Accelerating Food Safety: AI's Role in Ensuring Safer Additives
In the past century, the food and flavors industry has completely revolutionized the way that we grow, manufacture, store, and consume food. While the addition of herbs and spices to alter the taste of foodstuffs is a human experience that predates remembered history itself, the addition of synthetically produced molecules to fine-tune the qualities of our food is a much newer phenomenon. These ingredients, known as food additives, are extracted from natural sources or chemically synthesized and are used extensively across the food industry to alter food properties such as nutritional value, color, flavor, texture, and shelf-stability. Well known examples include natural and artificial sweeteners like stevia and aspartame to reduce caloric content, dyes like ‘Food Red No. 40’ and ‘Food Yellow No. 5’ that are used to make ‘visually desirable’ foods, or nitrates and nitrites added to preserve, stabilize color, and add flavor to meat products.
Despite widespread utilization of food additives, the safety of many have recently been called into question as mounting scientific evidence connects them to adverse health effects; artificial sweeteners have been implicated in misregulation of metabolism1 and to cause cardiovascular events like heart attack and stroke2; food dyes have been linked to myriad health problems including cancer and developmental problems3; and nitrates and nitrites have been linked to an increased risk of colorectal cancer4. Although many purported additive-disease linkages are still under professional consideration, the existing data and media scrutiny have been enough for the public and regulatory agencies to ask relevant questions and, in some cases, demand removal/replacement of some additives. It is anticipated that additional testing will lead to greater regulatory intervention and the need for ingredient replacement strategies to become a regular part of the product development process.
Unfortunately, identification and replacement of ‘toxic’ additives is no straightforward task. There are several associated challenges, not least because ‘toxic’ is a catch-all term that encompasses many potential hazards ranging from general food safety to environmental impacts. The problem is also more nuanced than just labeling something as ‘safe’ or ‘toxic’; relatively few molecules are safe at all exposure levels. As the old adage goes: the dose makes the poison. Accordingly, regulatory agencies create rules for minimum safety qualifications of flavors. In the United States, the Food and Drug Administration (FDA) labels additives that pass specific requirements as ‘generally recognized as safe’ (GRAS), but even these guidelines have not been without contention5,6. As new studies linking potential hazards emerge, regulatory bodies respond by re-evaluating and updating their safety guidelines. However, government agencies across the world move at different paces, complicating the matter further for companies who internationally distribute products. For example the European Food Safety Authority is known to ban flavors far quicker than the FDA7.
The need for food and flavor companies to adapt to the evolving regulatory landscape is clear. Unfortunately, solutions as nuanced as the problem itself may present a challenge. How do companies anticipate soon-to-be-banned ingredients and avoid reformulation downtime? If no currently approved additives are able to adequately replicate a banned ingredient’s properties, do companies begin developing new ones? Can additive manufacturers rely on traditional development approaches to create safe, new ingredients? Can newly developed technologies such as artificial intelligence (AI) be used to address these challenges faster, quicker, cheaper, and more reliably?
Computational Tools for Safety Evaluation
There are many resources that have been created to help simplify the process of evaluating the safety of new and existing molecules. For example, the EPA’s CompTox Chemical Dashboard8 helps users find information about existing molecules and suggests related molecules based on structure and other chemical characteristics. On one hand, well-curated dashboards like CompTox are of surprisingly great importance since the digital consolidation, storage, and organization of chemical data has been a historically challenging and monumental task. On the other hand, the sheer amount of information dashboards have can make them unwieldy and arduous to navigate.
Beyond curated datasets and dashboards, computational tools have been used for over 30 years to predict the safety of flavorings. The most popular computational methods used have been Quantitative Structure-Activity Relationship (QSAR) models, which take into account physico-chemical properties to predict a molecule's bioactivity and toxicity levels. QSARs are a standard metric accepted by many regulatory authorities, including the European Union and United States EPA9. The QSAR Toolbox10 is one such resource used to perform these analyses. However, QSAR models are vulnerable to misapplication and misinterpretation, especially among non-experts11. For example, QSAR models can be prone to error because of the quality of data used to create the model and are susceptible to outlier molecules that dramatically shift prediction values12. Put another way, a QSAR model's success largely depends on the scientist’s ability to navigate the pitfalls of model construction, and has been described as somewhat more of an ‘art than a science’13. Even with the most carefully considered model, structure-activity relationships are an incredibly subtle and complicated phenomenon; tiny differences between structurally similar molecules (known as bioisosteres) may vastly alter how a molecule interacts with the body14. While useful heuristics, QSAR models may inaccurately predict critical differences between bioisosteres and have also been found difficult to reproduce15. As such, researchers have spent considerable effort looking for new, more reliable predictive methods.
In the past decade, artificial intelligence-based models have become a prominent approach for chemical property prediction. What makes AI models so powerful is that they can learn non-linear trends and subtle correlations across hundreds or even thousands of variables, a feat that is all but impossible for human analysis or simple regression techniques alone. The types of data used to train AI models can come in many different forms, and can even integrate calculations predicted by simpler models, such as QSARs. However, the inclusion of calculations from other models does not mean that the AI model will succumb to similar biases, since it will also draw upon a wealth of other data during the learning process. Despite their tremendous potential, AI models still require careful construction, maintenance, and curation. Without proper guidance, AI models may not learn the desired trends.
A very effective solution has been the incorporation of scientific knowledge to constrain AI models, which helps keep predictions within the range of physical possibilities and reduces the total amount of data needed to train effective models. This approach, called Science-Based Artificial Intelligence (SBAI), leads to predictions far more accurate than either traditional AI or humans could achieve alone. Indeed, SBAI techniques have already been deployed to predict potential hazards ranging from environmental pollutants16 to biosecurity risks17. SBAI can be applied to all parts of the flavor development pipeline, from evaluating safety of existing flavors, suggesting regulation approved replacements, reformulating product recipes while maintaining the same food experience, and even suggesting brand-new molecules if no approved alternatives work.
NobleAI’s RAIR for Food and Flavors
At NobleAI, we have developed Risk Assessment services and Ingredient Replacement capability (RAIR) to address challenges exactly like those faced by the food and flavor industry. RAIR is part of the NobleAI Visualizations, Insights, and Predictions (VIP) Platform and uses SBAI models to evaluate the safety of ingredients and formulations and suggest safe alternatives. In contrast to other databases or specialized computational techniques that require expertise to use and interpret, the VIP platform can be easily used and interpreted by anyone.
By partnering with experts at NobleAI, organizations can identify at-risk additives in their existing formulations, rapidly replace ingredients in response to updated regulatory action, accelerate new formulation development, and even dramatically transform their R&D pipeline for new additives, all while saving time and money. For more information or to schedule a discovery call, visit our website or contact us today!
References:
[1] https://www.healthline.com/nutrition/artificial-food#potential-side-effects
[2] https://www.cnn.com/2023/02/27/health/zero-calorie-sweetener-heart-attack-stroke-wellness/index.html
[3] https://www.ewg.org/news-insights/news/2024/03/what-food-dye
[4] Bouvard, Véronique et al. “Carcinogenicity of consumption of red and processed meat.” The Lancet. Oncology vol. 16,16 (2015): 1599-600. doi:10.1016/S1470-2045(15)00444-1
[7] https://www.cbsnews.com/news/us-food-additives-banned-europe-making-americans-sick-expert-says/
[8] https://comptox.epa.gov/dashboard/
[9] Cronin, Mark TD, et al. "Use of QSARs in international decision-making frameworks to predict ecologic effects and environmental fate of chemical substances." Environmental Health Perspectives 111.10 (2003): 1376-1390.
[11] Cherkasov, Artem, et al. "QSAR modeling: where have you been? Where are you going to?." Journal of medicinal chemistry 57.12 (2014): 4977-5010.
[12] Zhao, Linlin, et al. "Experimental errors in QSAR modeling sets: what we can do and what we cannot do." ACS omega 2.6 (2017): 2805-2812.
[13] Cronin, Mark TD, and T. Wayne Schultz. "Pitfalls in QSAR." Journal of Molecular Structure: THEOCHEM 622.1-2 (2003): 39-51.
[14] Patani, George A., and Edmond J. LaVoie. "Bioisosterism: a rational approach in drug design." Chemical reviews 96.8 (1996): 3147-3176.
[15] Patel, Mukesh, et al. "Assessment and reproducibility of quantitative structure–activity relationship models by the nonexpert." Journal of Chemical Information and Modeling 58.3 (2018): 673-682.
[16] Huang, Lin, et al. "Exploring deep learning for air pollutant emission estimation." Geoscientific Model Development Discussions 2021 (2021): 1-22.
[17] Wittmann, Bruce B, et al. "Toward AI-Resilient Screening of Nucleic Acid Synthesis Orders: Process, Results, and Recommendations" bioRxiv. 4 Dec. 2024, doi: https://doi.org/10.1101/2024.12.02.626439
In the past century, the food and flavors industry has completely revolutionized the way that we grow, manufacture, store, and consume food. While the addition of herbs and spices to alter the taste of foodstuffs is a human experience that predates remembered history itself, the addition of synthetically produced molecules to fine-tune the qualities of our food is a much newer phenomenon. These ingredients, known as food additives, are extracted from natural sources or chemically synthesized and are used extensively across the food industry to alter food properties such as nutritional value, color, flavor, texture, and shelf-stability. Well known examples include natural and artificial sweeteners like stevia and aspartame to reduce caloric content, dyes like ‘Food Red No. 40’ and ‘Food Yellow No. 5’ that are used to make ‘visually desirable’ foods, or nitrates and nitrites added to preserve, stabilize color, and add flavor to meat products.
Despite widespread utilization of food additives, the safety of many have recently been called into question as mounting scientific evidence connects them to adverse health effects; artificial sweeteners have been implicated in misregulation of metabolism1 and to cause cardiovascular events like heart attack and stroke2; food dyes have been linked to myriad health problems including cancer and developmental problems3; and nitrates and nitrites have been linked to an increased risk of colorectal cancer4. Although many purported additive-disease linkages are still under professional consideration, the existing data and media scrutiny have been enough for the public and regulatory agencies to ask relevant questions and, in some cases, demand removal/replacement of some additives. It is anticipated that additional testing will lead to greater regulatory intervention and the need for ingredient replacement strategies to become a regular part of the product development process.
Unfortunately, identification and replacement of ‘toxic’ additives is no straightforward task. There are several associated challenges, not least because ‘toxic’ is a catch-all term that encompasses many potential hazards ranging from general food safety to environmental impacts. The problem is also more nuanced than just labeling something as ‘safe’ or ‘toxic’; relatively few molecules are safe at all exposure levels. As the old adage goes: the dose makes the poison. Accordingly, regulatory agencies create rules for minimum safety qualifications of flavors. In the United States, the Food and Drug Administration (FDA) labels additives that pass specific requirements as ‘generally recognized as safe’ (GRAS), but even these guidelines have not been without contention5,6. As new studies linking potential hazards emerge, regulatory bodies respond by re-evaluating and updating their safety guidelines. However, government agencies across the world move at different paces, complicating the matter further for companies who internationally distribute products. For example the European Food Safety Authority is known to ban flavors far quicker than the FDA7.
The need for food and flavor companies to adapt to the evolving regulatory landscape is clear. Unfortunately, solutions as nuanced as the problem itself may present a challenge. How do companies anticipate soon-to-be-banned ingredients and avoid reformulation downtime? If no currently approved additives are able to adequately replicate a banned ingredient’s properties, do companies begin developing new ones? Can additive manufacturers rely on traditional development approaches to create safe, new ingredients? Can newly developed technologies such as artificial intelligence (AI) be used to address these challenges faster, quicker, cheaper, and more reliably?
Computational Tools for Safety Evaluation
There are many resources that have been created to help simplify the process of evaluating the safety of new and existing molecules. For example, the EPA’s CompTox Chemical Dashboard8 helps users find information about existing molecules and suggests related molecules based on structure and other chemical characteristics. On one hand, well-curated dashboards like CompTox are of surprisingly great importance since the digital consolidation, storage, and organization of chemical data has been a historically challenging and monumental task. On the other hand, the sheer amount of information dashboards have can make them unwieldy and arduous to navigate.
Beyond curated datasets and dashboards, computational tools have been used for over 30 years to predict the safety of flavorings. The most popular computational methods used have been Quantitative Structure-Activity Relationship (QSAR) models, which take into account physico-chemical properties to predict a molecule's bioactivity and toxicity levels. QSARs are a standard metric accepted by many regulatory authorities, including the European Union and United States EPA9. The QSAR Toolbox10 is one such resource used to perform these analyses. However, QSAR models are vulnerable to misapplication and misinterpretation, especially among non-experts11. For example, QSAR models can be prone to error because of the quality of data used to create the model and are susceptible to outlier molecules that dramatically shift prediction values12. Put another way, a QSAR model's success largely depends on the scientist’s ability to navigate the pitfalls of model construction, and has been described as somewhat more of an ‘art than a science’13. Even with the most carefully considered model, structure-activity relationships are an incredibly subtle and complicated phenomenon; tiny differences between structurally similar molecules (known as bioisosteres) may vastly alter how a molecule interacts with the body14. While useful heuristics, QSAR models may inaccurately predict critical differences between bioisosteres and have also been found difficult to reproduce15. As such, researchers have spent considerable effort looking for new, more reliable predictive methods.
In the past decade, artificial intelligence-based models have become a prominent approach for chemical property prediction. What makes AI models so powerful is that they can learn non-linear trends and subtle correlations across hundreds or even thousands of variables, a feat that is all but impossible for human analysis or simple regression techniques alone. The types of data used to train AI models can come in many different forms, and can even integrate calculations predicted by simpler models, such as QSARs. However, the inclusion of calculations from other models does not mean that the AI model will succumb to similar biases, since it will also draw upon a wealth of other data during the learning process. Despite their tremendous potential, AI models still require careful construction, maintenance, and curation. Without proper guidance, AI models may not learn the desired trends.
A very effective solution has been the incorporation of scientific knowledge to constrain AI models, which helps keep predictions within the range of physical possibilities and reduces the total amount of data needed to train effective models. This approach, called Science-Based Artificial Intelligence (SBAI), leads to predictions far more accurate than either traditional AI or humans could achieve alone. Indeed, SBAI techniques have already been deployed to predict potential hazards ranging from environmental pollutants16 to biosecurity risks17. SBAI can be applied to all parts of the flavor development pipeline, from evaluating safety of existing flavors, suggesting regulation approved replacements, reformulating product recipes while maintaining the same food experience, and even suggesting brand-new molecules if no approved alternatives work.
NobleAI’s RAIR for Food and Flavors
At NobleAI, we have developed Risk Assessment services and Ingredient Replacement capability (RAIR) to address challenges exactly like those faced by the food and flavor industry. RAIR is part of the NobleAI Visualizations, Insights, and Predictions (VIP) Platform and uses SBAI models to evaluate the safety of ingredients and formulations and suggest safe alternatives. In contrast to other databases or specialized computational techniques that require expertise to use and interpret, the VIP platform can be easily used and interpreted by anyone.
By partnering with experts at NobleAI, organizations can identify at-risk additives in their existing formulations, rapidly replace ingredients in response to updated regulatory action, accelerate new formulation development, and even dramatically transform their R&D pipeline for new additives, all while saving time and money. For more information or to schedule a discovery call, visit our website or contact us today!
References:
[1] https://www.healthline.com/nutrition/artificial-food#potential-side-effects
[2] https://www.cnn.com/2023/02/27/health/zero-calorie-sweetener-heart-attack-stroke-wellness/index.html
[3] https://www.ewg.org/news-insights/news/2024/03/what-food-dye
[4] Bouvard, Véronique et al. “Carcinogenicity of consumption of red and processed meat.” The Lancet. Oncology vol. 16,16 (2015): 1599-600. doi:10.1016/S1470-2045(15)00444-1
[7] https://www.cbsnews.com/news/us-food-additives-banned-europe-making-americans-sick-expert-says/
[8] https://comptox.epa.gov/dashboard/
[9] Cronin, Mark TD, et al. "Use of QSARs in international decision-making frameworks to predict ecologic effects and environmental fate of chemical substances." Environmental Health Perspectives 111.10 (2003): 1376-1390.
[11] Cherkasov, Artem, et al. "QSAR modeling: where have you been? Where are you going to?." Journal of medicinal chemistry 57.12 (2014): 4977-5010.
[12] Zhao, Linlin, et al. "Experimental errors in QSAR modeling sets: what we can do and what we cannot do." ACS omega 2.6 (2017): 2805-2812.
[13] Cronin, Mark TD, and T. Wayne Schultz. "Pitfalls in QSAR." Journal of Molecular Structure: THEOCHEM 622.1-2 (2003): 39-51.
[14] Patani, George A., and Edmond J. LaVoie. "Bioisosterism: a rational approach in drug design." Chemical reviews 96.8 (1996): 3147-3176.
[15] Patel, Mukesh, et al. "Assessment and reproducibility of quantitative structure–activity relationship models by the nonexpert." Journal of Chemical Information and Modeling 58.3 (2018): 673-682.
[16] Huang, Lin, et al. "Exploring deep learning for air pollutant emission estimation." Geoscientific Model Development Discussions 2021 (2021): 1-22.
[17] Wittmann, Bruce B, et al. "Toward AI-Resilient Screening of Nucleic Acid Synthesis Orders: Process, Results, and Recommendations" bioRxiv. 4 Dec. 2024, doi: https://doi.org/10.1101/2024.12.02.626439
