Athena: Automated Tuning of k-mer based Genomic Error Correction Algorithms using Language Models

Area of Science:

• Bioinformatics and Computational Biology
• Genomics and Next-Generation Sequencing (NGS)
• Natural Language Processing (NLP) applications

Background:

• Error-correction (EC) algorithms for genomics reads critically depend on optimal configuration parameters (e.g., k-mer value).
• Optimal parameters vary across datasets (species, platforms) and EC tools, necessitating adaptive tuning.
• Current methods often require a reference genome for ground truth, limiting applicability.

Purpose of the Study:

• To develop an adaptive method for automatically optimizing EC configuration parameters to enhance genome assembly quality.
• To leverage NLP language modeling techniques for efficient and accurate parameter tuning.
• To validate the use of NLP's 'perplexity' metric for quantifying EC performance.

Main Methods:

• Developed 'Athena', an algorithmic suite employing N-Gram and Recurrent Neural Network (RNN) language models from NLP.
• Repurposed the 'perplexity' metric from NLP to quantitatively assess EC performance.
• Utilized a hill-climbing search guided by perplexity to find optimal configuration parameters (e.g., k-value).

Main Results:

• Athena accurately identified optimal k-values for 7 real datasets across 3 EC algorithms (Lighter, Blue, Racer).
• Perplexity showed a strong negative correlation (≥73%) with EC quality across diverse datasets and error types.
• Athena's selected k-values achieved assembly alignment rates within 0.53% of brute-force optimization, improving NG50 by 4.72X.

Conclusions:

• NLP-based language modeling provides an effective and efficient approach for adaptive EC parameter optimization.
• The perplexity metric reliably predicts EC performance, enabling automated tuning without reference genomes.
• Athena significantly enhances genome assembly quality by optimizing EC parameters, demonstrating broad applicability.