Data-Driven Prediction of Molecular Biotransformations in Food Fermentation

Abstract

Fermentation products, together with food components, determine the sense, nutrition, and safety of fermented foods. Traditional methods of fermentation product identification are time-consuming and cumbersome, which cannot meet the increasing need for the identification of the extensive bioactive metabolites produced during food fermentation. Hence, we propose a data-driven integrated platform (FFExplorer, http://www.rxnfinder.org/ffexplorer/) based on machine learning and data on 2,192,862 microbial sequence-encoded enzymes for computational prediction of fermentation products. Using FFExplorer, we explained the mechanism behind the disappearance of spicy taste during pepper fermentation and evaluated the detoxification effects of microbial fermentation for common food contaminants. FFExplorer will provide a valuable reference for inferring bioactive "dark matter" in fermented foods and exploring the application potential of microorganisms.

Discovery of Toxin-Degrading Enzymes with Positive Unlabeled Deep Learning

“…Despite successful application in many areas, interpretability remains challenging for deep learning-based methods . PU-EPP uses attention mechanisms to capture the importance and contributions of different input positions (e.g., atoms in substrates or residues in enzymes) to the final prediction, thereby inferring the knowledge the model has learned (Figure A,B). Though embeddings could not provide a perfect way to fully interpret the results, we mapped the attention weights to the substrates and enzymes and compared them with known data to verify whether PU-EPP could learn the catalytic mechanism behind the data.…”

“…We acquired 170,179 enzymes, 5837 substrates, and 606,555 corresponding enzyme–substrate pairs. Since negative data are rarely reported, previous studies usually augmented negative samples for training classification models based on activity thresholds or random sampling. , However, low activity is not equivalent to inactivity, and certain randomly sampled enzyme–substrate pairs may actually be functional (positive) because of substrate promiscuity, which would result in the deep learning models acquiring inaccurate results. Therefore, we proposed a strategy that combines weighted random sampling and positive unlabeled (PU) learning to minimize the impact of inaccurate negative samples (see the Methods section for details).…”

“…To interpret which residues on enzymes and which atoms on substrates are most important for the catalysis process, the multi-head self-attention mechanism was employed by assigning attention weights to the residues and atoms. For instance, a higher attention weight for a residue means that the residue is more important for enzyme activity toward the specific substrate; a higher attention weight for an atom means that it is important for binding with enzymes or means it may be present at the reaction sites (i.e., the atoms and bonds changed in the substrate during reactions) .…”

“…The multi-head self-attention mechanism was employed to interpret which subregions on enzymes, and substrates contributed more to the final prediction (Figure C). We used a multi-head self-attention layer in the encoder to evaluate the contribution of each point of the input enzyme sequence.…”

Predicting Antioxidant Synergism via Artificial Intelligence and Benchtop Data

“…Albeit effective for simple experimental designs, these one-dimensional methods often hinder the evaluation of nonlinear interactions − due to the complexity and multifaceted nature of antioxidant responses, which are affected by several factors such as their mechanism of action, structural properties, and matrix effects. On the contrary, machine learning approaches are particularly well suited to address these complex problems and have been recently applied to make predictions related to taste − and fermentation of foods, screen for Nrf2-agonists, identify flours infested by insects, and even discover green insecticides …”