MetaTransformer: deep metagenomic sequencing read classification using self-attention models.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Deep learning, particularly transformers, shows promise in analyzing genomic sequences.
• Existing tools like DeepMicrobes face challenges with slow runtimes and high memory usage.
• Metagenomic analysis requires efficient and accurate taxonomic prediction.

Purpose of the Study:

• To introduce MetaTransformer, a novel deep learning tool for metagenomic analysis.
• To improve upon the speed and memory efficiency of existing taxonomic classifiers.
• To evaluate the performance of transformer-encoder models and embedding schemes in metagenomics.

Main Methods:

• Developed MetaTransformer, a self-attention-based deep learning tool utilizing transformer-encoder models.
• Investigated different embedding schemes to optimize memory consumption and performance.
• Compared MetaTransformer's performance against DeepMicrobes for species and genus classification.

Main Results:

• MetaTransformer achieved superior species and genus classification accuracy compared to DeepMicrobes.
• The tool demonstrated a 2× to 5× speedup in inference time with a smaller memory footprint.
• Training times for MetaTransformer were 9 hours for genus and 16 hours for species prediction.

Conclusions:

• Self-attention models significantly improve performance in deep learning for metagenomic analysis.
• MetaTransformer offers an efficient and accurate solution for taxonomic prediction in metagenomics.
• Embedding schemes play a crucial role in optimizing deep learning models for genomic data.