Multi-label classification with XGBoost for metabolic pathway prediction.

Area of Science:

• Systems biology
• Computational biology
• Genomics

Background:

• Metabolic pathway prediction is crucial for reconstructing organism metabolic networks from genomic data.
• Machine learning (ML) methods show promise but previously underperformed rule-based methods like PathoLogic.
• Previous evaluations of PathoLogic omitted taxonomic pruning, negatively impacting its performance.

Purpose of the Study:

• To re-evaluate PathoLogic with taxonomic pruning and compare it against ML approaches.
• To introduce mlXGPR, an improved XGBoost-based metabolic pathway prediction method.
• To enhance ML pathway prediction by incorporating label correlations via classifier chains.

Main Methods:

• Updated evaluation of PathoLogic incorporating taxonomic pruning.
• Development of mlXGPR using XGBoost and a multi-label classification framework.
• Implementation of classifier chains with a novel ranking method to leverage label correlations.
• Benchmarking mlXGPR against existing methods on single- and multi-organism datasets.

Main Results:

• PathoLogic with taxonomic pruning outperforms previous ML methods.
• mlXGPR, particularly with classifier chains, surpasses PathoLogic and other ML methods.
• mlXGPR achieves superior performance in terms of Hamming loss, precision, and F1 score on single-organism benchmarks.

Conclusions:

• ML-based metabolic pathway prediction can achieve high performance.
• Optimized ML methods, like mlXGPR, can outperform established tools such as PathoLogic with taxonomic pruning.
• Further improvements in ML approaches are viable for competitive metabolic pathway prediction.