Improving candidate Biosynthetic Gene Clusters in fungi through reinforcement learning

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate identification of Biosynthetic Gene Clusters (BGCs) is crucial for natural product discovery.
• Current BGC discovery tools often struggle with precise prediction of cluster boundaries and components, leading to overestimation.
• Manual curation by domain experts is time-consuming and essential for refining candidate BGCs.

Purpose of the Study:

• To develop a reinforcement learning (RL) approach to optimize the composition and boundaries of candidate BGCs.
• To enhance the performance of existing state-of-the-art BGC discovery tools.
• To reduce the manual curation effort required for BGC identification.

Main Methods:

• A novel reinforcement learning (RL) method was proposed, utilizing protein domain and functional annotation data from expert-curated BGCs.
• The RL approach was applied to refine candidate BGCs generated by established tools.
• The method was evaluated on BGC predictions for two fungal genomes: Aspergillus niger and Aspergillus nidulans.

Main Results:

• The RL method demonstrated significant improvements in gene precision (over 15%) for TOUCAN, fungiSMASH, and DeepBGC.
• Cluster precision saw improvements exceeding 25% for fungiSMASH and DeepBGC.
• These enhancements led to near-perfect precision in cluster prediction for the evaluated tools.
• The approach effectively minimized the overestimation of BGC boundaries by existing algorithms.

Conclusions:

• The proposed reinforcement learning approach offers a powerful strategy for optimizing BGC prediction in fungi.
• This method enhances the accuracy of BGC discovery tools, reducing the need for extensive manual curation.
• The findings pave the way for more efficient and precise identification of BGCs, accelerating natural product research.