Nathalie Perrot

This chapter presents three examples of data-based machine learning on time series. The common denominator of these case studies is the sparseness of data, making machine learning results fragile and inaccurate. We show how human expertise can be effectively mobilized for building useful systems, for instance useful decision support systems, able to better meet the needs of the agri-food chain. The design and analysis of different features of machine learning coupled with human knowledge enables us to sketch future human-centered machine learning systems. This approach is very relevant for the modeling of agri-food systems, because human expertise, skills and know-how are rich and numerous, but often implicit, data are heterogeneous-big and sparse-and processes are complex and deeply conditioned by human needs and interactions.

In this paper, we present a test of an interactive modelling scheme in real conditions. The aim is to use this scheme to identify the physiological responses of microorganisms at different scales in a real industrial application context. The originality of the proposed tool, Biosys-LiDeOGraM, is to generate through a human–machine cooperation a consistent and concise model from molecules to microbial population scales: If multi-omics measurements can be connected relatively easily to the response of the biological system at the molecular scale, connecting them to the macroscopic level of the biosystem remains a difficult task, where human knowledge plays a crucial role. The use-case considered here pertains to an engineering process of freeze-drying and storage of Lactic Acid Bacteria. Producing a satisfying model of this process is a challenge due to (i) the scarcity and variability of the experimental dataset, (ii) the complexity and multi-scale nature of biological phenomena, and (iii) the wide knowledge about the biological mechanisms involved in this process. The Biosys-LiDeOGraM tool has two main components that can have to be utilized in an iterative manner: the Genomic Interactive Clustering (GIC) module and the Interactive Multi-Scale modellIng Exploration (IMSIE) module, both involve users in their learning loops. Applying our approach to a dataset of 2,741 genes, an initial model, as a graph involving 33 variables and 165 equations, was first built. Then the system was able to interactively improve a synthetic version of this model using only 27 variables and 16 equations. The final graph providing a consistent and explainable biological model. This graphical representation allows various user interpretations at local and global scales, an easy confrontation with data, and an exploration of various assumptions. Finally Biosys-LiDeOGraM is easily transferable to other use-cases of multi-scale modelling using ‘functional’ graphs. Author summary: The use of “omics” data for understanding biological systems has become prevalent in several research domains. However, the data generated from diverse macroscopic scales used for this purpose is highly heterogeneous and challenging to integrate. Yet, it is crucial to incorporate this information to gain a comprehensive understanding of the underlying biological system. Although various integrative analysis methods that have been developed provide predictive molecular-scale models, they only offer a mechanistic view of the biological system at the cellular level. In addition, they often focus on specific biological hypotheses through dedicated case studies, making it difficult to apply their results to other scientific problems. To address these issues, we propose an interactive multi-scale modelling approach to integrate cross-scale relationships providing predictive and potentially explanatory models. A proof-of-concept tool has been developed and was validated in the context of the bioproduction of Lactococcus lactis , a bacterial species of high economic interest in the food industry and for which the control of the bioprocess is essential to guarantee its viability and functionality. Our approach can be applied to any biological system that can be defined through a set of variables, constraints and scales.

Odor is central to food quality. Still, a major challenge is to understand how the odorants present in a given food contribute to its specific odor profile, and how to predict this olfactory outcome from the chemical composition. In this proof-of-concept study, we seek to develop an integrative model that combines expert knowledge, fuzzy logic, and machine learning to predict the quantitative odor description of complex mixtures of odorants. The model output is the intensity of relevant odor sensory attributes calculated on the basis of the content in odor-active comounds. The core of the model is the mathematically formalized knowledge of four senior flavorists, which provided a set of optimized rules describing the sensory-relevant combinations of odor qualities the experts have in mind to elaborate the target odor sensory attributes. The model first queries analytical and sensory databases in order to standardize, homogenize, and quantitatively code the odor descriptors of the odorants. Then the standardized odor descriptors are translated into a limited number of odor qualities used by the experts thanks to an ontology. A third step consists of aggregating all the information in terms of odor qualities across all the odorants found in a given product. The final step is a set of knowledge-based fuzzy membership functions representing the flavorist expertise and ensuring the prediction of the intensity of the target odor sensory descriptors on the basis of the products' aggregated odor qualities; several methods of optimization of the fuzzy membership functions have been tested. Finally, the model was applied to predict the odor profile of 16 red wines from two grape varieties for which the content in odorants was available. The results showed that the model can predict the perceptual outcome of food odor with a certain level of accuracy, and may also provide insights into combinations of odorants not mentioned by the experts.

When creating a flavor to elicit a specific odor object characterized by odor sensory attributes (OSA), expert perfumers or flavorists use mental combinations of odor qualities (OQ) such as Fruity, Green, and Smoky. However, OSA and OQ are not directly related to the molecular composition in terms of odorants that constitute the chemical stimuli supporting odor object perception because of the complex non-linear integration of odor mixtures within the olfactory system. Indeed, single odorants are described with odor descriptors (OD), which can be found in various databases. Although classifications and aroma wheels studied the relationships between OD and OQ, the results were highly dependent on the studied products. Nevertheless, ontologies have proven to be very useful in sharing concepts across applications in a generic way and to allow experts’ knowledge integration, implying non-linear cognitive processes. In this paper, we constructed the Ontology for Odor Perceptual Space (OOPS) to merge OD into a set of OQ best characterizing the odor, further translated into a set of OSA thanks to expert knowledge integration. Results showed that OOPS can help bridge molecular composition to odor perception and description, as demonstrated in the case of wines.