Application of Machine Learning Algorithms to Estimate Enzyme Loading, Immobilization Yield, Activity Retention, and Reusability of Enzyme–Metal–Organic Framework Biocatalysts

Abstract

The ability to predict enzyme−metal−organic framework (MOF) properties such as enzyme loading, immobilization yield, activity retention, and reusability can maximize product yield and extend the operational life of enzyme−MOF biocatalysts. However, this is challenging due to the vast combinations of available metal and ligand building blocks for MOF and enzymes. Therefore, several machine learning (ML) algorithms are applied in this study using data collected from 127 journal articles to estimate these biocatalyst parameters. Twelve input variables, including the metal and ligand properties of MOF, as well as the enzyme properties, were integrated and fed into two ML algorithmsrandom forest and Gaussian process regression (GPR)to predict model outputs. A 10-fold cross-validation approach with grid search was applied to obtain the optimal hyperparameter values. The random forest model (RFM) provided more accurate estimates of the enzyme loading, immobilization yield, and reusability of the biocatalyst than the GPR model, with relatively high R 2 values of 0.85, 0.77, and 0.91, respectively. Both models are less effective in predicting the enzyme activity retention, however, with R 2 values of 0.63 or lower. Sensitivity analysis of the input variables revealed the most significant variables for each corresponding output parameter, allowing further optimization of the RFM. The final RFM was then tested with a second unseen dataset collected from experiments. Findings confirmed the validity of the predictive model, including a relative error of less than 25%. Our model can aid in the synthesis of enzyme−MOF biocatalysts by providing valuable estimates of these output parameters for different MOF precursors and enzymes, saving experimental time and cost.

Putting precision and elegance in enzyme immobilisation with bio-orthogonal chemistry

“…To save experimental time and cost, several machine learning algorithms (random forest and Gaussian process regression) were used to predict enzyme-MOF properties, including enzyme loading, activity retention, immobilisation yield and reusability. 243 Recently, Raman hyperspectral imaging with principal component analysis has been developed to discern the spatial and chemical distribution of immobilised enzymes. 244 Furthermore, chemical imaging has been combined with machine learning to elucidate the immobilisation of enzymes for biocatalysis.…”

Machine Learning in the Development of Adsorbents for Clean Energy Application and Greenhouse Gas Capture

“…[116][117][118] SHAP can also give unreliable results when features are correlated, and thus the results should be scrutinized by domain specialists. For NN models, salience methods (e.g., class activation maps), [119] attention masks, [120] and partial derivatives (sensitivity analysis) [121] are used to interpret these "black box" models. Interested readers can find more details about model interpretations in the recent review by Oviedo et al [122] Good practice in ML research requires good quality data.…”

Unified Roadmap for ZIF-8 Nucleation and Growth: Machine Learning Analysis of Synthetic Variables and Their Impact on Particle Size and Morphology

“…A successful ML approach relies on the existence of appropriate databases. ,− It can provide interesting alternatives for the identification of suitable candidates for various applications or the determination of key structural and functional features, and it has been successfully employed in MOF research over the last years. − Another possible alternative is to use ML for systematization and validation of multiple reported synthesis methods for a particular material, as it was recently demonstrated for an archetypal Copper-trimesic acid MOF, HKUST-1. − Another example is the use of linear models to investigate the formation of defects on UiO-66 MOF and their impact on its catalytic and adsorptive performance . This approach provided valuable insights into optimizing defect formation through various synthetic parameters.…”