Prediction of laccase activity in deep eutectic solvents by combining COSMO-RS and machine learning techniques

• Machine learning and COSMO-RS model predict laccase activity in DES. • HBA descriptors strongly influence laccase activity in DES aqueous solutions. • ANN model showed the best performance with an RMSD of 0.958. • Optimized ANN accurately predicts laccase activity in DES aqueous solutions. • AI-driven approach accelerates selection of DES for enzymatic processes. Laccases are enzymes with significant roles in biotechnological processes across industries like food, textiles, chemistry and cosmetics. Recent studies have explored combining enzymes with deep eutectic solvents (DES) and their aqueous solutions to enhance enzymatic activity. However, the vast number of possible combinations of hydrogen bond acceptor (HBA) and hydrogen bond donor (HBD) to form DES makes their selection challenging and time-consuming. This study addresses this gap by showing the development of a predictive model that integrates molecular descriptors (activity coefficient, chemical potential, Misfit, vdW, and H-Bond) of solvent components assessed by conductor-like screening model for real solvent (COSMO-RS) with machine learning (ML) techniques to forecast laccase activity in DES aqueous solutions. Orange software was utilized to screen various supervised ML models, including Support Vector Machine, AdaBoost, Decision Trees, Gradient Boosting, Logistic Regression optimized with Stochastic Gradient Descent, K-Nearest Neighbors, and Artificial Neural Networks (ANN). Among those, ANN was selected for further optimization using MATLAB 2023b with the Deep Learning Toolbox™, applying the Levenberg-Marquardt backpropagation algorithm for network training. The model used five COSMO-RS molecular descriptors for each mixture component (HBA, HBD, and water), generating 15 inputs. The ANN architecture demonstrated robust predictive capabilities for relative laccase activity in DES solutions, achieving an impressive RMSE value of 0.958, with further validation against experimental data. Notably, the glycerol:xylitol DES exhibited an absolute error below 0.5%. These results highlight the potential of artificial intelligence tools in predicting enzyme activity in environmentally friendly solvents, with applicability to other fields.