A new paradigm in lignocellulolytic enzyme cocktail optimization: Free from expert-level prior knowledge and experimental datasets

Abstract

Effectively pairing diverse lignocellulolytic enzyme cocktails with intricately structured lignocellulosic substrates is an enduring challenge for science and technology. To date, extensive trial-and-error remains the primary approach and no deep-learning methods were developed to address it due to limited experimental data and incomplete expert-level knowledge of enzyme-cocktail-substrate structure-dynamics-function relationships. Here, a novel model is developed to tackle this issue in efficient, cost-effective, and high-throughput manners. It needs no pre-labeled datasets, instead utilizing simple features, eliminating the reliance on expert-level prior knowledge of reaction mechanisms. Experimentally optimal combinations were found within predicted ranges of tailor-made combinations with precision of 91.98%, covering 80.00% of overall top-100. Practical tests demonstrated its effectiveness in narrowing down potential optimal combinations, speeding up targeted screening, and enabling efficient degradation of lignocellulosic biomass. The method has good applications in artificial proteins biosynthesis from low-value lignocellulosic straw, providing alternative solutions for biomass biorefining challenges in complex enzyme-cocktail-substrate interactions.