Lerna: transformer architectures for configuring error correction tools for short- and long-read genome sequencing.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Sequencing technologies introduce errors, necessitating error correction (EC) for accurate downstream analysis.
• Manual configuration of EC tools is suboptimal, as optimal parameters are dataset- and tool-dependent.
• Evaluating EC performance typically requires a reference genome, which is not always available.

Purpose of the Study:

• To introduce Lerna, a tool for automated configuration of k-mer-based EC tools.
• To develop a method for optimizing EC parameters without relying on a reference genome.
• To improve de novo genome assembly quality through automated EC optimization.

Main Methods:

• Lerna constructs a language model (LM) of uncorrected genomic reads.
• It calculates a perplexity metric using the LM to evaluate corrected reads for various parameter settings.
• The tool identifies parameters yielding the highest alignment rate by correlating perplexity with assembly quality, bypassing the need for a reference genome.

Main Results:

• Optimal k-mer values are dataset-specific, validating the need for automated selection.
• The perplexity metric, estimated by attention-based transformers, effectively predicts EC performance.
• Lerna demonstrates strong negative correlation between perplexity and alignment rate/assembly quality, enabling automated k-mer tuning.
• The approach achieves significant runtime improvements (18x faster) due to optimized model architecture and GPU utilization.

Conclusions:

• Lerna enhances de novo genome assembly by optimizing EC tools.
• The developed method automates parameter tuning without requiring a reference genome.
• The code is publicly available for broader research use.