Integrating metagenomics and explainable artificial intelligence for modeling of food waste treatment using full-scale anaerobic digestion

Anaerobic digestion (AD), a biochemical process that can convert food waste (FW) into methane, offers great promise as a sustainable form of energy production. While several attempts have been made to optimize AD systems using various mathematical models, more precise modeling approaches that fully consider the complexity of the AD process are required, leading to the adoption of artificial intelligence (AI) as a suitable alternative to numerical modeling. In line with this, the present study tested 11 AI-based models on their prediction of the methane yield for a full-scale AD process using FW as a feedstock. The models incorporated operational parameters, environmental conditions, and microbial information to improve their predictive performance. Although a one-dimensional convolutional neural network (1D-CNN) was the most precise, random forest regression (RFR) was selected as the optimal model for further analysis due to its superior interpretability and stability. Explainable AI (XAI) was then used to determine the most important input features contributing to the predictions of the optimal AI model, thus allowing for detailed model interpretation. Methanothrix was identified as a key predictor of methane yield, and metagenomic analysis provided independent genome-level evidence broadly consistent with the XAI results. Overall, this study proposes a novel approach to the interpretation and optimization of AD performance, rather than focusing only on enhancing the predictive performance of a discrete model.