Impact of microbial profile integration on machine learning predictions of methane production: synergies and...

Microbial sequencing data were rarely integrated into the prediction of methane production using machine learning (ML) models because of high dimensionality and the lack of a systematic way to evaluate the change of insight gained from modelling with only physicochemical information. Here, key taxa were extracted with co-occurrence network analysis to reduce the dimension of the microbial profile. With 101 datasets with paired microbial and physiochemical features, integrating microbial features significantly enhanced accuracy for predicting methane production, increasing average R² from 0.73 to 0.79 and reducing mean absolute error from 8.7 to 8.0. Notably, integrating microbial features altered physicochemical feature impacts, shifting both their importance and directional effects. This underscores how microbial data refine mechanistic understanding and synergistically improve prediction accuracy, addressing a key gap left by models relying solely on physicochemical parameters. The work advocates for systematic microbial feature inclusion to advance methane production modelling with ML frameworks.