Optimisation and Extension of the Structural Variant Database for Long-Read Sequencing Data

Sammanfattning

Structural variant databases are widely used in research and clinical settings to build, merge and query structural variants from variant call files. These databases help distinguish rare from common variants and support the identification of disease-associated alterations. The toolkit SVDB applies the DBSCAN algorithm and uses positional parameters to cluster variants based on their start and end coordinates. It helps gather information such as frequency and occur- rences of structural variants. With growing interest in long-read sequencing, positional cluster- ing alone has become insufficient, particularly for insertions, where sequence composition must be considered to detect rare structural variants. Alternative clustering strategies were explored , including OPTICS and interval trees, together with two different sequence distance metrics, Hamming and Levenshtein. Two datasets were used: a small dataset serving as a truth-set to evaluate clustering accuracy and larger datasets to assess performance and scalability. The evaluation applied an initial positional clustering step, followed by sequence-based clustering for insertions. The results showed that DBSCAN produced the most accurate clusters when compared with the truth-set created from the Platinum Pedigree files. Databases built from the Platinum Pedigree samples showed expected biological patterns: parents were least similar, parent–child pairs more similar and siblings the most similar, although this analysis was restricted to chromo- some 1. Sequence comparison experiments showed that Hamming distance was faster, whereas Levenshtein distance gave more accurate similarity assessments, highlighting a trade-off be- tween speed and accuracy. Additionally, cluster content showed that Hamming produced more clusters with single sequences compared to Levenshtein which clustered more clusters contain- ing at least two sequences. Levenshtein had more flexibility and could distinguish more SVs compared to Hamming. Hamming was more suitable for equal length strings and hence only substitutions. Large scale evaluation using samples from the 1000 Genomes Project showed that DBSCAN was the fastest algorithm and used the least memory; however, because this dataset contained only deletions and duplications, it could not be used to assess distance metrics for insertions. Overall, DBSCAN produced clusters closest to the truth-set, though not perfectly. Graph-based strategy may therefore be more suitable for SVDB. Interval trees were tested as a graph-based method but requires further evaluation. A promising future direction for the toolkit is to combine DBSCAN with Levenshtein distance for a reclustering of insertions after positional clustering. Additionally, evaluate the graph-based approach with the interval tree for long-read sequences.