In this research, machine learning models were used to predict the output current of electrochemical sensors and pKa values in proteins. The aim of the first work is to design an advanced machine learning model to predict the current value in an electrochemical sensor, and in the second work, a model is presented to predict the pKa values, which is very important for understanding the acidic dissociation behavior of proteins. In the field of electrochemical sensing, machine learning models have been focused on to enhance the performance and reliability of an electrochemical sensor. The aim is to design a machine learning model to predict the current in a molecularly imprinted polymers based sensor. Molecularly imprinted polymers-based sensors are made from polymers that are specifically designed to detect target molecules. These sensors can identify and measure the target molecule by creating holes in the polymer structure that match the shape and size of the target molecule. These features cause high performance and selectivity in diagnosis. The characteristics used in this study include loading of multi-walled carbon nanotubes on the surface, doxorubicin concentration, pyrrole concentration, number of electropolymerization cyclic voltammetry cycles, electropolymerization cyclic voltammetry scanning speed, number of cyclic voltammetry scanning cycles during overoxidation and incubation time. Four machine learning models have been used to eva‎luate the effect of each parameter on the prediction performance, using the SHAP method to determine the importance of the feature. Based on the analysis, removing a less influential feature and introducing a new feature significantly improved the predictive capabilities of the model. In addition, the use of the Min-Max scaling technique ensured that all features contributed proportionally to the model learning process. Different machine learning models have been applied to the dataset and their performance in predicting the sensor output current has been compared. To further increase the predictive performance, the stacking regressor technique was used, in which basic regression models are combined, their strengths are used, and their individual weaknesses are compensated. This stacking regressor ultimately resulted in a model that outperformed the other individual models in terms of both performance and reliability. Similarly, in predicting protein pKa values, a machine learning-based approach was developed by carefully selecting a set of features that represent the intrinsic properties of molecular residues and the effects of the molecular environment. These features include residue name, B factor, surface accessibility, number of hydrogen bonds, number of heavy atoms, and number of ionizable residues, which are related to factors affecting pKa such as chemical identity, solvent exposure, and hydrogen bond networks. Various regression models including linear regression, support vector regression, and random forests have been applied and their performance was eva‎luated using cross-validation techniques based on criteria such as mean absolute error (MAE) and root mean square error (RMSE). Similar to the electrochemical sensor work, the stacking regressor technique resulted in the best model that achieved the highest level of predictive performance by effectively capturing the complex relationships between selected features and pKa values. Overall, our research demonstrates the effectiveness of machine learning, particularly through ensemble approaches such as stacking regressor, in improving predictive modeling in electrochemical and biochemical sensing.

اسماعيل حيدري

نفيسه فهيمي كاشاني , حسين فرخ پور