Integrating in vivo characterization and machine learning for prediction of novel outer membrane-permeabilizing peptides

Abstract

The increasing prevalence of antibiotic resistance has created an urgent need for new strategies to treat bacterial infections. Gram-negative bacteria are particularly difficult to target due to their outer membrane (OM), which acts as an additional barrier limiting the uptake of many antibiotics. One strategy to overcome this barrier is the use of OM permeabilizing peptides (OMPPs), designed to selectively interact with and permeabilize the OM, thereby potentiating existing antibiotics without necessarily exhibiting direct antibacterial activity. In this thesis, a set of novel OMPPs was investigated using experimental characterization and in silico modeling. Experimentally, eight peptides were evaluated in Escherichia coli for antibacterial activity and potentiation of mupirocin using minimal inhibitory concentration assays and OM permeabilization using the fluorescent OM reporter NPN. In this set, one peptide emerged as the most promising OMPP candidate, facilitating a 2000-fold mupirocin potentiation and high NPN uptake. This suggests that it can permeabilize the OM and increase antibiotic entry. Meanwhile, another peptide showed strong antibacterial activity on its own, but no potentiation at sub-inhibitory concentrations. The in silico part evaluated descriptor-based machine learning models for the prediction of peptide performance. Molecular and physicochemical descriptors were combined with feature selection and nested cross-validation. These models were able to capture the training data well, but performance on unseen data was limited, indicating challenges with generalization. This may be explained by feature instability, correlated descriptors, small dataset size, and biological complexity. However, predictions for potentiation and NPN uptake showed more promising performance when evaluated using Spearman correlation. The selected descriptors were found to be related to biological properties such as hydrophobicity, amphipathicity, charge, and topology, which suggest that the descriptors contain meaningful signals relevant to membrane interactions.