Task geometry alignment enables parameter independent and accurate genomic search

Area of Science:

• Bioinformatics
• Computational Biology
• Genomic Analysis

Background:

• Standard sequence alignment relies on edit distance, which struggles with biological data complexities like insertions/deletions (indels).
• Existing methods often use heuristic gap penalties, limiting accuracy for certain biological sequence types.

Purpose of the Study:

• Introduce Task-Geometry Alignment (TGA) as a new design principle for sequence alignment.
• Develop and evaluate TGALIGN, a tool implementing TGA for improved genomic search.

Main Methods:

• TGALIGN encodes sequences into gap-robust syncmer profiles and uses Approximate Nearest Neighbor (ANN) search.
• The tool tiles reference databases to match query lengths, enabling efficient indexing.
• Expert-parameter independence is maintained throughout the process.

Main Results:

• TGAlign achieves performance comparable to state-of-the-art aligners on substitution-heavy markers (e.g., COI).
• Significant accuracy improvements (up to 10% on 16S) are observed for sequence fragments and indel-heavy markers.
• TGAlign matches MMseqs2 performance on highly variable ITS datasets while offering sub-millisecond query latency.

Conclusions:

• TGA provides a robust and scalable framework for post-alignment genomic search by aligning algorithmic structure with biological task geometry.
• TGAlign offers substantial accuracy and speed advantages, particularly for challenging sequence types.
• The tool's efficiency is driven by translating sequence comparison into dense matrix operations via ANN indexing.