A1153
Title: MV-FS: A feature identification and prediction ensemble for complex soft matter systems
Authors: Nawisa Jullapech - University of Reading (United Kingdom) [presenting]
Fazil Baksh - University of Reading (United Kingdom)
Zuowei Wang - University of Reading (United Kingdom)
Abstract: A novel ensemble method is developed and evaluated that combines feature selection and prediction based on majority vote (MV-FS) to address the limitations of base predictors and ensure robust performance across diverse datasets. Theoretical analyses and simulations reveal how variations in model performance and inter-model relationships affect overall effectiveness, which in turn can be used to ensure generalization and stability, even in challenging classification scenarios. To demonstrate its practical utility, the proposed methodology is applied to two distinct soft-matter systems. First, it extends current machine-learning approaches in drug discovery by investigating the link between antimicrobial peptide (AMP) structure and antibacterial activity, a key area in antibiotic discovery. By integrating domain knowledge of physicochemical and structural features with a data-driven ensemble, MV-FS identifies critical factors influencing AMP activity and predicts promising regions in the physicochemical space for active peptides. Second, the method is used in a preliminary study on predicting conformational transitions of single-charged polymers, relevant for polymer physics and materials science. By combining multiple models, MV-FS is shown to overcome the limitations of individual methods and achieve results consistent with theoretical expectations. The proposed ensemble's ability to enhance interpretability and performance underscores its value in soft matter systems.
