Predicting the Textural Properties of Plant-Based Meat Analogs with Machine Learning.

Plant-based meat analogs are food products that mimic the appearance, texture, and taste of real meat. The development process requires laborious experimental iterations and expert knowledge to meet consumer expectations. To address these problems, we propose a machine learning (ML)-based framework to predict the textural properties of meat analogs. We introduce the proximate compositions of the raw materials, namely protein, fat, carbohydrate, fibre, ash, and moisture, in percentages and the "targeted moisture contents" of the meat analogs as input features of the ML models, such as Ridge, XGBoost, and MLP, adopting a build-in feature selection mechanism for predicting "Hardness" and "Chewiness". We achieved a mean absolute percentage error (MAPE) of 22.9%, root mean square error (RMSE) of 10.101 for Hardness, MAPE of 14.5%, and RMSE of 6.035 for Chewiness. In addition, carbohydrates, fat and targeted moisture content are found to be the most important factors in determining textural properties. We also investigate multicollinearity among the features, linearity of the designed model, and inconsistent food compositions for validation of the experimental design. Our results have shown that ML is an effective aid in formulating plant-based meat analogs, laying out the groundwork to expediently optimize product development cycles to reduce costs.

Disease gene classification with metagraph representations.

Protein-protein interaction (PPI) networks play an important role in studying the functional roles of proteins, including their association with diseases. However, protein interaction networks are not sufficient without the support of additional biological knowledge for proteins such as their molecular functions and biological processes. To complement and enrich PPI networks, we propose to exploit biological properties of individual proteins. More specifically, we integrate keywords describing protein properties into the PPI network, and construct a novel PPI-Keywords (PPIK) network consisting of both proteins and keywords as two different types of nodes. As disease proteins tend to have a similar topological characteristics on the PPIK network, we further propose to represent proteins with metagraphs. Different from a traditional network motif or subgraph, a metagraph can capture a particular topological arrangement involving the interactions/associations between both proteins and keywords. Based on the novel metagraph representations for proteins, we further build classifiers for disease protein classification through supervised learning. Our experiments on three different PPI databases demonstrate that the proposed method consistently improves disease protein prediction across various classifiers, by 15.3% in AUC on average. It outperforms the baselines including the diffusion-based methods (e.g., RWR) and the module-based methods by 13.8-32.9% for overall disease protein prediction. For predicting breast cancer genes, it outperforms RWR, PRINCE and the module-based baselines by 6.6-14.2%. Finally, our predictions also turn out to have better correlations with literature findings from PubMed.