Proof-of-concept prediction model: Hierarchical-Bayesian MCMC inference of Saccharomyces cerevisiae metabolic products in Very High Gravity fermentations

The role of stochastic dynamic systems modelling are ideal candidates for being able to reduce lab trials and provide an increased accuracy in predicting and or characterising the inherently statistically noisy environment of a complex biological system. A proof-of-concept approach using a direct variable prediction model is presented, rather than modelling the entire fermentation dynamic. A series of Artificial Neural Networks were utilised to build and validate a hierarchical-Bayesian Markov Chain Monte Carlo simulation. The model was built using the smoothed output from an n=11 ANN, validated using an n=12 ANN, and a n=6 minimal input prediction ANN. The concentration of glycerol, maltose, esters, and higher alcohols were then predicted at 1.078 original gravity and validated with experimental data. Dynamic time warping (DTW) distance (0.52) and z-scores of Vmax (hrs) (-1.121) showed that a minimal input prediction ANN (n=6) could be used to extract reliable kinetic inference simulation predictors to simulate the key compounds for an unfermented gravity, reducing laboratory demand to obtain predictions. The simulation showed a cohesive prediction pattern for esters and higher alcohols, the key flavour compound groups in whisky products, and maltose using CO2 kinetic inference predictors. The work highlighted that the underlying biochemical mechanisms can be captured by the inference predictors. This research highlights the benefit of low maintenance stochastic modelling with predictive ANNs to reduce the architectural limitations and demands of modelling complex biological systems under stress, in a proof-of-concept research study.