MetageNN: a memory-efficient neural network taxonomic classifier robust to sequencing errors and missing genomes.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Long-read sequencing technologies offer high throughput but present challenges for accurate taxonomic classification due to higher error rates.
• Alignment-based methods are accurate but slow, while k-mer methods are faster but may lack sensitivity for underrepresented taxa.

Purpose of the Study:

• To develop and evaluate MetageNN, a memory-efficient, machine learning-based taxonomic classifier designed for long-read sequencing data.
• To address limitations of existing methods regarding sequencing errors, incomplete reference databases, and computational efficiency.

Main Methods:

• MetageNN utilizes a neural network model that processes short k-mer profiles of sequences to mitigate the impact of errors in long reads.
• The model was benchmarked against alignment-based (MetaMaps, MEGAN-LR) and k-mer-based (Kraken2) classifiers, as well as another machine learning approach (GeNet).

Main Results:

• MetageNN demonstrated substantial improvements in F1 score (20%) compared to GeNet on long-read data.
• It achieved significant sensitivity gains over MetaMaps (100%), MEGAN-LR (36%), and Kraken2 (23%) at the read level, particularly with incomplete reference databases.
• MetageNN requires less than 1/4 of the database storage of Kraken2, MEGAN-LR, and MMseqs2, and is substantially faster than existing tools.

Conclusions:

• Machine learning-based methods show significant promise for accurate and efficient taxonomic classification of long-read sequencing data.
• MetageNN provides a robust, memory-efficient alternative for classifying challenging sequences and can be further optimized.