MACHINE LEARNING-GUIDED IDENTIFICATION OF ANTIMICROBIAL PEPTIDES FOR BONE AND JOINT INFECTIONS

Aim Bacterial infections in joints and bones are becoming increasingly challenging to treat due to the rise in antibiotic resistance. This limits the effectiveness of conventional antibiotics, thus prolonging infection and decreasing the patients’ life quality. Antimicrobial peptides (AMPs) are natural occurring peptides, which have been shown as an alternative due to their broad-spectrum activity and lower risk for resistance development. Yet, the discovery of highly active AMPs is a timeconsuming and costly process. This study aims to develop a machine learning based approach to predict and identify highly active AMPs based only on the peptide sequence and optimizing the screening process. Method Our model was developed and trained on an existing AMP database, 1500 peptides were used for the training, with 200 known peptides used for its validation. The model was designed to categorize peptides based on their half-maximal inhibitory concentration (IC50). Peptides were classified as active if the predicted IC50 values, relative to the control peptide Bac2A, were below 0.5. The model was used to predict active AMPs in a dataset of 50.000 peptides with unknown IC50. The best ranked peptides identified will be validated in vitro by determining the minimum inhibitory concentration, minimal bactericidal concentration and minimal biofilm inhibitory concentrations. Cytocompatibility in various bone-relevant cell lines will likewise be evaluated. Results The preliminary computational analysis had a training accuracy of 97%, and a precession of 87%. The validation metrics likewise showed acceptable results with an accuracy of 87% and a precession of 83%. More interestingly, an internal validation of the model was carried out with a dataset of 40 peptides with known activity level. Here the model has 95% accuracy and 100% precision. This shows that while the model predicts active peptides with high accuracy it is prone to false negatives. Conclusions This study demonstrates how machine learning can aid in finding novel AMPs. By computationally predicting active peptides, the initial experimental screening time is reduced. Future work will focus on validating the peptides found using the model.