Authors:

Elliot Nick Brajkovich

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Peptide natural products have long been utilized as drugs, with many having antibacterial, antifungal, and anticancer activity. Testing for these activities requires an intense investment into expression and purification, prompting bioinformatics-based approaches to prioritize promising compounds. This project aims to use machine learning to streamline the discovery of bioactive non-ribosomal peptides by predicting macrocyclization type from predicted biosynthetic gene cluster sequences. The genetic organization of non- ribosomal peptide synthetases has inspired many tools for predicting peptide structure. However, these tools have yet to reliably predict final macrocyclization by thioesterase domains despite the documented importance of macrocyclization for bioactivity. Utilizing data from the MIBiG 3.0 repository, multiple models were trained to predict macrocyclization type from synthetase derived features, including the DNA sequence of the thioesterase domain. Best results have been achieved using a boosted decision tree model, yielding PR-AUC and ROC-AUC scores of 0.81 and 0.74, respectively. Input data was initially restricted to only 115 entries with experimentally confirmed features to preserve fidelity; however final evaluations show no significant impact on performance utilizing DNA sequences alone. This initial discovery both reduces the anticipated need for transfer learning upon deployment and increases the number of usable entries for additional model training to improve performance. Keywords: Natural Products; Machine Learning; Peptide

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Elliot Nick Brajkovich