Prediction and validation of nanowire proteins in Oleidesulfovibrio alaskensis G20 using machine learning and feature engineering

The type-IV bacterial pili and multiheme c-type cytochrome protein family have gained significant attention due to their role in extracellular electron transfer (EET), which defines their electrogenic properties. These electrogenic behaviors play a crucial role in interspecies microbial communication, essential for microbial biofilm formation and the development of robust technologies such as biosensors. Given the technological and ecological significance of electron transfer mechanisms, this study presents NanowireML (NWML), a 2-staged machine learning (ML) system to identify and analyze nanowire (NW) proteins. Stage 1 predicts NW proteins using minimal features, and Stage 2 leverages graphical knowledge representation to predict the most relevant NW mechanism governing candidates in specific experimental conditions. To train the stage 1 model, we used a comprehensive dataset of 999 proteins from a public database specializing in NW development. The primary objective of the NWML model is to identify and validate microbial proteins involved in the biogenesis of NW. Protein feature, such as dipeptide amino acid composition, transition, and distribution enhance the model's performance. Gene ontology (GO) analysis revealed that NW are structural extrusions, part of membranal proteins, with several exposed metal ion binding motifs. The predicted NW protein collection advances to stage 2, where their GO knowledge is stored using a graphical representation. A customized deep neural network (biologically influenced neural networks: BINN) then predicts the most relevant biological pathways governing NW formation using experimental gene expression data. Our study provides detailed insights into gene sets, unveiling the mechanistic networks and pathways crucial for NW formation. These findings enable data-driven decisions for biomedical and biotechnological applications. The NWML model demonstrated high accuracy achieving 94.87 %, 96.68 %, 96.65 %, 96.05 %, and 96.13 % on support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), logistic regression (LR), and artificial neural network (ANN).