BELLA: Berkeley Efficient Long-Read to Long-Read Aligner and Overlapper

Abstract

Recent advances in long-read sequencing allow characterization of genome structure and its variation within and between species at a resolution not previously possible. Detection of overlap between reads is an essential component of many long read genome pipelines, such as de novo genome assembly. Longer reads simplify genome assembly and improve reconstruction contiguity, but current long read technologies are associated with moderate to high error rates.In this work, we present Berkeley Efficient Long-Read to Long-Read Aligner and Overlapper (BELLA), a novel overlap detection and alignment algorithm using sparse matrixmatrix multiplication. In addition, we present a probabilistic model that demonstrates the feasibility of using k-mers for overlap candidate detection and shows its flexibility when applied to different k-mer selection strategies. Based on such a model, we introduce a notion of reliable k-mers. Combining reliable k-mers with our binning mechanism increases the computational efficiency and accuracy of our algorithm. Finally, we present a new method based on Chernoff bounds to separate true overlaps from false positives by combining alignment techniques and probabilistic modeling. Our goal is to maximize the balance of precision and recall.For both real and synthetic data, BELLA is among the best F1 scores, showing a stability of performance that is often lacking in competing software. BELLA's F1 score is consistently within 1.7% of the top performer. In particular, we show improved de novo assembly quality on synthetic data when BELLA is coupled with the miniasm assembler.

Asynchronous Distributed Actor-Based Approach to Jaccard Similarity for Genome Comparisons

“…It uses cardinality estimation methods highly specialized for set union and set intersection. BELLA [40] is an overlap detection and alignment algorithm that defines genome similarity calculations using sparse matrix-matrix multiplication. Similarly, GenomeAtScale [10], the SOTA algorithm, is a matrix-matrix multiplication approach to genome similarity that allows for parallel data compression and batch computation.…”

A simple theory for finding related sequences by adding probabilities of alternative alignments

“…The range is adjusted as the algorithm proceeds, attempting to follow the highest score. The best way to do this is unclear: several ways have been suggested [7,[25][26][27][28].…”

An Efficient Parallel Sketch-based Algorithmic Workflow for Mapping Long Reads

“…The rationale for choosing a region with least 'N' content is simple-to improve the mapping quality (recall, in particular) for long reads that have repetitive content. We use ℓ = 2000 bp as our segment length for L2L, based on prior works that have suggested similar lengths [8,37].…”

“…However, this approach achieves a very low F1 score [8]. So, the balance between precision and recall is not maintained for complex genomes.…”

“…Broadly speaking, two classes of approaches exist for using long reads—standalone and hybrid. Standalone long read assemblers [6, 7] produce a de novo assembly from long reads using the Overlap-Layout-Consensus (OLC) paradigm [8]. HiFi long reads significantly facilitate the task of standalone assemblers and with higher accuracy in the reads, the overlap graph becomes smaller as compared to the overlap graphs generated from error-prone long reads.…”