Progress 08/01/24 to 07/31/25

Outputs
Target Audience:This report pertains to multiple groups, including researchers in the field of food chemistry, food producers,and general public/consumers. In the first place, the activities reported during the first year of the project have been directed to understanding the molecular interactions behind the differential activity of mixtures of antioxidants, leading to synergistic, additive, or antagonistic behavior. To this end, our project hypothesized that subtle, non-covalent, interactions like hydrogen-bonding are determining the behavior of the mixtures. Changes/Problems:Although we have encountered no major problems during the first year of the project, it is important to note that many of the initial experiments (collected using olive oil) turned out to be unreliable due to large changes in stability between batches. Considering this, we are now using oleic acid as a model fatty acid, enabling the comparison of various batches of experiments. What opportunities for training and professional development has the project provided?The project is partly supportingone graduate student, who has recently advanced to candidacy. We are now completing the required steps to incorporate a recent Clemson graduate, who will assist in the data collection. How have the results been disseminated to communities of interest?We have published two papers, we are working on two additional manuscripts, and attended several conferences. What do you plan to do during the next reporting period to accomplish the goals?We intend to continue with the activities proposed in our application.

Impacts
What was accomplished under these goals? Activities performing under thegoals of the grant can be broadly grouped in three categories: 1- Publications: With the support of the grant, our group was able to complete the publication of two papers (citations provided in products) addressingthe oxidation of lipids in food samples. While most consumer products contain antioxidants, the most efficient strategy is to incorporate combinations of two or more compounds, boosting the total antioxidant capacity. Unfortunately, the reasons for observing synergistic / antagonistic / additive effects in food samples are still unclear, and it is common to observe very different responses even for similar mixtures. Aiming to identify chemical features that can be correlated with specific responses, wereported an analysis of 1243 mixtures of antioxidants reported in the literature, how their chemical structure affects the response, and the determining role of hydrogen bonding. Also following the hypothesis linked to hydrogen bonding, our groupdeveloped and presented thefirst example of a functional deep eutectic solvent, defined as a solvent formed exclusively with molecules with specific functionality (antioxidants, in this case). Specifically, the strategy utilizedtwo machine learning models to sequentially predict which antioxidants are likely to form synergistic mixtures and then predict which mixtures would also form stable deep eutectic solvents (DES). As a proof of concept, we presented results obtained with mixtures of BHA (tert-butyl-4-hydroxyanisole) and BHT (3,5-di-tert-butyl-4-hydroxytoluene), which are not only known to form synergistic mixtures but also identified by our model as top candidates for the proposed strategy. To investigate the structure and interaction energy between the selected compounds, a computational analysis using density functional theory was also implemented. The antioxidant capacity of the formulated DES was then assessed (by the thiobarbituric acid reactive substances assay) using oleic acid as well as commercial samples of olive oil, pork lard, and duck fat. We found that the new mixtures ledto systems that feature a total antioxidant capacity that is superior to both the individual components and the non-DES forms of the same antioxidants. These differences, attributed to significant decreases in the bond dissociation enthalpy of the antioxidants involved in the DES, highlighted the importance of promoting (and preserving) the formation of intermolecular hydrogen bonds between antioxidants to boost their efficiency.We are currently working in one additional manuscript, describing the use of AI to predict the oxidation of fatty acids. 2- New data: As described in the proposal, our group heavily relied on the use of traditional analytical approaches to evaluate antioxidant capacity (TBARS, PV, etc). While potentially useful, these methodologies also present several practical challenges that have limited their adotpion in industrial settings. Aiming to address this need (and to be able to translate the data), we decided to purchase a rancimat (Metrohm 892 Professional Rancimat) and supplement our database with induction times. This parameter is obtained byaccelerating the oxidation process of a sample (raising its temperature and passing a continuous stream of air) leading to the release of volative organic acids and serving as a proxy for the stability of that fat sample. In fact, the induction time is a standard parameter in quality testing of oils and fats in the food industry and gives indications about the remaining shelf life of a product. To ensure adequate reproducibility and comparability of the data, our lab focused on the use of oleic acid as a model fatty acid to evaluate the antioxidant capacity of the investigated samples. In this scenario, we have evaluated the antioxidant capacity of binary mixtures ofBHA, BHT, LG, OG, PG, PHENOL , PY, PROTOCAT, GALLIC ACID, and CANNABIDIOL. These experiments have been performed at 1:1, 1:3, and 3:1 ratios of the corresponding antioxidants and used to calculate a synergism factor (Xsyn), defined as the change in induction time of a sample in the presence of an antioxidant mixture/inductiontime of a sample without antioxidants. Although we currently have almost 200 experimental datapoints (obtained in the lab), we are currently working to translate an existing database (containing CI values) intoXsyn values. Besides these experiments, our team is also working to obtain rancimat data on various mixtures of fatty acids, to challenge a model trained using literature data. 3- New models:Although several studies have attempted to model the chemical relationship between constituent fatty acids and oxidative stability, the developed frameworks rely on simplified empirical models or multivariate-based correlations dependent solely on fatty acid percentage composition or unsaturation metrics.Additional parameters such as iodine value, Allylic Position equivalent (APE), and Bis Allylic Position Equivalent (BAPE) have been employed to measure stability. Knothe and Dunn (2003)applied a simple linear regressor to investigate the effect of iodine values, catalytic metals, APE and BAPE under the AOCS OSI Method. The results highlight BAPE as an essential descriptor with an inverse relationship with oxidative stability, demonstrating an R² value of 0.983. It is, however, pertinent to note the emergence of lipid oxidation analysis based on machine learning (ML) models. Bukkarapu and Krishnasamy (2022)conducted a comprehensive review that collates developed ML models for predicting lipid system properties, especially oxidative stability. It is essential to mention that Suvarna et al. (2022)developed classifier model on fatty acid methyl ester (FAME) profiles to predict sample compliance with ASTM and EN oxidative stability limits. The curated database from the study provided a correlation between fatty acid methyl ester (FAME) profile and oxidative stability metrics. Also, the transesterification process does not alter the fatty acid profile of lipid-based systems; hence, predictive modeling based on FAME composition remains valid across different fatty acid matrices.Nevertheless, percentage composition data does not represent intrinsic chemical reactivity, which determines oxidative stability.In addition to descriptors linked to bond dissociation energy (BDE), it is worth noting that other features, such as polarizability and molecular surface area, significantly affect oxidation kinetics, with extended surface area/ molecule being an indicator of auto-oxidation.Despite the significance of previously developed theoretical and statistical models, they are, however, limited in their percentage composition-only format and inability to capture structural, electronic, and steric factors that affect oxidative mechanisms at the molecular level. Considering the established comprehensive understanding of lipid oxidation and the persistent gap in its application due to difficulties in bridging experimental and theoretical approaches, we developed a novel, descriptor-informed inference model to gain further insight into the oxidative behavior of lipid matrices. To achieve this, we developed and trained three distinct machine learning models: Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost), to leverage the unique interpretability strengths of ML models. The results from this multi-model approach provide a clear understanding and complementary patterns of chemical descriptors that affect oxidative stability.

Publications

• Type: Peer Reviewed Journal Articles Status: Published Year Published: 2025 Citation: Deciphering Antioxidant Interactions via Data Mining and RDKit Lucas B. Ayres, Justin T. Furgala, and Carlos D. Garcia Scientific Reports 15 (2025) 670
• Type: Peer Reviewed Journal Articles Status: Published Year Published: 2025 Citation: Functional Deep Eutectic Solvents to Boost Antioxidant Synergism in Edible Fats Emmanuel Dike, Lucas Ayres, Tomas Benavidez, Jorge Barroso, Vagner dos Santos, and Carlos D. Garcia ACS Sustainable Chemistry & Engineering 13 (2025) 44274438
• Type: Other Status: Published Year Published: 2025 Citation: Cover image for 10.1021/acssuschemeng.4c09183 https://pubs.acs.org/toc/ascecg/13/11#:~:text=Download-,Cover,-In%20this%20issue
• Type: Conference Papers and Presentations Status: Published Year Published: 2025 Citation: Functional DES: Where adventure and imagination meet the final frontier Invited seminar, Department of Chemistry, University of Texas El Paso (El Paso, TX  04/2025)
• Type: Conference Papers and Presentations Status: Published Year Published: 2025 Citation: Big Data for a Deep Problem: Understanding and Predicting the formation of DES via Machine Learning Lucas B. Ayres, Emmanuel Dike, Jorge Barroso, Connor Parker, Vagner B. dos Santos and Carlos D. Garcia 4th International Meeting on Deep Eutectic Solvents (Lisbon, Portugal  06/2025)  Keynote Speaker
• Type: Conference Papers and Presentations Status: Published Year Published: 2025 Citation: When 1+1 ? 2: Deciphering Antioxidant Interactions via Machine Learning Carlos D. Garcia Webinar AI in Oilseeds & Oils Industry (Toulouse, France  05/2025)
• Type: Conference Papers and Presentations Status: Published Year Published: 2025 Citation: Big Data for a Deep Problem: Understanding and Predicting the formation of DES via Machine Learning Emmanuel Dike, Lucas B. Ayres, Jorge Barroso, and Carlos D. Garcia 2025 AOAC Annual Meeting (Portland, OR  04/2025)  Featured Speaker
• Type: Conference Papers and Presentations Status: Published Year Published: 2025 Citation: Functional Deep Eutectic Solvents to Boost Antioxidant Synergism in Edible Fats Emmanuel D. Dike, Lucas B. Ayres, Tomas E. Benavidez, Jorge Barroso, Vagner B. dos Santos and Carlos D. Garcia 10th Annual Chemistry Research Symposium (Clemson, SC - 02/2025)
• Type: Conference Papers and Presentations Status: Accepted Year Published: 2025 Citation: Functional Deep Eutectic Solvents to Boost Antioxidant Synergism in Edible Fats Emmanuel Dike, Lucas Ayres, Tomas Benavidez, Jorge Barroso, Vagner dos Santos, and Carlos D. Garcia ACS Fall Meeting (Washington, DC - 08/2025)