementary Fig. 6). Most K-Nearest Neighbors models had a bias of less than 10%, while most Random Forest models had a bias of over 20%. Bias and overfitting were reduced due to a significant increase in the recall after applying the resampling strategies. Although the bias and overfitting were still slightly high, this represented a significant improvement compared to the performance with the original data. The recall of all the algorithms trained with resampled data was over 50%, while the specificity of most of these models remained above 70%. Multiple studies have shown that both oversampling and undersampling can be used to correct the problems caused by class imbalance in machine learning approaches[ 33 , 34 , 40 ]. In the context of flavor prediction, several studies have investigated the effect of SMOTE oversampling[ 19 , 20 , 28 ]. These studies have focused mainly on sweet, bitter, and sour flavors, and have obtained results similar to those of the present work[ 19 , 25 , 38 ]. On the other hand, using a balancing transformer on the molecular graph to train a convolutional graph neural network significantly improved the recall, but also significantly reduced the specificity. The recall for classes with more class imbalance improved by 73–99%, but the specificity dropped by a similar proportion (Supplementary Fig. 7). Additionally, the recall for classes such as sweet and bitter decreased. Consequently, the bias and overfitting increased for all models trained with the balanced molecular graph compared to the original data. The bias was higher than 50% for most target flavors and was as high as 90% for fruity, off-flavor, nutty, and sour flavors. This indicates that the balancing transformer had a significant negative effect on the specificity of the models. The overfitting for bitterness and sweetness predictions increased with the balanced data. For fruity, off-flavor, nutty, and sour, the recall decreased to negative values by more than 20%. A negative recall change value indicates underfitting, which occurs when the model does not learn a strong enough pattern from the training data[ 30 , 31 ]. This can be solved by performing a more intense hyperparameter optimization, but this may come at a considerable computational cost compared to Random Forest and K-Nearest Neighbors algorithms. The balancing transformer and resampling techniques (SMOTE and cluster centroid) differ in how they address class imbalance. The balancing transformer focuses on the weights of positive and negative examples in the neural network, while resampling techniques focus on the feature space[ 30 , 40 , 41 ]. The balancing transformer does not change the input data or the number of examples in each class[ 30 ]. The poor results obtained with this strategy demonstrate that this is insufficient to solve the severe class imbalance of the input data. Resampling techniques, on the other hand, change the input data by creating new synthetic examples in the minority class (SMOTE) or by removing examples from the majority class and replacing them with cluster centroids[ 40 ]. Considering the significant improvement in the performance of the algorithms trained with resampling techniques, this seems to be the best approach to balance the flavor compound database. Unfortunately, it is challenging to implement resampling strategies on molecular graphs, and only possible with molecular descriptors and fingerprints. This is because clustering molecular graphs without affecting their structure and losing valuable information is nearly impossible. Also, in flavor studies, minor changes in structure (graphs) can cause severe changes in perception. Thus, synthetic filling can cause more rather than fewer problems. Although other balance methods are available for graph data, their usefulness with molecular graphs remains to be evaluated[ 41 ]. 2.2. FlavorMiner combines the best ML models for prediction of different flavor classes. Random Forest outperformed the K-Nearest Neighbors algorithm for most target flavor notes, except sour (see below for discussion). Random Forest trained with ECFP oversampled with SMOTE performed best for bitter, fruity, sweet, and off-flavor notes. Random Forest trained with RDKit descriptors performed best for floral and nutty notes. K-Nearest Neighbors trained with ECFP oversampled with SMOTE performed best for sour notes. In general, K-Nearest Neighbors had similar recall to Random Forest with the same input datasets, but slightly lower specificity. Also, algorithms trained with data resampled with the cluster centroid algorithm had slightly better recall, but a higher drop in specificity compared to datasets resampled with SMOTE. These results are consistent with previous studies, which found that Random Forest outperforms other algorithms for predicting sweet and bitter flavors[ 3 , 22 , 26 ]. A correlation was observed between the amount of positive data available and the performance of the classifiers. Sweet, the class with the highest number of positive instances, had the best overall performance, with an ROC AUC score of 0.97. Sour, the class with the lowest number of positive instances, had the lowest performance, with a ROC AUC score of 0.78. These results suggest a superior performance of algorithms trained with resampled datasets compared to those trained with the original data. The performance of the seven final predictors selected for the FlavorMiner backbone is shown in Fig. 3. The average ROC score, specificity, and recall of these classifiers were 0.88, 0.82, and 0.77, respectively. The performance of FlavorMiner for bitter and sweet prediction was comparable to that of existing predictors[ 20 , 22 , 26 ]. For fruity and floral prediction, FlavorMiner achieved recalls of 0.71 and 0.76, respectively, representing an improvement of over 50% compared to previous studies[ 18 , 42 ]. FlavorMiner is the first model to predict nutty and off-flavor notes from molecular structures. For sour prediction, FlavorMiner was outperformed by a previously published tool[ 25 ] by about 15%. However, the dataset, composition of positive and negative examples, and code used in this study are not publicly available, making it difficult to assess the reasons for the observed difference. Figure 3. Performance of the Optimized Classifiers for Target Flavor Notes in FlavorMiner. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. (a) Classification metrics obtained during training using 5-fold cross-validation. (b) Classification metrics obtained using the test set. Random Forest was used for bitter, fruity, sweet, off-flavor, floral and nutty. K-Nearest Neighbors for sour notes. Variable Importance Plots (VIP) scores[ 31 ] revealed the most important features for predicting floral, off-flavor, and nutty notes (Supplementary Fig. 8) with RDKit molecular descriptors. Six descriptors appeared repeatedly in all three cases, accounting for around 45% of the classification. These descriptors measure properties such as the size and polarity of molecules (TPSA), their electronic structure (PEOE_VSA and EState_VSA) and stability (SMR_VSA1 and MinEStateIndex), and their tendency to partition into a hydrophobic environment (MolLogP). Supplementary Fig. 9 shows the trend of the five most relevant features for positive and negative examples of each flavor note. Off-flavor molecules tend to be smaller and less polar than non-off-flavor molecules, with a higher tendency to partition into hydrophobic environments. Floral molecules tend to be smaller and more flexible than non-floral molecules, with a higher tendency to partition into hydrophilic environments. Finally, nutty molecules tend to be smaller and less flexible than non-nutty molecules, with a higher electronic stability. These results are new for these flavor notes and provide a basis for future research to select more specific mathematical representations and use data mining techniques to better understand why molecules have these flavors. Supplementary Fig. 10 shows the VIP scores for the Random Forest models trained on oversampled ECFP descriptors for predicting bitterness, fruitiness, and sweetness. The four most important bits for the binary classifiers predicting these flavor notes were 897, 314, 489, and 463. The fragments corresponding to these bits are shown in Supplementary Fig. 4. For the K-Nearest Neighbors algorithm, the permutation importance score[ 43 ] was used to estimate feature importance because in this case it is not possible to use the VIP score (Fig. 4 ). Interestingly, most of the top five fingerprints for these notes corresponded to fragments that were absent in the positive compounds. This is likely due to the higher chemical diversity of the negative compounds. E.g. many typical bitter compounds contain an (alkaloid) nitrogen, but no N-containing fragment appeared in the top 5 for bitter, actually many top fragments like bit 897 (C-O-C - moiety) appeared in bitter, fruity and sweet, i.e. they are of universal flavor relevance, but for note determination probably play there role obviously only in the context with other features (e.g. in esters for fruity or sweet in cyclic sugars). Even though resampling strategies were implemented to improve the overall performance of the models, this did not necessarily enhance the chemical diversity of the positive examples. Fingerprints have two main advantages over molecular descriptors. First, they can provide information about the structural features that lead an algorithm to a certain decision. Future work could involve a deeper analysis of the fragments that play a central role in the classification to better understand the structural features that underlie these flavor notes. Second, they can be calculated from MS-spectra data, even when the structure of a compound is not fully elucidated[ 44 , 45 ]. This makes fingerprints useful for accelerating the discovery of new flavor molecules in metabolomics experiments. Metabolomics experiments typically involve the analysis of many compounds and ECFP can help concentrate the annotation and structural elucidation on the most promising candidates. This can save time and money by focusing efforts on the most promising and likely flavor-active compounds. The CGNModel combined with molecular graphs showed poor performance, even with a balancing transformer. This is likely due to the inherent noisiness of the data, which is exacerbated by the susceptibility of Graph Neural Networks to noisy data[ 46 , 47 ]. This noisiness arises from the heavy dependence of flavor characterization on human tasters and the influence of genetic, sensory, and environmental factors on flavor perception[ 3 , 16 , 48 , 49 ]. It is challenging to implement a denoising strategy without losing valuable information. Therefore, the CGNModel was discarded for FlavorMiner, given the limitations of the current data and the better performance of other algorithms. The flavor profile of a molecule also depends on its concentration and the surrounding matrix[ 50 , 51 ]. This is related to the concept of flavor threshold and the synergistic and antagonistic effects of flavor molecules in complex mixtures. The flavor threshold is the minimum concentration at which the flavor is detectable[ 50 , 51 ]. This version of FlavorMiner only performs binary prediction, and intensity data is not yet incorporated. Although some data is available, it is not readily accessible, as there is no standardized database of threshold concentrations for molecules with known flavor profiles. Some databases such as FlavorDB[ 6 , 7 ] and LSB@TUM Odorant Database ( https://www.leibniz-lsb.de/en/databases/leibniz-lsbtum-odorant-database/start/ ) contain information on flavor thresholds. However, there is a lack of standardization in the thresholds reported in these databases. This means that a method is needed to unify and make this data comparable. Also, most information on flavor thresholds is available in unstructured format (text). Therefore, an intense text mining process is required to extract this data and make it usable for machine learning purposes. Additionally, some studies have shown that combining several molecules with different flavor profiles can enhance the flavor profile of a mixture or block certain notes[ 50 , 51 ]. However, data in this area is limited, and any effort in this direction will require a preliminary experimental process to generate it. Overcoming these challenges could lead to the development of regression algorithms that can be combined with flavor classifiers to predict not only the flavor profile of a molecule but also its threshold concentration and matrix effect. Sour (like salty not evaluated here) is a special flavor note, as it relies on the smallest available “molecule”, the proton. Also, it is not activating a classical GPCR like the other taste (T1R and T2R) or the olfactory receptors. Only quite recently the responsible Otop1 ion channels were assigned[ 52 ]. Thus, typical structural features of a molecule might be considered irrelevant, except for its pKa properties, i.e. its ability to lower pH, an effect that will strongly depend on the matrix’ overall pH, buffer capacity and may be the proton relay/ion transport capacity. Thus, predicting sour taste from structure might be considered impossible if only the pH change is sensed. However, like GPCRs, ion channels can be influenced by more than the ion it is selective for for various reasons, including ion pairing and matrix/mucosa effects or directly at the ion channel by secondary interactions and additional binding sites which will have selective structural preferences as every protein does. In conclusion, structure-based predictions for ion channel-based tastes (here sour, but also salty) have to be considered with caution, as slight changes in the tasting parameters, e.g. of the matrix (pH, buffer capacity) can thwart results and thus all ML. To understand, if there is sour taste influence on the anionic, organic (i.e. structurally influenced) part, such taste experiments must run with a standardized, high-capacity buffered matrix, neutral pH or better at 2–3 different pH values. Only this can reveal any possible structural influences of the organic counterion or a neutral molecule influencing or mimicking sour taste. Otherwise, it will not be better than a standard pKa prediction which does not require ML. Independent of this, perception is also influenced by the other receptors. A classic example is of course the action of Miraculin. 2.3. Molecular flavor prediction for compounds involved in the processing of cocoa. Previous studies have annotated around 210 compounds during the fermentation, drying, and roasting of fine-flavor cocoa[ 53 , 54 ]. However, for less than half of these compounds a flavor profile has been reported. The existing data were analyzed with FlavorMiner to predict the flavor profile of these compounds. After the prediction, the compounds with “known” flavor profiles increased to 92%. The newly predicted compounds include 12 floral, 8 fruity, and 4 compounds with unknown fine-flavor attributes that are potentially linked to positive impacts on quality and price. Additionally, 2 compounds linked to off-flavors and 27 unknown potentially sweet compounds were suggested by the model. These predictions represent an important step forward in closing the gap between cocoa metabolic fingerprint variation during processing and flavor quality. Figure 5 shows the frequency of compounds increasing in association with each of the seven target flavors at the end of every cocoa processing stage (Fermentation, Drying, and Roasting). In general, the frequency of compounds for the different target flavors is similar during fermentation and drying. The most relevant change through the processing chain is in sweet compounds, which decrease considerably during the process. This drop is associated with a decrease in the carbohydrate content during the processing chain[ 53 , 54 ], as most of these molecules are reported as sweet agents. In the roasted samples some compounds linked to sour and bitter showed a higher abundance, but the real impact of these suggested flavor molecules still needs to be elucidated. For example, some degradation products of more complex compounds have a lower biological activity (e.i., antioxidant activity) than their precursors[ 55 ]. If a similar trend occurs with respect to flavor will require further investigations. In contrast, most compounds linked to fine flavor notes (fruity, floral, and nutty) show a relatively constant frequency throughout the cocoa processing chain. These results provide further suggestions into flavor development from biochemistry to processing, which was a missing component until now. CONCLUSIONS The present work led to the development of FlavorMiner, an open-access flavor ML predictor for fruity, floral, off-flavor, nutty, sour, sweet, and bitter notes. Access sites are: https://github.com/ipb-halle/FlavorMiner . This fills a gap in the current landscape of flavor prediction tools, as there are few open-access options available for predicting these specific flavor notes. FlavorMiner has an average ROC AUC score of 0.88. This algorithm performs similarly to other bitterness and sweetness predictors but outperforms existing floral and fruity predictors. It is also the first tool to predict nutty and off-flavor notes. This work also showed the superiority of RDKit descriptors and ECFP over molecular graphs (combined with Convolutional Graph Neural Network) as mathematical representations to predict flavor notes from molecular structures. Molecular graphs had a recall lower than 0.5 in most cases, even after balancing the classes with a transformer. This means that they currently fail to accurately identify molecular structures with the target flavors. On the other hand, models trained with RDKit descriptors and ECFP had a recall over 0.6 in most cases, especially after resampling with the SMOTE algorithm. The best performers were selected to assemble the backbone of FlavorMiner, reaching a recall of over 0.7 in most cases. Further analyses of the best performers revealed the following properties that can define a molecule as positive or negative for floral, off-flavor, and nutty notes such as the size, polarity, electronic structure, stability, and their tendency to partition into a hydrophobic environment. Similarly, this analysis revealed the most critical fragments to define as positive or negative in a molecule for bitter, fruity, sour, and sweet, whereby predictions for sour must be considered with special caution. These results offer a solid base for future studies aiming to dissect the chemistry of flavor properties. Using FlavorMiner with cocoa metabolomics data highlighted its potential to elucidate the molecular basis of flavor development in processed food products. To gain further insights into the chemistry and pharmacology of flavor, additional functionalities such as flavor threshold and interaction prediction should be included. However, several challenges must be overcome, starting with (unified) data availability. It is also important to note that machine learning techniques should complement traditional experimental methods in flavor prediction. Integrating the strengths of both approaches can enhance the overall accuracy and efficiency of flavor prediction and lead to new discoveries in the field. MATERIALS AND METHODS 5.1. Data collection and preprocessing The data for this study were collected from multiple sources, including FlavorDB, Sigma-Aldrich catalog, BitterSweet, ChemTastesDB, Flavornet, Off-flavor, AromaDB, OlfactionBase, and Natural_tAS2r_agonists[ 16 ]. The data were cleaned by removing duplicate entries, molecules without reported flavor profiles, invalid PubChem identifiers or names, molecules with fewer than 2 heavy atoms, and salts, knowing that especially the latter fact may depent on representation (in ionic or non-ionic form) of some molecules, which may also have contributed to the problems with the sour set. The linear structures of the remaining molecules were retrieved from the PubChem database using the Python library PubChemPy. The compounds were labeled with the flavor information retrieved from the databases. The flavor notes were grouped into seven categories: bitter, floral, fruity, off-flavor, nutty, sour, and sweet. The labeling process was automated in a Python script and a flavor wheel[ 56 ] to facilitate future relabeling or adding new data. The seven target labels were then converted into binary values using the One Hot Encoding method[ 30 ]. This information was stored in an Excel file containing the compound name, PubChem ID, flavor profile, isomeric smiles, data source, and labels (Supplementary File 1). 5.2. Mathematical representation of molecular structures Three mathematical representations of the molecular structures were generated. First, 200 molecular descriptors were calculated using the RDKit library. Descriptors with invalid data, more than 97% of unique values, and descriptors highly correlated were removed. For this, the Pearson correlation index was used with a threshold of 0.95[ 57 ]. Second, an Extended Connectivity Fingerprint (ECFP) with a radius of 2 was generated for all compounds using the RDKit library. Fingerprints with more than 97% unique values were also removed. The datasets corresponding to the RDKit molecular descriptors and ECFP were split into training and testing sets using a random partition of 80:20. A third mathematical representation was generated by creating molecular graphs from the molecular structures using the DeepChem library. The DeepChem MolGraphConvFeaturizer was used to generate the molecular graphs. The graphs were then labeled for each target flavor with a binary label and converted into NumPy datasets. This dataset was divided into a training, validation, and testing set using a random partition of 70:10:20. 5.3. Machine Learning algorithms training, optimization, and testing Independent binary classifiers were used to predict each flavor category. Initially, a Random Forest and K-Nearest Neighbors algorithm were trained on the RDKit molecular descriptors and ECFP datasets using scikit-learn. Hyperparameter optimization was performed using the Grid Search method and 5-fold cross-validation. Class imbalance was addressed by oversampling the minority classes using SMOTE (Synthetic Minority Oversampling Technique) and undersampling the majority classes using the cluster centroid of a KMeans algorithm. A Graph Convolutional Neural Network (GCNModel) was then trained on the molecular graphs using DeepChem. Hyperparameter optimization of the GCNModel was performed using the hyperopt library. The balance of the molecular graph data was done using the DeepChem balancing transformer. The performance of the algorithms was evaluated using the recall, specificity and Receiving Operator Characteristic (ROC) curves[ 30 , 31 ]. 5.4. Using cocoa processing metabolomics data as a case study Two previous datasets from metabolomics studies of fine-flavor cocoa processing were used for this case study[ 53 , 54 ]. The first study used LC-QTOF-MS and GC-QTOF-MS to analyze the fermentation of fine-flavored cocoa[ 53 ]. The second study used the same analytical platforms to assess the evolution of flavor during post-fermentation processing (drying and roasting) of fine-flavor cocoa[ 54 ]. The PubChem ID for each molecule was obtained from the PubChem database. This data was stored in an Excel file and fed to FlavorMiner to assign flavor profiles. Declarations Availability of data and materials All the datasets, and scripts used in this work are publicly available in a GitHub repository (https://github.com/ipb-halle/FlavorMiner). FlavorMiner and all Python scripts are available in Jupyter Notebooks. These Jupyter Notebooks contain a detailed explanation of the code and the data, which is integrated with the GitHub repository, reducing the need to download files locally to run the scripts. Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this manuscript. Funding The authors thank the program Impacto País of the Vice-presidency of Research and Creation of the University of Los Andes and the Jeff Schell Fellowship of the Bayer Foundation, for the funding provided to Fabio Herrera-Rocha. IATA-CSIC received funding from the Spanish government, ref. MCIN/AEI/10.13039/501100011033, as a ‘Severo Ochoa’ Center of Excellence (CEX2021-001189-S). Finally, the authors acknowledge the Alexander von Humboldt Foundation of Germany for the financial support to Miguel Angel Fernández-Niño and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - SPP2363 for funding provided to Mehdi D. Davari. Authors contributions FHR: manuscript writing, code development, repository construction, investigation, validation, benchmarking, and machine learning. JD: conceptualization, manuscript writing, code development, validation, and supervision. MFN: manuscript writing, conceptualization. MPC: conceptualization, manuscript writing, validation, and supervision. MJC: conceptualization, manuscript writing, validation, and supervision. LAW: conceptualization, manuscript writing, validation, and supervision. MD: conceptualization, manuscript writing, validation, and supervision. AFGB: conceptualization, manuscript writing, validation, and supervision. All authors have reviewed and approved the final version of the manuscript. Acknowledgments The authors would like to acknowledge the Group of Product and Process Design (GDPP), in the Department of Chemical and Food Engineering of the University of Los Andes, and the Department of Bioorganic Chemistry for their valuable insights and thoughtful recommendations. Finally, the authors thank Exacore of the University of Los Andes for the High-Performance Computing (HPC) services provided to this project. References Mouritsen OG (2015) The science of taste. Flavour 4:10–11. https://doi.org/10.1186/s13411-014-0028-3 Wang YT, Yang ZX, Piao ZH, et al (2021) Prediction of flavor and retention index for compounds in beer depending on molecular structure using a machine learning method. RSC Adv 11:36942–36950. https://doi.org/10.1039/d1ra06551c Keller A, Gerkin RC, Guan Y, et al (2017) Predicting human olfactory perception from chemical features of odor molecules. Science (80- ) 355:820–826. https://doi.org/10.1126/science.aal2014 Queiroz LP, Rebello CM, Costa EA, et al (2023) Generating Flavor Molecules Using Scientific Machine Learning. ACS Omega 10875–108887. https://doi.org/10.1021/acsomega.2c07176 Anastasiadi M, Mohareb F, Redfern SP, et al (2017) Biochemical Profile of Heritage and Modern Apple Cultivars and Application of Machine Learning Methods to Predict Usage, Age, and Harvest Season. J Agric Food Chem 65:5339–5356. https://doi.org/10.1021/acs.jafc.7b00500 Garg N, Sethupathy A, Tuwani R, et al (2018) FlavorDB: A database of flavor molecules. Nucleic Acids Res 46:1210–1216. https://doi.org/10.1093/nar/gkx957 Grover N, Goel M, Batra D, et al (2022) FlavorDB2: An Updated Database of Flavor Molecules. arXiv 1–5 Naveja JJ, Rico-Hidalgo MP, Medina-Franco JL (2018) Analysis of a large food chemical database: chemical space, diversity, and complexity. F1000Research 7:993. https://doi.org/10.12688/F1000RESEARCH.15440.2 Yu M, Yang P, Song H, Guan X (2022) Research progress in comprehensive two-dimensional gas chromatography-mass spectrometry and its combination with olfactometry systems in the flavor analysis field. J Food Compos Anal 114:104790. https://doi.org/10.1016/j.jfca.2022.104790 Li Y, Yuan L, Liu H, et al (2022) Analysis of the changes of volatile flavor compounds in a traditional Chinese shrimp paste during fermentation based on electronic nose, SPME-GC-MS and HS-GC-IMS. Food Sci Hum Wellness 12:173–182. https://doi.org/10.1016/j.fshw.2022.07.035 Degenhardt A, Wittlake R, Seilwind S, et al (2014) Quantification of Important Flavor Compounds in Beef Stocks and Correlation to Sensory Results by “Reverse Metabolomics.” Elsevier Inc. Wardencki W, Chmiel T, Dymerski T (2013) Gas chromatography-olfactometry (GC-O), electronic noses (e-noses) and electronic tongues (e-tongues) for in vivo food flavour measurement. Woodhead Publishing Limited Yimenu SM, Kim JY, Kim BS (2017) Prediction of egg freshness during storage using electronic nose. Poult Sci 96:3733–3746. https://doi.org/10.3382/ps/pex193 Rottiers H, Tzompa Sosa DA, Van de Vyver L, et al (2019) Discrimination of Cocoa Liquors Based on Their Odor Fingerprint: a Fast GC Electronic Nose Suitability Study. Food Anal Methods 12:475–488. https://doi.org/10.1007/s12161-018-1379-7 Tseng YJ, Chuang PJ, Appell M (2023) When Machine Learning and Deep Learning Come to the Big Data in Food Chemistry. ACS Omega 15854–15864. https://doi.org/10.1021/acsomega.2c07722 Kou X, Shi P, Gao C, et al (2023) Data-Driven Elucidation of Flavor Chemistry. J Agric Food Chem 6789–6802. https://doi.org/10.1021/acs.jafc.3c00909 Queiroz LP, Rebello CM, Costa EA, et al (2023) Transfer Learning Approach to Develop Natural Molecules with Specific Flavor Requirements. 9062–9076. https://doi.org/10.1021/acs.iecr.3c00722 Sanchez-Lengeling B, Wei JN, Lee BK, et al (2019) Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules. arXiv Dutta P, Jain D, Gupta R, Rai B (2022) Classification of Tastants: A Deep Learning Based Approach. ChemRxiv Rojas C, Ballabio D, Consonni V, et al (2023) Classification-based machine learning approaches to predict the taste of molecules: A review. Food Res Int 171:. https://doi.org/10.1016/j.foodres.2023.113036 Zheng S, Jiang M, Zhao C, et al (2018) e-Bitter: Bitterant prediction by the consensus voting from the machine-learning methods. Front Chem 6:1–18. https://doi.org/10.3389/fchem.2018.00082 Tuwani R, Wadhwa S, Bagler G (2019) BitterSweet: Building machine learning models for predicting the bitter and sweet taste of small molecules. Sci Rep 9:1–13. https://doi.org/10.1038/s41598-019-43664-y Martinez-Mayorga K, Peppard TL, López-Vallejo F, et al (2013) Systematic mining of generally recognized as safe (GRAS) flavor chemicals for bioactive compounds. J Agric Food Chem 61:7507–7514. https://doi.org/10.1021/jf401019b Dagan-Wiener A, Nissim I, Ben Abu N, et al (2017) Bitter or not? BitterPredict, a tool for predicting taste from chemical structure. Sci Rep 7:1–13. https://doi.org/10.1038/s41598-017-12359-7 Fritz F, Preissner R, Banerjee P (2021) VirtualTaste: A web server for the prediction of organoleptic properties of chemical compounds. Nucleic Acids Res 49:W679–W684. https://doi.org/10.1093/nar/gkab292 Banerjee P, Preissner R (2018) Bitter sweet forest: A Random Forest based binary classifier to predict bitterness and sweetness of chemical compounds. Front Chem 6:1–10. https://doi.org/10.3389/fchem.2018.00093 Nozaki Y, Nakamoto T (2018) Correction: Predictive modeling for odor character of a chemical using machine learning combined with natural language processing (PLoS ONE (2018) 13, 6 (e0198475) DOI: 10.1371/journal.pone.0198475). PLoS One 13:1–13. https://doi.org/10.1371/journal.pone.0208962 Shang L, Liu C, Tomiura Y, Hayashi K (2017) Machine-Learning-Based Olfactometer: Prediction of Odor Perception from Physicochemical Features of Odorant Molecules. Anal Chem 89:11999–12005. https://doi.org/10.1021/acs.analchem.7b02389 Sharma A, Kumar R, Ranjta S, Varadwaj PK (2021) SMILES to Smell: Decoding the Structure-Odor Relationship of Chemical Compounds Using the Deep Neural Network Approach. J Chem Inf Model 61:676–688. https://doi.org/10.1021/acs.jcim.0c01288 Ramsundar B, Eastman P, Walters P, Pande V (2019) Deep Learning for the Life Sciences Géron A (2019) Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow, 2nd Editio Lo YC, Rensi SE, Torng W, Altman RB (2018) Machine learning in chemoinformatics and drug discovery. Drug Discov Today 23:1538–1546. https://doi.org/10.1016/j.drudis.2018.05.010 Sun Y, Wong AKC, Kamel MS (2009) Classification of imbalanced data: A review. Int J Pattern Recognit Artif Intell 23:687–719. https://doi.org/10.1142/S0218001409007326 Banerjee P, Dehnbostel FO, Preissner R (2018) Prediction is a balancing act: Importance of sampling methods to balance sensitivity and specificity of predictive models based on imbalanced chemical data sets. Front Chem 6:387941. https://doi.org/10.3389/FCHEM.2018.00362/BIBTEX Reiser P, Neubert M, Eberhard A, et al (2022) Graph neural networks for materials science and chemistry. Commun Mater 2022 31 3:1–18. https://doi.org/10.1038/s43246-022-00315-6 von Lilienfeld OA, Burke K (2020) Retrospective on a decade of machine learning for chemical discovery. Nat Commun 2020 111 11:1–4. https://doi.org/10.1038/s41467-020-18556-9 Leevy JL, Khoshgoftaar TM, Bauder RA, Seliya N (2018) A survey on addressing high-class imbalance in big data. J Big Data 5:. https://doi.org/10.1186/s40537-018-0151-6 Malavolta M, Pallante L, Mavkov B, et al (2022) A survey on computational taste predictors. Eur Food Res Technol 248:2215–2235. https://doi.org/10.1007/s00217-022-04044-5 De La Fuente A, Zamberlan F, Sánchez Ferrán A, et al (2020) Relationship among subjective responses, flavor, and chemical composition across more than 800 commercial cannabis varieties. J Cannabis Res 2:. https://doi.org/10.1186/S42238-020-00028-Y Yen SJ, Lee YS (2009) Cluster-based under-sampling approaches for imbalanced data distributions. Expert Syst Appl 36:5718–5727. https://doi.org/10.1016/J.ESWA.2008.06.108 Hu F, Wang L, Liu Q, et al (2022) GraphDIVE: Graph Classification by Mixture of Diverse Experts. IJCAI Int Jt Conf Artif Intell 3:2080–2086. https://doi.org/10.24963/IJCAI.2022/289 Saini K, Ramanathan V (2022) Predicting odor from molecular structure: a multi-label classification approach. Sci Rep 12:1–11. https://doi.org/10.1038/s41598-022-18086-y Altmann A, Toloşi L, Sander O, Lengauer T (2010) Permutation importance: a corrected feature importance measure. Bioinformatics 26:1340–1347. https://doi.org/10.1093/BIOINFORMATICS/BTQ134 Dührkop K, Nothias LF, Fleischauer M, et al (2021) Systematic classification of unknown metabolites using high-resolution fragmentation mass spectra. Nat Biotechnol 39:462–471. https://doi.org/10.1038/s41587-020-0740-8 Ji H, Deng H, Lu H, Zhang Z (2020) Predicting a Molecular Fingerprint from an Electron Ionization Mass Spectrum with Deep Neural Networks. Anal Chem 92:8649–8653. https://doi.org/10.1021/ACS.ANALCHEM.0C01450/SUPPL_FILE/AC0C01450_SI_002.XLS Dai E, Jin W, Liu H, Wang S (2022) Towards robust graph neural networks for noisy graphs with sparse labels. WSDM 2022 - Proc 15th ACM Int Conf Web Search Data Min 181–191. https://doi.org/10.1145/3488560.3498408 Godwin J, Schaarschmidt M, Gaunt A, et al (2022) Simple Gnn Regularisation for 3D Molecular Property Prediction & Beyond. ICLR 2022 - 10th Int Conf Learn Represent 1–23 Poivet E, Tahirova N, Peterlin Z, et al (2018) Functional odor classification through a medicinal chemistry approach. Sci Adv 4:. https://doi.org/10.1126/sciadv.aao6086 Khan B (2018) Application of pattern mining on data of flavor molecules , their percepts and molecular features Mojet J, Heidema J, Christ-Hazelhof E (2004) Effect of concentration on taste-taste interactions in foods for elderly and young subjects. Chem Senses 29:671–681. https://doi.org/10.1093/CHEMSE/BJH070 Marks LE, Shepard TG, Burger K, Chakwin EM (2012) Flavor-Intensity Perception: Effects of Stimulus Context. Physiol Behav 105:443. https://doi.org/10.1016/J.PHYSBEH.2011.08.039 Turner HN, Liman ER (2022) The Cellular and Molecular Basis of Sour Taste. Annu Rev Physiol 84:41–58. https://doi.org/10.1146/annurev-physiol-060121-041637 Herrera-Rocha F, Cala MP, Mejía Aguirre LJ, et al (2021) Dissecting fine-flavor cocoa bean fermentation through metabolomics analysis to break down the current metabolic paradigm. Sci Rep 11:. https://doi.org/10.1038/s41598-021-01427-8 Herrera-Rocha F, León-Inga AM, Aguirre Mejía JL, et al (2024) Bioactive and flavor compounds in cocoa liquor and their traceability over the major steps of cocoa post-harvesting processes. Food Chem 435:. https://doi.org/10.1016/j.foodchem.2023.137529 Mcshea A, Ramiro-Puig E, Munro SB, et al (2008) Clinical benefit and preservation of flavonols in dark chocolate manufacturing. https://doi.org/10.1111/j.1753-4887.2008.00114.x Spencer M, Sage E, Velez M, Guinard JX (2016) Using Single Free Sorting and Multivariate Exploratory Methods to Design a New Coffee Taster’s Flavor Wheel. J Food Sci 81:S2997–S3005. https://doi.org/10.1111/1750-3841.13555 Rácz A, Bajusz D, Héberger K (2019) Intercorrelation Limits in Molecular Descriptor Preselection for QSAR/QSPR. Mol Inform 38:1–6. https://doi.org/10.1002/minf.201800154 Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformationFlavorMiner06272024.docx Cite Share Download PDF Status: Published Journal Publication published 10 Dec, 2024 Read the published version in Journal of Cheminformatics → Version 1 posted Editorial decision: Revision requested 20 Jul, 2024 Editor assigned by journal 20 Jul, 2024 Submission checks completed at journal 20 Jul, 2024 First submitted to journal 19 Jul, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4768238","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":329597958,"identity":"96ed57ce-6641-4d70-86b7-15e710db083a","order_by":0,"name":"Fabio Herrera-Rocha","email":"","orcid":"","institution":"Leibniz Institute of Plant Biochemistry","correspondingAuthor":false,"prefix":"","firstName":"Fabio","middleName":"","lastName":"Herrera-Rocha","suffix":""},{"id":329597959,"identity":"0be3718e-790c-4c06-a23f-cce30e396e4c","order_by":1,"name":"Miguel Fernández-Niño","email":"","orcid":"","institution":"Leibniz Institute of Plant Biochemistry","correspondingAuthor":false,"prefix":"","firstName":"Miguel","middleName":"","lastName":"Fernández-Niño","suffix":""},{"id":329597960,"identity":"98cd8a2a-486f-4252-92e3-2d59ed380176","order_by":2,"name":"Jorge Duitama","email":"","orcid":"","institution":"Universidad de Los Andes","correspondingAuthor":false,"prefix":"","firstName":"Jorge","middleName":"","lastName":"Duitama","suffix":""},{"id":329597961,"identity":"79c5653d-12be-4bc1-b4fd-e26179169c93","order_by":3,"name":"Mónica P. Cala","email":"","orcid":"","institution":"Universidad de Los Andes","correspondingAuthor":false,"prefix":"","firstName":"Mónica","middleName":"P.","lastName":"Cala","suffix":""},{"id":329597962,"identity":"7dc4f1f8-f3a8-4cf3-bc21-9f1f7b72158b","order_by":4,"name":"María José Chica","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"María","middleName":"José","lastName":"Chica","suffix":""},{"id":329597963,"identity":"a64721dd-aaa3-428b-a521-8dce55de419c","order_by":5,"name":"Ludger A. Wessjohann","email":"","orcid":"","institution":"Leibniz Institute of Plant Biochemistry","correspondingAuthor":false,"prefix":"","firstName":"Ludger","middleName":"A.","lastName":"Wessjohann","suffix":""},{"id":329597964,"identity":"f08cbc84-0a12-44f2-b4db-fb5c3fe62d49","order_by":6,"name":"Mehdi D. Davari","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABEklEQVRIie2QMUvDQBiG35A14HqgxV8gXAmEQCv+lTsKveWQgksHh4BQF+XW5l/UX9ALB+0SnSMIxiWTQ8YOIp6pFBxOOwreM33D99z7vQd4PH+SCIQh7EbdUjAQBLoGgmwfpZhvFWj2m4IvJYywh3JyfVtUNQbHdH3/YoaTp3McXtWaTdFTDiUpH0Ypg+gvSkGNpM0FjlZUsxJx7ohJKpnYLibIszGsYviyYlTzGfhCO5Tn1045y1UDk1olI6LV/B186VKqqFO4IjYFnSJtSmZTnF1knDIqRoo0KG4+uxA50WxF4rkrZV32HzfTwensYBy2mzf7Y0Tc1e3lsKcc9bfQ3bR7mfy0/w3HMR6Px/Ov+QBguGZLnra/rgAAAABJRU5ErkJggg==","orcid":"","institution":"Leibniz Institute of Plant Biochemistry","correspondingAuthor":true,"prefix":"","firstName":"Mehdi","middleName":"D.","lastName":"Davari","suffix":""},{"id":329597965,"identity":"ffc942a4-6c13-469d-8682-83952782c74b","order_by":7,"name":"Andres Fernando González Barrios","email":"","orcid":"","institution":"Universidad de Los Andes","correspondingAuthor":false,"prefix":"","firstName":"Andres","middleName":"Fernando González","lastName":"Barrios","suffix":""}],"badges":[],"createdAt":"2024-07-19 14:22:04","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4768238/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4768238/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s13321-024-00935-9","type":"published","date":"2024-12-10T15:58:02+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":62731654,"identity":"48860530-707d-4939-aefe-f6938a610ebd","added_by":"auto","created_at":"2024-08-18 23:33:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":91436,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the FlavorMiner algorithm for flavor prediction and data visualization. The algorithm requires input compounds structured in Isomeric SMILES format, essential for both database matching and prediction steps.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/33b7b8a67751ff207acf947b.png"},{"id":62732186,"identity":"c9610677-6971-4f18-8d13-ab37bb525cee","added_by":"auto","created_at":"2024-08-18 23:41:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":47770,"visible":true,"origin":"","legend":"\u003cp\u003eClassification Metrics for Algorithms Trained with Original Descriptor Data on Test Set. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. \u003cstrong\u003e(a)\u003c/strong\u003e Random Forest trained with molecular descriptors. \u003cstrong\u003e(b) \u003c/strong\u003eRandom Forest trained with extended connectivity fingerprint. \u003cstrong\u003e(c)\u003c/strong\u003e K-Nearest Neighbors trained with molecular descriptors. \u003cstrong\u003e(d) \u003c/strong\u003eK-Nearest Neighbors trained with extended connectivity fingerprint. \u003cstrong\u003e(e)\u003c/strong\u003e Convolutional Graph Neural Network trained with molecular graph.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/b6f41d5b6730818a4ecddd00.png"},{"id":62731659,"identity":"a1a56b43-de02-48fe-b67a-b95f6980e22c","added_by":"auto","created_at":"2024-08-18 23:33:22","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":35534,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of the\u003cstrong\u003e \u003c/strong\u003eOptimized Classifiers for Target Flavor Notes in FlavorMiner. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. \u003cstrong\u003e(a) \u003c/strong\u003eClassification metrics obtained during training using 5-fold cross-validation. \u003cstrong\u003e(b) \u003c/strong\u003eClassification metrics obtained using the test set. Random Forest was used for bitter, fruity, sweet, off-flavor, floral and nutty. K-Nearest Neighbors for sour notes.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/1b089947db3c4c992af144f7.png"},{"id":62731656,"identity":"25ee3d91-49ec-41e2-a967-32cbc682d272","added_by":"auto","created_at":"2024-08-18 23:33:21","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":219807,"visible":true,"origin":"","legend":"\u003cp\u003eFragments Representing High VIP Scores in Random Forest Predictors with ECFP. This figure displays fragments corresponding to the 5 bits with the highest Variable Importance (VIP) scores in Random Forest predictors trained with Extended Connectivity Fingerprint (ECFP). (a) Bitter, (b) Fruity, (c) Sweet.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/a3632f43ee494d08ee152287.png"},{"id":62731658,"identity":"1a8f705d-e503-4d20-842a-02bb9ddbb31e","added_by":"auto","created_at":"2024-08-18 23:33:22","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":123971,"visible":true,"origin":"","legend":"\u003cp\u003eMolecule Counts for Each Target Flavor Across Cocoa Processing Stages\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/a4eb422bcdaaa7177294fea9.png"},{"id":71552506,"identity":"134a9e7a-9ba3-4172-a261-02507760e413","added_by":"auto","created_at":"2024-12-16 16:06:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1059428,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/78bfb711-b274-43f4-beb7-be2a045320f5.pdf"},{"id":62731657,"identity":"ea30c725-5658-4459-bccb-602cd00336e8","added_by":"auto","created_at":"2024-08-18 23:33:21","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":864598,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformationFlavorMiner06272024.docx","url":"https://assets-eu.researchsquare.com/files/rs-4768238/v1/7966312925b856d116ba9b2f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eFlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data\u003c/p\u003e","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eFlavor is defined as the combination between taste and odor, without distinction, of a substance or product[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. This feature plays an essential role in defining consumer's acceptance of foods and beverages. The flavor properties of most processed food products are readily manipulable by formulating the proper ingredients to meet the desired sensory quality. Nonetheless, in the case of fermented products, matured foods, fruits, and vegetables, the flavor is purely determined by their chemical composition (including that of the matrix) [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The flavor molecules are produced in complex biological processes or chemical reactions and may be modified during processing, causing a pleasant or unpleasant sensation depending on their interaction with flavor receptors[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Hence, understanding the impact of these compounds on the flavor profile is essential for engineering and decision-making in food and beverage production. This knowledge is critical for discovering and developing new flavors, standardizing and optimizing processing conditions, and variety selection to produce plant-based foods with a better flavor, among other applications[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe FlavorDB database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cosylab.iiitd.edu.in/flavordb/\u003c/span\u003e\u003cspan address=\"https://cosylab.iiitd.edu.in/flavordb/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) is the most comprehensive collection of molecules with an experimentally validated flavor profile[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. This database also contains information from other flavor databases and external sources such as PubChem. Although this database contains more than 25,500 compounds, only 2,254 metabolites (~\u0026thinsp;9%) are linked to the 936 food products[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. In contrast, FooDB (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://foodb.ca/\u003c/span\u003e\u003cspan address=\"https://foodb.ca/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), the largest food metabolomics database, contains more than 24,000 compounds identified in foods[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. The number of food-related molecules without flavor profiles will probably continue increasing in the future, boosted by the growing power of high-throughput metabolomics techniques and the complexity of testing the flavor profile of individual molecules. In addition, for most synthetic compounds no flavor profile is known.\u003c/p\u003e \u003cp\u003eThe current experimental methodologies to assess the flavor profile of individual molecules require the compound to be isolated or synthesized[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Alternatively, the flavor is inferred by correlating metabolomics data with sensory results or tested using trained panelists [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This procedure is considerably time-consuming and expensive, especially considering the complex composition of most food products. A recent alternative to streamline this process is the procedure called ultra-fast GC E-nose, which is an automated methodology to discriminate samples based on their odor fingerprint[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The main drawback of this approach is that it does not assign a flavor profile to single compounds. Similarly, Gas Chromatography-olfactometry (GC-O) combines the power of gas chromatography with the sensitivity of the human nose to assign a flavor profile to individual compounds[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. However, only volatile compounds can be tested following this technique, also perception even of an identical compound can be different when ingested orally in a matrix.\u003c/p\u003e \u003cp\u003eSimilar to several other knowledge domains, different Machine Learning (ML) models have been developed to perform \u003cem\u003ein-silico\u003c/em\u003e flavor prediction from molecular structures with available flavor profiles[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. This trend has been prompted by advancements in ML algorithms and the availability of large-scale molecular data[\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. This approach enables an efficient screening of potential flavors to prioritize compounds for validation using traditional experimental methods, saving time and resources[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Nevertheless, ML models require diverse datasets to learn effectively. This variability is even more relevant for flavor prediction because molecules with similar structures can have completely different flavors or divergent molecular structures can have a similar flavor profile[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Obtaining high-quality flavor data with sufficient coverage of different molecular structures can be challenging. In some cases, data may be limited or biased, leading to potential inaccuracies in predictions.\u003c/p\u003e \u003cp\u003eMost studies in this area used binary classifiers concentrated on predicting sweet and bitter flavors, mainly because of the availability of large datasets with labeled sweet and bitter molecules[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. These developments were also fostered by the need to address consumer preferences and health concerns. With increasing interest in reduced sugar consumption, the food industry is demanding alternative sweeteners or taste enhancers mimicking the desired tastes while reducing sugar content. Similarly, predicting bitterness or off-flavors is crucial for avoiding unpleasant taste experiences. Another reason boosting bitterness prediction is the wide prevalence of this flavor note in natural bioactive compounds, serving then as a potential screening tool[\u003cspan additionalcitationids=\"CR23\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Additionally, some predictors for sourness and umami have been published, although with less success because the availability of labeled data is limited for these flavor notes[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe existing public flavor predictors cannot predict notes like floral, nutty, fruity, or off-flavor[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Some attempts have been undertaken in this area, but the performance is still poor and the code and data are not publicly available[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan additionalcitationids=\"CR28\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], leading to a lack of available multilabel tools for flavor prediction. These flavor notes are of capital importance in fermented food processing (including coffee, beer, wine, chocolate, bread and others)[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Also, some models based on Generative Artificial Intelligence have been trained to generate new molecules with flavors potentially interesting for the food industry (including the above-mentioned flavors), but they lack any classification capability[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. A major challenge for predicting these flavor notes is the class imbalance, as the number of positive examples is significantly lower than the negatives[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Even though some class balancing methods are available (e.g., over-sampling and under-sampling)[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], it is unclear what approach works better for flavor prediction.\u003c/p\u003e \u003cp\u003eThe most widespread approaches for flavor prediction are based on Quantitative Structure-Activity Relationship (QSAR) models, which have been extensively used in drug discovery[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. These models correlate the chemical structure and properties of molecules with their biological activity, including flavor. The main ML algorithms used have been Random Forest, Support Vector Machines, K-Nearest Neighbors, Deep Neural Networks[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], and more recently Graph Neural Networks[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. These models use mathematical features that capture molecular properties, such as atomic composition, connectivity, and physicochemical properties[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Molecular descriptors followed by molecular fingerprints are the most widespread mathematical representation of molecular structure to train flavor predictors[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Recently, molecular graph representations were also implemented for flavor prediction[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. However, to date, no study comparing the performance of these mathematical representations has been reported.\u003c/p\u003e \u003cp\u003eIn response to the identified gaps, the present study aimed to develop FlavorMiner, a flexible and retrainable flavor ML predictor for seven flavor categories critical for quality control within the food industry (i.e., floral, fruity, sour, sweet, bitter, off-flavor, and nutty). This research comprises an exploration of the performance of different mathematical representations of molecular structures and classification algorithms for flavor prediction. Demonstrating its prowess, FlavorMiner was applied successfully to cocoa metabolomics data, showcasing its ability to unlock valuable insights from food metabolomics data. It can be utilized for flavor analysis in various food products, thanks to a diverse training dataset encompassing over 934 different food products.\u003c/p\u003e"},{"header":"RESULTS AND DISCUSSION","content":"\u003cp\u003eIn this work we introduce the FlavorMiner algorithm, which takes as input the Isomeric Smiles of a set of molecules and produces as output their flavor profile (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The first step is to query a database of 13,387 molecules with known flavor profiles. Only the set of molecules with no database match pass to the prediction step. Then, the respective mathematical representation of the molecules is generated. In the next step, this mathematical representation is fed to seven independent binary classifiers. The average prediction capability of these predictors is 0.88 (ROC AUC Score). Each classifier predicts one of the seven target flavor categories (bitter, floral, fruity, off-flavor, nutty, sour, and sweet). The results are provided in a table, including the predicted flavor profile for each compound and the source of the flavor profile (database match or prediction). The probability values are also provided, indicating the confidence level of each prediction. Finally, a radar chart showing the recurrence of the molecules with each target flavor is also generated.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Development of ML models for flavor prediction, including management of class imbalance biases\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo train the classifiers incorporated in FlavorMiner, a flavor molecule dataset was assembled, containing 13,387 compounds with experimentally validated flavor profiles. The positive examples (those with a specific flavor) represent on average 20% of the dataset, while the negative examples (those without a specific flavor) represent 80% (Supplementary Fig.\u0026nbsp;1). This is a sign of class imbalance, which is an important challenge in the development of ML models, as it can lead to bias towards the majority class[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Due to this class imbalance, all algorithms trained on the original data, except for those trained on sweet molecules, had poor recall (Fig.\u0026nbsp;2), which measures the ability to correctly identify positive examples[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. On the other hand, the specificity, which measures the ability to correctly identify negative examples[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], was significantly high. This bias towards the majority class was consistently observed regardless of the target flavor, algorithm, or mathematical representation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFigure 2.\u003c/b\u003e Classification Metrics for Algorithms Trained with Original Descriptor Data on Test Set. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. \u003cb\u003e(a)\u003c/b\u003e Random Forest trained with molecular descriptors. \u003cb\u003e(b)\u003c/b\u003e Random Forest trained with extended connectivity fingerprint. \u003cb\u003e(c)\u003c/b\u003e K-Nearest Neighbors trained with molecular descriptors. \u003cb\u003e(d)\u003c/b\u003e K-Nearest Neighbors trained with extended connectivity fingerprint. \u003cb\u003e(e)\u003c/b\u003e Convolutional Graph Neural Network trained with molecular graph.\u003c/p\u003e \u003cp\u003eMost algorithms trained on the original data showed a specificity higher than 0.9 during the test. Nonetheless, these models had a recall lower than 0.5. This performance evidence a bias towards the majority class higher than 40% for most algorithms. The Convolutional Graph Neural Network trained with the original molecular graph had the lowest recall for most target flavors (close to zero) (Fig.\u0026nbsp;2c). This is likely because this algorithm is more complex (larger number of parameters) and hence it requires more data to be trained effectively[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. On the other hand, the sweet category has a bias of less than 10% with Random Forest and K-Nearest Neighbors, trained either with RDKit descriptors or ECFP, which can be explained because it had the smallest class imbalance. The number of sweet positives is only 2% lower than the negative examples. Conversely, the sour category has a class imbalance of 97% and showed the highest bias towards the majority class (\u0026gt;\u0026thinsp;85%). This is a common issue in ML models dealing with imbalanced data[\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAdditionally, overfitting was observed in models using Random Forest and K-Nearest Neighbors after the first training iteration with the original data. This is likely due to the limited number of positive data, which can lead the model to exceedingly consider specific features from the negative training set that constrain the ability to adequately generalize and predict when presented with previously unseen data[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. The difference between the train and test specificity was under 10% in most cases, but the recall showed a considerable drop between 20\u0026ndash;90% from train to test for most algorithms trained with the original data (Supplementary Fig.\u0026nbsp;2). This is a clear sign of overfitting[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Only some models obtained with the Convolutional Graph Neural Network showed no sign of overfitting. This is only because the recall was near to zero both during training and testing. Similarly, there was a proportional relationship between class imbalance and overfitting. For example, the models for the sweet flavor (the class with no imbalance) showed the lowest overfitting, while the sour flavor (the class with a high imbalance and special problems associated with its perception, v.i.) had the highest overfitting percentage.\u003c/p\u003e \u003cp\u003eSMOTE and Cluster Centroid sampling techniques were implemented to address the class imbalance. These strategies significantly reduced bias and overfitting. SMOTE, an oversampling technique previously used in flavor predictors[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], was applied to the minority class to increase the number of positive examples. This resulted in a bias of less than 20% for most algorithms (Supplementary Fig.\u0026nbsp;3). The overfitting level was also reduced to less than 30% for most algorithms (Supplementary Fig.\u0026nbsp;4). Under-sampling with Cluster Centroid [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] was also applied to reduce the number of negative examples (Supplementary Fig.\u0026nbsp;5). This resulted in an overfitting reduction to less than 30% for K-Nearest Neighbors models and less than 15% for Random Forest models (Supplementary Fig.\u0026nbsp;6). Most K-Nearest Neighbors models had a bias of less than 10%, while most Random Forest models had a bias of over 20%.\u003c/p\u003e \u003cp\u003eBias and overfitting were reduced due to a significant increase in the recall after applying the resampling strategies. Although the bias and overfitting were still slightly high, this represented a significant improvement compared to the performance with the original data. The recall of all the algorithms trained with resampled data was over 50%, while the specificity of most of these models remained above 70%. Multiple studies have shown that both oversampling and undersampling can be used to correct the problems caused by class imbalance in machine learning approaches[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. In the context of flavor prediction, several studies have investigated the effect of SMOTE oversampling[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. These studies have focused mainly on sweet, bitter, and sour flavors, and have obtained results similar to those of the present work[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOn the other hand, using a balancing transformer on the molecular graph to train a convolutional graph neural network significantly improved the recall, but also significantly reduced the specificity. The recall for classes with more class imbalance improved by 73\u0026ndash;99%, but the specificity dropped by a similar proportion (Supplementary Fig.\u0026nbsp;7). Additionally, the recall for classes such as sweet and bitter decreased. Consequently, the bias and overfitting increased for all models trained with the balanced molecular graph compared to the original data. The bias was higher than 50% for most target flavors and was as high as 90% for fruity, off-flavor, nutty, and sour flavors. This indicates that the balancing transformer had a significant negative effect on the specificity of the models. The overfitting for bitterness and sweetness predictions increased with the balanced data. For fruity, off-flavor, nutty, and sour, the recall decreased to negative values by more than 20%. A negative recall change value indicates underfitting, which occurs when the model does not learn a strong enough pattern from the training data[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. This can be solved by performing a more intense hyperparameter optimization, but this may come at a considerable computational cost compared to Random Forest and K-Nearest Neighbors algorithms.\u003c/p\u003e \u003cp\u003eThe balancing transformer and resampling techniques (SMOTE and cluster centroid) differ in how they address class imbalance. The balancing transformer focuses on the weights of positive and negative examples in the neural network, while resampling techniques focus on the feature space[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. The balancing transformer does not change the input data or the number of examples in each class[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. The poor results obtained with this strategy demonstrate that this is insufficient to solve the severe class imbalance of the input data. Resampling techniques, on the other hand, change the input data by creating new synthetic examples in the minority class (SMOTE) or by removing examples from the majority class and replacing them with cluster centroids[\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. Considering the significant improvement in the performance of the algorithms trained with resampling techniques, this seems to be the best approach to balance the flavor compound database. Unfortunately, it is challenging to implement resampling strategies on molecular graphs, and only possible with molecular descriptors and fingerprints. This is because clustering molecular graphs without affecting their structure and losing valuable information is nearly impossible. Also, in flavor studies, minor changes in structure (graphs) can cause severe changes in perception. Thus, synthetic filling can cause more rather than fewer problems. Although other balance methods are available for graph data, their usefulness with molecular graphs remains to be evaluated[\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. FlavorMiner combines the best ML models for prediction of different flavor classes.\u003c/h2\u003e \u003cp\u003eRandom Forest outperformed the K-Nearest Neighbors algorithm for most target flavor notes, except sour (see below for discussion). Random Forest trained with ECFP oversampled with SMOTE performed best for bitter, fruity, sweet, and off-flavor notes. Random Forest trained with RDKit descriptors performed best for floral and nutty notes. K-Nearest Neighbors trained with ECFP oversampled with SMOTE performed best for sour notes. In general, K-Nearest Neighbors had similar recall to Random Forest with the same input datasets, but slightly lower specificity. Also, algorithms trained with data resampled with the cluster centroid algorithm had slightly better recall, but a higher drop in specificity compared to datasets resampled with SMOTE. These results are consistent with previous studies, which found that Random Forest outperforms other algorithms for predicting sweet and bitter flavors[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. A correlation was observed between the amount of positive data available and the performance of the classifiers. Sweet, the class with the highest number of positive instances, had the best overall performance, with an ROC AUC score of 0.97. Sour, the class with the lowest number of positive instances, had the lowest performance, with a ROC AUC score of 0.78. These results suggest a superior performance of algorithms trained with resampled datasets compared to those trained with the original data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe performance of the seven final predictors selected for the FlavorMiner backbone is shown in Fig.\u0026nbsp;3. The average ROC score, specificity, and recall of these classifiers were 0.88, 0.82, and 0.77, respectively. The performance of FlavorMiner for bitter and sweet prediction was comparable to that of existing predictors[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. For fruity and floral prediction, FlavorMiner achieved recalls of 0.71 and 0.76, respectively, representing an improvement of over 50% compared to previous studies[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. FlavorMiner is the first model to predict nutty and off-flavor notes from molecular structures. For sour prediction, FlavorMiner was outperformed by a previously published tool[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] by about 15%. However, the dataset, composition of positive and negative examples, and code used in this study are not publicly available, making it difficult to assess the reasons for the observed difference.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFigure 3.\u003c/b\u003e Performance of the Optimized Classifiers for Target Flavor Notes in FlavorMiner. The metrics include Recall (blue bar), Specificity (orange bar), and ROC AUC Score (green bar) for each algorithm. \u003cb\u003e(a)\u003c/b\u003e Classification metrics obtained during training using 5-fold cross-validation. \u003cb\u003e(b)\u003c/b\u003e Classification metrics obtained using the test set. Random Forest was used for bitter, fruity, sweet, off-flavor, floral and nutty. K-Nearest Neighbors for sour notes.\u003c/p\u003e \u003cp\u003eVariable Importance Plots (VIP) scores[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] revealed the most important features for predicting floral, off-flavor, and nutty notes (Supplementary Fig.\u0026nbsp;8) with RDKit molecular descriptors. Six descriptors appeared repeatedly in all three cases, accounting for around 45% of the classification. These descriptors measure properties such as the size and polarity of molecules (TPSA), their electronic structure (PEOE_VSA and EState_VSA) and stability (SMR_VSA1 and MinEStateIndex), and their tendency to partition into a hydrophobic environment (MolLogP). Supplementary Fig.\u0026nbsp;9 shows the trend of the five most relevant features for positive and negative examples of each flavor note. Off-flavor molecules tend to be smaller and less polar than non-off-flavor molecules, with a higher tendency to partition into hydrophobic environments. Floral molecules tend to be smaller and more flexible than non-floral molecules, with a higher tendency to partition into hydrophilic environments. Finally, nutty molecules tend to be smaller and less flexible than non-nutty molecules, with a higher electronic stability. These results are new for these flavor notes and provide a basis for future research to select more specific mathematical representations and use data mining techniques to better understand why molecules have these flavors.\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;10 shows the VIP scores for the Random Forest models trained on oversampled ECFP descriptors for predicting bitterness, fruitiness, and sweetness. The four most important bits for the binary classifiers predicting these flavor notes were 897, 314, 489, and 463. The fragments corresponding to these bits are shown in Supplementary Fig.\u0026nbsp;4. For the K-Nearest Neighbors algorithm, the permutation importance score[\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] was used to estimate feature importance because in this case it is not possible to use the VIP score (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Interestingly, most of the top five fingerprints for these notes corresponded to fragments that were absent in the positive compounds. This is likely due to the higher chemical diversity of the negative compounds. E.g. many typical bitter compounds contain an (alkaloid) nitrogen, but no N-containing fragment appeared in the top 5 for bitter, actually many top fragments like bit 897 (C-O-C - moiety) appeared in bitter, fruity and sweet, i.e. they are of universal flavor relevance, but for note determination probably play there role obviously only in the context with other features (e.g. in esters for fruity or sweet in cyclic sugars). Even though resampling strategies were implemented to improve the overall performance of the models, this did not necessarily enhance the chemical diversity of the positive examples.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFingerprints have two main advantages over molecular descriptors. First, they can provide information about the structural features that lead an algorithm to a certain decision. Future work could involve a deeper analysis of the fragments that play a central role in the classification to better understand the structural features that underlie these flavor notes. Second, they can be calculated from MS-spectra data, even when the structure of a compound is not fully elucidated[\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. This makes fingerprints useful for accelerating the discovery of new flavor molecules in metabolomics experiments. Metabolomics experiments typically involve the analysis of many compounds and ECFP can help concentrate the annotation and structural elucidation on the most promising candidates. This can save time and money by focusing efforts on the most promising and likely flavor-active compounds.\u003c/p\u003e \u003cp\u003eThe CGNModel combined with molecular graphs showed poor performance, even with a balancing transformer. This is likely due to the inherent noisiness of the data, which is exacerbated by the susceptibility of Graph Neural Networks to noisy data[\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. This noisiness arises from the heavy dependence of flavor characterization on human tasters and the influence of genetic, sensory, and environmental factors on flavor perception[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. It is challenging to implement a denoising strategy without losing valuable information. Therefore, the CGNModel was discarded for FlavorMiner, given the limitations of the current data and the better performance of other algorithms.\u003c/p\u003e \u003cp\u003eThe flavor profile of a molecule also depends on its concentration and the surrounding matrix[\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. This is related to the concept of flavor threshold and the synergistic and antagonistic effects of flavor molecules in complex mixtures. The flavor threshold is the minimum concentration at which the flavor is detectable[\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. This version of FlavorMiner only performs binary prediction, and intensity data is not yet incorporated. Although some data is available, it is not readily accessible, as there is no standardized database of threshold concentrations for molecules with known flavor profiles. Some databases such as FlavorDB[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] and LSB@TUM Odorant Database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.leibniz-lsb.de/en/databases/leibniz-lsbtum-odorant-database/start/\u003c/span\u003e\u003cspan address=\"https://www.leibniz-lsb.de/en/databases/leibniz-lsbtum-odorant-database/start/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) contain information on flavor thresholds. However, there is a lack of standardization in the thresholds reported in these databases. This means that a method is needed to unify and make this data comparable. Also, most information on flavor thresholds is available in unstructured format (text). Therefore, an intense text mining process is required to extract this data and make it usable for machine learning purposes.\u003c/p\u003e \u003cp\u003eAdditionally, some studies have shown that combining several molecules with different flavor profiles can enhance the flavor profile of a mixture or block certain notes[\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. However, data in this area is limited, and any effort in this direction will require a preliminary experimental process to generate it. Overcoming these challenges could lead to the development of regression algorithms that can be combined with flavor classifiers to predict not only the flavor profile of a molecule but also its threshold concentration and matrix effect.\u003c/p\u003e \u003cp\u003eSour (like salty not evaluated here) is a special flavor note, as it relies on the smallest available \u0026ldquo;molecule\u0026rdquo;, the proton. Also, it is not activating a classical GPCR like the other taste (T1R and T2R) or the olfactory receptors. Only quite recently the responsible Otop1 ion channels were assigned[\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. Thus, typical structural features of a molecule might be considered irrelevant, except for its pKa properties, i.e. its ability to lower pH, an effect that will strongly depend on the matrix\u0026rsquo; overall pH, buffer capacity and may be the proton relay/ion transport capacity. Thus, predicting sour taste from structure might be considered impossible if only the pH change is sensed. However, like GPCRs, ion channels can be influenced by more than the ion it is selective for for various reasons, including ion pairing and matrix/mucosa effects or directly at the ion channel by secondary interactions and additional binding sites which will have selective structural preferences as every protein does. In conclusion, structure-based predictions for ion channel-based tastes (here sour, but also salty) have to be considered with caution, as slight changes in the tasting parameters, e.g. of the matrix (pH, buffer capacity) can thwart results and thus all ML. To understand, if there is sour taste influence on the anionic, organic (i.e. structurally influenced) part, such taste experiments must run with a standardized, high-capacity buffered matrix, neutral pH or better at 2\u0026ndash;3 different pH values. Only this can reveal any possible structural influences of the organic counterion or a neutral molecule influencing or mimicking sour taste. Otherwise, it will not be better than a standard pKa prediction which does not require ML. Independent of this, perception is also influenced by the other receptors. A classic example is of course the action of Miraculin.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Molecular flavor prediction for compounds involved in the processing of cocoa.\u003c/h2\u003e \u003cp\u003ePrevious studies have annotated around 210 compounds during the fermentation, drying, and roasting of fine-flavor cocoa[\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. However, for less than half of these compounds a flavor profile has been reported. The existing data were analyzed with FlavorMiner to predict the flavor profile of these compounds. After the prediction, the compounds with \u0026ldquo;known\u0026rdquo; flavor profiles increased to 92%. The newly predicted compounds include 12 floral, 8 fruity, and 4 compounds with unknown fine-flavor attributes that are potentially linked to positive impacts on quality and price. Additionally, 2 compounds linked to off-flavors and 27 unknown potentially sweet compounds were suggested by the model. These predictions represent an important step forward in closing the gap between cocoa metabolic fingerprint variation during processing and flavor quality.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the frequency of compounds increasing in association with each of the seven target flavors at the end of every cocoa processing stage (Fermentation, Drying, and Roasting). In general, the frequency of compounds for the different target flavors is similar during fermentation and drying. The most relevant change through the processing chain is in sweet compounds, which decrease considerably during the process. This drop is associated with a decrease in the carbohydrate content during the processing chain[\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e], as most of these molecules are reported as sweet agents. In the roasted samples some compounds linked to sour and bitter showed a higher abundance, but the real impact of these suggested flavor molecules still needs to be elucidated. For example, some degradation products of more complex compounds have a lower biological activity (e.i., antioxidant activity) than their precursors[\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e]. If a similar trend occurs with respect to flavor will require further investigations. In contrast, most compounds linked to fine flavor notes (fruity, floral, and nutty) show a relatively constant frequency throughout the cocoa processing chain. These results provide further suggestions into flavor development from biochemistry to processing, which was a missing component until now.\u003c/p\u003e \u003c/div\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eThe present work led to the development of FlavorMiner, an open-access flavor ML predictor for fruity, floral, off-flavor, nutty, sour, sweet, and bitter notes. Access sites are: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/ipb-halle/FlavorMiner\u003c/span\u003e\u003cspan address=\"https://github.com/ipb-halle/FlavorMiner\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. This fills a gap in the current landscape of flavor prediction tools, as there are few open-access options available for predicting these specific flavor notes. FlavorMiner has an average ROC AUC score of 0.88. This algorithm performs similarly to other bitterness and sweetness predictors but outperforms existing floral and fruity predictors. It is also the first tool to predict nutty and off-flavor notes. This work also showed the superiority of RDKit descriptors and ECFP over molecular graphs (combined with Convolutional Graph Neural Network) as mathematical representations to predict flavor notes from molecular structures. Molecular graphs had a recall lower than 0.5 in most cases, even after balancing the classes with a transformer. This means that they currently fail to accurately identify molecular structures with the target flavors. On the other hand, models trained with RDKit descriptors and ECFP had a recall over 0.6 in most cases, especially after resampling with the SMOTE algorithm.\u003c/p\u003e \u003cp\u003eThe best performers were selected to assemble the backbone of FlavorMiner, reaching a recall of over 0.7 in most cases. Further analyses of the best performers revealed the following properties that can define a molecule as positive or negative for floral, off-flavor, and nutty notes such as the size, polarity, electronic structure, stability, and their tendency to partition into a hydrophobic environment. Similarly, this analysis revealed the most critical fragments to define as positive or negative in a molecule for bitter, fruity, sour, and sweet, whereby predictions for sour must be considered with special caution. These results offer a solid base for future studies aiming to dissect the chemistry of flavor properties.\u003c/p\u003e \u003cp\u003eUsing FlavorMiner with cocoa metabolomics data highlighted its potential to elucidate the molecular basis of flavor development in processed food products. To gain further insights into the chemistry and pharmacology of flavor, additional functionalities such as flavor threshold and interaction prediction should be included. However, several challenges must be overcome, starting with (unified) data availability. It is also important to note that machine learning techniques should complement traditional experimental methods in flavor prediction. Integrating the strengths of both approaches can enhance the overall accuracy and efficiency of flavor prediction and lead to new discoveries in the field.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e5.1. Data collection and preprocessing\u003c/h2\u003e \u003cp\u003eThe data for this study were collected from multiple sources, including FlavorDB, Sigma-Aldrich catalog, BitterSweet, ChemTastesDB, Flavornet, Off-flavor, AromaDB, OlfactionBase, and Natural_tAS2r_agonists[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. The data were cleaned by removing duplicate entries, molecules without reported flavor profiles, invalid PubChem identifiers or names, molecules with fewer than 2 heavy atoms, and salts, knowing that especially the latter fact may depent on representation (in ionic or non-ionic form) of some molecules, which may also have contributed to the problems with the sour set. The linear structures of the remaining molecules were retrieved from the PubChem database using the Python library PubChemPy. The compounds were labeled with the flavor information retrieved from the databases. The flavor notes were grouped into seven categories: bitter, floral, fruity, off-flavor, nutty, sour, and sweet. The labeling process was automated in a Python script and a flavor wheel[\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e] to facilitate future relabeling or adding new data. The seven target labels were then converted into binary values using the One Hot Encoding method[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. This information was stored in an Excel file containing the compound name, PubChem ID, flavor profile, isomeric smiles, data source, and labels (Supplementary File 1).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e5.2. Mathematical representation of molecular structures\u003c/h2\u003e \u003cp\u003eThree mathematical representations of the molecular structures were generated. First, 200 molecular descriptors were calculated using the RDKit library. Descriptors with invalid data, more than 97% of unique values, and descriptors highly correlated were removed. For this, the Pearson correlation index was used with a threshold of 0.95[\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]. Second, an Extended Connectivity Fingerprint (ECFP) with a radius of 2 was generated for all compounds using the RDKit library. Fingerprints with more than 97% unique values were also removed. The datasets corresponding to the RDKit molecular descriptors and ECFP were split into training and testing sets using a random partition of 80:20. A third mathematical representation was generated by creating molecular graphs from the molecular structures using the DeepChem library. The DeepChem MolGraphConvFeaturizer was used to generate the molecular graphs. The graphs were then labeled for each target flavor with a binary label and converted into NumPy datasets. This dataset was divided into a training, validation, and testing set using a random partition of 70:10:20.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e5.3. Machine Learning algorithms training, optimization, and testing\u003c/h2\u003e \u003cp\u003eIndependent binary classifiers were used to predict each flavor category. Initially, a Random Forest and K-Nearest Neighbors algorithm were trained on the RDKit molecular descriptors and ECFP datasets using scikit-learn. Hyperparameter optimization was performed using the Grid Search method and 5-fold cross-validation. Class imbalance was addressed by oversampling the minority classes using SMOTE (Synthetic Minority Oversampling Technique) and undersampling the majority classes using the cluster centroid of a KMeans algorithm. A Graph Convolutional Neural Network (GCNModel) was then trained on the molecular graphs using DeepChem. Hyperparameter optimization of the GCNModel was performed using the hyperopt library. The balance of the molecular graph data was done using the DeepChem balancing transformer. The performance of the algorithms was evaluated using the recall, specificity and Receiving Operator Characteristic (ROC) curves[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e5.4. Using cocoa processing metabolomics data as a case study\u003c/h2\u003e \u003cp\u003eTwo previous datasets from metabolomics studies of fine-flavor cocoa processing were used for this case study[\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. The first study used LC-QTOF-MS and GC-QTOF-MS to analyze the fermentation of fine-flavored cocoa[\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e]. The second study used the same analytical platforms to assess the evolution of flavor during post-fermentation processing (drying and roasting) of fine-flavor cocoa[\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. The PubChem ID for each molecule was obtained from the PubChem database. This data was stored in an Excel file and fed to FlavorMiner to assign flavor profiles.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll the datasets, and scripts used in this work are publicly available in a GitHub repository (https://github.com/ipb-halle/FlavorMiner). FlavorMiner and all Python scripts are available in Jupyter Notebooks. These Jupyter Notebooks contain a detailed explanation of the code and the data, which is integrated with the GitHub repository, reducing the need to download files locally to run the scripts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this manuscript.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors thank the program Impacto Pa\u0026iacute;s of the Vice-presidency of Research and Creation of the University of Los Andes and the Jeff Schell Fellowship of the Bayer Foundation, for the funding provided to Fabio Herrera-Rocha. IATA-CSIC received funding from the Spanish government, ref. MCIN/AEI/10.13039/501100011033, as a \u0026lsquo;Severo Ochoa\u0026rsquo; Center of Excellence (CEX2021-001189-S). Finally, the authors acknowledge the Alexander von Humboldt Foundation of Germany for the financial support to Miguel Angel Fern\u0026aacute;ndez-Ni\u0026ntilde;o and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - SPP2363 for funding provided to Mehdi D. Davari.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFHR: manuscript writing, code development, repository construction, investigation, validation, benchmarking, and machine learning. JD: conceptualization, manuscript writing, code development, validation, and supervision. MFN: manuscript writing, conceptualization. MPC: conceptualization, manuscript writing, validation, and supervision. MJC: conceptualization, manuscript writing, validation, and supervision. \u0026nbsp;LAW: conceptualization, manuscript writing, validation, and supervision. MD: conceptualization, manuscript writing, validation, and supervision. AFGB: conceptualization, manuscript writing, validation, and supervision. All authors have reviewed and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to acknowledge the Group of Product and Process Design (GDPP), in the Department of Chemical and Food Engineering of the University of Los Andes, and the Department of Bioorganic Chemistry for their valuable insights and thoughtful recommendations. Finally, the authors thank Exacore of the University of Los Andes for the High-Performance Computing (HPC) services provided to this project. \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMouritsen OG (2015) The science of taste. Flavour 4:10\u0026ndash;11. https://doi.org/10.1186/s13411-014-0028-3\u003c/li\u003e\n\u003cli\u003eWang YT, Yang ZX, Piao ZH, et al (2021) Prediction of flavor and retention index for compounds in beer depending on molecular structure using a machine learning method. RSC Adv 11:36942\u0026ndash;36950. https://doi.org/10.1039/d1ra06551c\u003c/li\u003e\n\u003cli\u003eKeller A, Gerkin RC, Guan Y, et al (2017) Predicting human olfactory perception from chemical features of odor molecules. Science (80- ) 355:820\u0026ndash;826. https://doi.org/10.1126/science.aal2014\u003c/li\u003e\n\u003cli\u003eQueiroz LP, Rebello CM, Costa EA, et al (2023) Generating Flavor Molecules Using Scientific Machine Learning. ACS Omega 10875\u0026ndash;108887. https://doi.org/10.1021/acsomega.2c07176\u003c/li\u003e\n\u003cli\u003eAnastasiadi M, Mohareb F, Redfern SP, et al (2017) Biochemical Profile of Heritage and Modern Apple Cultivars and Application of Machine Learning Methods to Predict Usage, Age, and Harvest Season. J Agric Food Chem 65:5339\u0026ndash;5356. https://doi.org/10.1021/acs.jafc.7b00500\u003c/li\u003e\n\u003cli\u003eGarg N, Sethupathy A, Tuwani R, et al (2018) FlavorDB: A database of flavor molecules. Nucleic Acids Res 46:1210\u0026ndash;1216. https://doi.org/10.1093/nar/gkx957\u003c/li\u003e\n\u003cli\u003eGrover N, Goel M, Batra D, et al (2022) FlavorDB2: An Updated Database of Flavor Molecules. arXiv 1\u0026ndash;5\u003c/li\u003e\n\u003cli\u003eNaveja JJ, Rico-Hidalgo MP, Medina-Franco JL (2018) Analysis of a large food chemical database: chemical space, diversity, and complexity. F1000Research 7:993. https://doi.org/10.12688/F1000RESEARCH.15440.2\u003c/li\u003e\n\u003cli\u003eYu M, Yang P, Song H, Guan X (2022) Research progress in comprehensive two-dimensional gas chromatography-mass spectrometry and its combination with olfactometry systems in the flavor analysis field. J Food Compos Anal 114:104790. https://doi.org/10.1016/j.jfca.2022.104790\u003c/li\u003e\n\u003cli\u003eLi Y, Yuan L, Liu H, et al (2022) Analysis of the changes of volatile flavor compounds in a traditional Chinese shrimp paste during fermentation based on electronic nose, SPME-GC-MS and HS-GC-IMS. Food Sci Hum Wellness 12:173\u0026ndash;182. https://doi.org/10.1016/j.fshw.2022.07.035\u003c/li\u003e\n\u003cli\u003eDegenhardt A, Wittlake R, Seilwind S, et al (2014) Quantification of Important Flavor Compounds in Beef Stocks and Correlation to Sensory Results by \u0026ldquo;Reverse Metabolomics.\u0026rdquo; Elsevier Inc.\u003c/li\u003e\n\u003cli\u003eWardencki W, Chmiel T, Dymerski T (2013) Gas chromatography-olfactometry (GC-O), electronic noses (e-noses) and electronic tongues (e-tongues) for in vivo food flavour measurement. Woodhead Publishing Limited\u003c/li\u003e\n\u003cli\u003eYimenu SM, Kim JY, Kim BS (2017) Prediction of egg freshness during storage using electronic nose. Poult Sci 96:3733\u0026ndash;3746. https://doi.org/10.3382/ps/pex193\u003c/li\u003e\n\u003cli\u003eRottiers H, Tzompa Sosa DA, Van de Vyver L, et al (2019) Discrimination of Cocoa Liquors Based on Their Odor Fingerprint: a Fast GC Electronic Nose Suitability Study. Food Anal Methods 12:475\u0026ndash;488. https://doi.org/10.1007/s12161-018-1379-7\u003c/li\u003e\n\u003cli\u003eTseng YJ, Chuang PJ, Appell M (2023) When Machine Learning and Deep Learning Come to the Big Data in Food Chemistry. ACS Omega 15854\u0026ndash;15864. https://doi.org/10.1021/acsomega.2c07722\u003c/li\u003e\n\u003cli\u003eKou X, Shi P, Gao C, et al (2023) Data-Driven Elucidation of Flavor Chemistry. J Agric Food Chem 6789\u0026ndash;6802. https://doi.org/10.1021/acs.jafc.3c00909\u003c/li\u003e\n\u003cli\u003eQueiroz LP, Rebello CM, Costa EA, et al (2023) Transfer Learning Approach to Develop Natural Molecules with Specific Flavor Requirements. 9062\u0026ndash;9076. https://doi.org/10.1021/acs.iecr.3c00722\u003c/li\u003e\n\u003cli\u003eSanchez-Lengeling B, Wei JN, Lee BK, et al (2019) Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules. arXiv\u003c/li\u003e\n\u003cli\u003eDutta P, Jain D, Gupta R, Rai B (2022) Classification of Tastants: A Deep Learning Based Approach. ChemRxiv\u003c/li\u003e\n\u003cli\u003eRojas C, Ballabio D, Consonni V, et al (2023) Classification-based machine learning approaches to predict the taste of molecules: A review. Food Res Int 171:. https://doi.org/10.1016/j.foodres.2023.113036\u003c/li\u003e\n\u003cli\u003eZheng S, Jiang M, Zhao C, et al (2018) e-Bitter: Bitterant prediction by the consensus voting from the machine-learning methods. Front Chem 6:1\u0026ndash;18. https://doi.org/10.3389/fchem.2018.00082\u003c/li\u003e\n\u003cli\u003eTuwani R, Wadhwa S, Bagler G (2019) BitterSweet: Building machine learning models for predicting the bitter and sweet taste of small molecules. Sci Rep 9:1\u0026ndash;13. https://doi.org/10.1038/s41598-019-43664-y\u003c/li\u003e\n\u003cli\u003eMartinez-Mayorga K, Peppard TL, L\u0026oacute;pez-Vallejo F, et al (2013) Systematic mining of generally recognized as safe (GRAS) flavor chemicals for bioactive compounds. J Agric Food Chem 61:7507\u0026ndash;7514. https://doi.org/10.1021/jf401019b\u003c/li\u003e\n\u003cli\u003eDagan-Wiener A, Nissim I, Ben Abu N, et al (2017) Bitter or not? BitterPredict, a tool for predicting taste from chemical structure. Sci Rep 7:1\u0026ndash;13. https://doi.org/10.1038/s41598-017-12359-7\u003c/li\u003e\n\u003cli\u003eFritz F, Preissner R, Banerjee P (2021) VirtualTaste: A web server for the prediction of organoleptic properties of chemical compounds. Nucleic Acids Res 49:W679\u0026ndash;W684. https://doi.org/10.1093/nar/gkab292\u003c/li\u003e\n\u003cli\u003eBanerjee P, Preissner R (2018) Bitter sweet forest: A Random Forest based binary classifier to predict bitterness and sweetness of chemical compounds. Front Chem 6:1\u0026ndash;10. https://doi.org/10.3389/fchem.2018.00093\u003c/li\u003e\n\u003cli\u003eNozaki Y, Nakamoto T (2018) Correction: Predictive modeling for odor character of a chemical using machine learning combined with natural language processing (PLoS ONE (2018) 13, 6 (e0198475) DOI: 10.1371/journal.pone.0198475). PLoS One 13:1\u0026ndash;13. https://doi.org/10.1371/journal.pone.0208962\u003c/li\u003e\n\u003cli\u003eShang L, Liu C, Tomiura Y, Hayashi K (2017) Machine-Learning-Based Olfactometer: Prediction of Odor Perception from Physicochemical Features of Odorant Molecules. Anal Chem 89:11999\u0026ndash;12005. https://doi.org/10.1021/acs.analchem.7b02389\u003c/li\u003e\n\u003cli\u003eSharma A, Kumar R, Ranjta S, Varadwaj PK (2021) SMILES to Smell: Decoding the Structure-Odor Relationship of Chemical Compounds Using the Deep Neural Network Approach. J Chem Inf Model 61:676\u0026ndash;688. https://doi.org/10.1021/acs.jcim.0c01288\u003c/li\u003e\n\u003cli\u003eRamsundar B, Eastman P, Walters P, Pande V (2019) Deep Learning for the Life Sciences\u003c/li\u003e\n\u003cli\u003eG\u0026eacute;ron A (2019) Hands-On Machine Learning with Scikit-Learn, Keras \u0026amp; TensorFlow, 2nd Editio\u003c/li\u003e\n\u003cli\u003eLo YC, Rensi SE, Torng W, Altman RB (2018) Machine learning in chemoinformatics and drug discovery. Drug Discov Today 23:1538\u0026ndash;1546. https://doi.org/10.1016/j.drudis.2018.05.010\u003c/li\u003e\n\u003cli\u003eSun Y, Wong AKC, Kamel MS (2009) Classification of imbalanced data: A review. Int J Pattern Recognit Artif Intell 23:687\u0026ndash;719. https://doi.org/10.1142/S0218001409007326\u003c/li\u003e\n\u003cli\u003eBanerjee P, Dehnbostel FO, Preissner R (2018) Prediction is a balancing act: Importance of sampling methods to balance sensitivity and specificity of predictive models based on imbalanced chemical data sets. Front Chem 6:387941. https://doi.org/10.3389/FCHEM.2018.00362/BIBTEX\u003c/li\u003e\n\u003cli\u003eReiser P, Neubert M, Eberhard A, et al (2022) Graph neural networks for materials science and chemistry. Commun Mater 2022 31 3:1\u0026ndash;18. https://doi.org/10.1038/s43246-022-00315-6\u003c/li\u003e\n\u003cli\u003evon Lilienfeld OA, Burke K (2020) Retrospective on a decade of machine learning for chemical discovery. Nat Commun 2020 111 11:1\u0026ndash;4. https://doi.org/10.1038/s41467-020-18556-9\u003c/li\u003e\n\u003cli\u003eLeevy JL, Khoshgoftaar TM, Bauder RA, Seliya N (2018) A survey on addressing high-class imbalance in big data. J Big Data 5:. https://doi.org/10.1186/s40537-018-0151-6\u003c/li\u003e\n\u003cli\u003eMalavolta M, Pallante L, Mavkov B, et al (2022) A survey on computational taste predictors. Eur Food Res Technol 248:2215\u0026ndash;2235. https://doi.org/10.1007/s00217-022-04044-5\u003c/li\u003e\n\u003cli\u003eDe La Fuente A, Zamberlan F, S\u0026aacute;nchez Ferr\u0026aacute;n A, et al (2020) Relationship among subjective responses, flavor, and chemical composition across more than 800 commercial cannabis varieties. J Cannabis Res 2:. https://doi.org/10.1186/S42238-020-00028-Y\u003c/li\u003e\n\u003cli\u003eYen SJ, Lee YS (2009) Cluster-based under-sampling approaches for imbalanced data distributions. Expert Syst Appl 36:5718\u0026ndash;5727. https://doi.org/10.1016/J.ESWA.2008.06.108\u003c/li\u003e\n\u003cli\u003eHu F, Wang L, Liu Q, et al (2022) GraphDIVE: Graph Classification by Mixture of Diverse Experts. IJCAI Int Jt Conf Artif Intell 3:2080\u0026ndash;2086. https://doi.org/10.24963/IJCAI.2022/289\u003c/li\u003e\n\u003cli\u003eSaini K, Ramanathan V (2022) Predicting odor from molecular structure: a multi-label classification approach. Sci Rep 12:1\u0026ndash;11. https://doi.org/10.1038/s41598-022-18086-y\u003c/li\u003e\n\u003cli\u003eAltmann A, Toloşi L, Sander O, Lengauer T (2010) Permutation importance: a corrected feature importance measure. Bioinformatics 26:1340\u0026ndash;1347. https://doi.org/10.1093/BIOINFORMATICS/BTQ134\u003c/li\u003e\n\u003cli\u003eD\u0026uuml;hrkop K, Nothias LF, Fleischauer M, et al (2021) Systematic classification of unknown metabolites using high-resolution fragmentation mass spectra. Nat Biotechnol 39:462\u0026ndash;471. https://doi.org/10.1038/s41587-020-0740-8\u003c/li\u003e\n\u003cli\u003eJi H, Deng H, Lu H, Zhang Z (2020) Predicting a Molecular Fingerprint from an Electron Ionization Mass Spectrum with Deep Neural Networks. Anal Chem 92:8649\u0026ndash;8653. https://doi.org/10.1021/ACS.ANALCHEM.0C01450/SUPPL_FILE/AC0C01450_SI_002.XLS\u003c/li\u003e\n\u003cli\u003eDai E, Jin W, Liu H, Wang S (2022) Towards robust graph neural networks for noisy graphs with sparse labels. WSDM 2022 - Proc 15th ACM Int Conf Web Search Data Min 181\u0026ndash;191. https://doi.org/10.1145/3488560.3498408\u003c/li\u003e\n\u003cli\u003eGodwin J, Schaarschmidt M, Gaunt A, et al (2022) Simple Gnn Regularisation for 3D Molecular Property Prediction \u0026amp; Beyond. ICLR 2022 - 10th Int Conf Learn Represent 1\u0026ndash;23\u003c/li\u003e\n\u003cli\u003ePoivet E, Tahirova N, Peterlin Z, et al (2018) Functional odor classification through a medicinal chemistry approach. Sci Adv 4:. https://doi.org/10.1126/sciadv.aao6086\u003c/li\u003e\n\u003cli\u003eKhan B (2018) Application of pattern mining on data of flavor molecules , their percepts and molecular features\u003c/li\u003e\n\u003cli\u003eMojet J, Heidema J, Christ-Hazelhof E (2004) Effect of concentration on taste-taste interactions in foods for elderly and young subjects. Chem Senses 29:671\u0026ndash;681. https://doi.org/10.1093/CHEMSE/BJH070\u003c/li\u003e\n\u003cli\u003eMarks LE, Shepard TG, Burger K, Chakwin EM (2012) Flavor-Intensity Perception: Effects of Stimulus Context. Physiol Behav 105:443. https://doi.org/10.1016/J.PHYSBEH.2011.08.039\u003c/li\u003e\n\u003cli\u003eTurner HN, Liman ER (2022) The Cellular and Molecular Basis of Sour Taste. Annu Rev Physiol 84:41\u0026ndash;58. https://doi.org/10.1146/annurev-physiol-060121-041637\u003c/li\u003e\n\u003cli\u003eHerrera-Rocha F, Cala MP, Mej\u0026iacute;a Aguirre LJ, et al (2021) Dissecting fine-flavor cocoa bean fermentation through metabolomics analysis to break down the current metabolic paradigm. Sci Rep 11:. https://doi.org/10.1038/s41598-021-01427-8\u003c/li\u003e\n\u003cli\u003eHerrera-Rocha F, Le\u0026oacute;n-Inga AM, Aguirre Mej\u0026iacute;a JL, et al (2024) Bioactive and flavor compounds in cocoa liquor and their traceability over the major steps of cocoa post-harvesting processes. Food Chem 435:. https://doi.org/10.1016/j.foodchem.2023.137529\u003c/li\u003e\n\u003cli\u003eMcshea A, Ramiro-Puig E, Munro SB, et al (2008) Clinical benefit and preservation of flavonols in dark chocolate manufacturing. https://doi.org/10.1111/j.1753-4887.2008.00114.x\u003c/li\u003e\n\u003cli\u003eSpencer M, Sage E, Velez M, Guinard JX (2016) Using Single Free Sorting and Multivariate Exploratory Methods to Design a New Coffee Taster\u0026rsquo;s Flavor Wheel. J Food Sci 81:S2997\u0026ndash;S3005. https://doi.org/10.1111/1750-3841.13555\u003c/li\u003e\n\u003cli\u003eR\u0026aacute;cz A, Bajusz D, H\u0026eacute;berger K (2019) Intercorrelation Limits in Molecular Descriptor Preselection for QSAR/QSPR. Mol Inform 38:1\u0026ndash;6. https://doi.org/10.1002/minf.201800154\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-cheminformatics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"chin","sideBox":"Learn more about [Journal of Cheminformatics](https://jcheminf.biomedcentral.com/)","snPcode":"13321","submissionUrl":"https://submission.nature.com/new-submission/13321/3","title":"Journal of Cheminformatics","twitterHandle":"@jcheminf","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Flavor Chemistry, Molecular Machine Learning, Molecular Representation, Deep Learning, Cocoa","lastPublishedDoi":"10.21203/rs.3.rs-4768238/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4768238/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eFlavor is the main factor driving consumers acceptance of food products. However, tracking the biochemistry of flavor is a formidable challenge due to the complexity of food composition. Current methodologies for linking individual molecules to flavor in foods and beverages are expensive and time-consuming. Predictive models based on machine learning (ML) are emerging as an alternative to speed up this process. Nonetheless, the optimal approach to predict flavor features of molecules remains elusive. In this work we present FlavorMiner, an ML-based multilabel flavor predictor. FlavorMiner seamlessly integrates different combinations of algorithms and mathematical representations, augmented with class balance strategies to address the inherent class of the input dataset. Notably, Random Forest and K-Nearest Neighbors combined with Extended Connectivity Fingerprint and RDKit molecular descriptors consistently outperform other combinations in most cases. Resampling strategies surpass weight balance methods in mitigating bias associated with class imbalance. FlavorMiner exhibits remarkable accuracy, with an average ROC AUC score of 0.88. This algorithm was used to analyze cocoa metabolomics data, unveiling its profound potential to help extract valuable insights from intricate food metabolomics data. FlavorMiner can be used for flavor mining in any food product, drawing from a diverse training dataset that spans over 934 distinct food products.\u003c/p\u003e","manuscriptTitle":"FlavorMiner: A Machine Learning Platform for Extracting Molecular Flavor Profiles from Structural Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-08-18 23:33:16","doi":"10.21203/rs.3.rs-4768238/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-07-20T21:11:35+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-07-20T10:23:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-07-20T10:22:29+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Cheminformatics","date":"2024-07-19T14:20:48+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-cheminformatics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"chin","sideBox":"Learn more about [Journal of Cheminformatics](https://jcheminf.biomedcentral.com/)","snPcode":"13321","submissionUrl":"https://submission.nature.com/new-submission/13321/3","title":"Journal of Cheminformatics","twitterHandle":"@jcheminf","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c198f63d-462d-4b74-b9a2-a75f52ce8385","owner":[],"postedDate":"August 18th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2024-12-16T16:03:39+00:00","versionOfRecord":{"articleIdentity":"rs-4768238","link":"https://doi.org/10.1186/s13321-024-00935-9","journal":{"identity":"journal-of-cheminformatics","isVorOnly":false,"title":"Journal of Cheminformatics"},"publishedOn":"2024-12-10 15:58:02","publishedOnDateReadable":"December 10th, 2024"},"versionCreatedAt":"2024-08-18 23:33:16","video":"","vorDoi":"10.1186/s13321-024-00935-9","vorDoiUrl":"https://doi.org/10.1186/s13321-024-00935-9","workflowStages":[]},"version":"v1","identity":"rs-4768238","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4768238","identity":"rs-4768238","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}