RFBGCpred: A Random forest based tool for prediction of biosynthetic gene clusters.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Biosynthetic gene clusters (BGCs) are crucial for producing natural products with significant applications.
• Existing tools like antiSMASH and DeepBGC have limitations in identifying complex or atypical BGC architectures.

Purpose of the Study:

• To develop and present RFBGCPred, an open-source machine learning classifier.
• To enhance the classification accuracy for five specific, high-impact BGC classes: PKS, NRPS, RiPPs, terpenes, and PKS-NRPS hybrids.
• To complement existing BGC detection pipelines by improving class-level discrimination.

Main Methods:

• Utilized curated data from the MiBIG database for training.
• Employed Word2Vec for feature extraction and supervised UMAP for dimensionality reduction.
• Applied SMOTE to handle class imbalance and Random Forest as the final classifier, selected via the TOPSIS criterion.

Main Results:

• The Random Forest model achieved high performance: 98.0% accuracy (MCC: 0.9752, AUC: 0.9928) on a balanced test set.
• Demonstrated strong generalization on an unbalanced validation set with 94.8% accuracy (MCC: 0.89, AUC: 0.96).
• Showed improved recall for hybrid PKS-NRPS clusters compared to antiSMASH and DeepBGC, with competitive precision.

Conclusions:

• RFBGCPred effectively classifies key BGC types, particularly hybrid architectures, reducing misclassification errors.
• The tool offers a valuable addition to the bioinformatics toolkit for natural product discovery.
• RFBGCPred is accessible, supporting multiple input formats and providing all resources via GitHub.