TRGT-denovo: accurate detection of de novo tandem repeat mutations

Abstract

Motivation: Identifying de novo tandem repeat (TR) mutations on a genome-wide scale is essential for understanding genetic variability and its implications in rare diseases. While PacBio HiFi sequencing data enhances the accessibility of the genome's TR regions for genotyping, simple de novo calling strategies often generate an excess of likely false positives, which can obscure true positive findings, particularly as the number of surveyed genomic regions increases.

Results: We developed TRGT-denovo, a computational method designed to accurately identify all types of de novo TR mutations-including expansions, contractions, and compositional changes-within family trios. TRGT-denovo directly interrogates read evidence, allowing for the detection of subtle variations often overlooked in variant call format (VCF) files. TRGT-denovo improves the precision and specificity of de novo mutation (DNM) identification, reducing the number of de novo candidates by an order of magnitude compared to genotype-based approaches. In our experiments involving eight rare disease trios previously studiedTRGT-denovo correctly reclassified all false positive DNM candidates as true negatives. Using an expanded repeat catalog, it identified new candidates, of which 95% (19/20) were experimentally validated, demonstrating its effectiveness in minimizing likely false positives while maintaining high sensitivity for true discoveries.

Availability and implementation: Built in Rust, TRGT-denovo is available as source code and a pre-compiled Linux binary along with a user guide at: https://github.com/PacificBiosciences/trgt-denovo.

Fig 1.. Overview of TRGT-denovo

(full details in Methods). (a) TRGT pre-processing, which requires aligned PacBio HiFi reads, a repeat definition catalog, and a reference genome. (b) TRGT-denovo uses TRGT output, specifically spanning reads and genotyping data, along with the reference genome and repeat definitions. (c) By matching repeat definitions and corresponding allele sequences, reads are partitioned and assigned to alleles. This is achieved via TRGT-obtained classifications, consensus allele alignment, or phasing, thus determining the allele sequence each read best supports. (d) Allele partitioned reads are realigned to child allele consensus sequences for comparison purposes. (e) Potential DNMs are identified by examining discrepancies in alignment score distributions among candidate de novo alleles.

Fig 2.. TRGT-denovo metrics.

De novo coverage relative to the (a) allele de novo ratio; (b) child de novo ratio; (c) mean absolute difference between the reads with de novo evidence and P_u. Each point represents a potential de novo allele. Horizontal and vertical lines indicate thresholds for minimal de novo coverage, allele de novo ratio, and a range for the child de novo ratio, creating shaded boxes where true de novo mutations are more likely.

Fig 4.. Alignment score distributions.

Distributions of alignment scores for reads spanning alleles (${\text{M}}_0$, ${\text{M}}_1$, ${\text{F}}_0$, ${\text{F}}_1$, ${\text{C}}_0$, ${\text{C}}_1$) when aligned to alleles ${\text{C}}_0$ (a) and ${\text{C}}_1$ (b). WFA alignment scores range from negative, less similar, to zero, perfect match. Inheritance patterns, as inferred from surrounding genetic variation, are: ${\text{M}}_0\to {\text{C}}_0$(inherited) and ${\text{F}}_1\to {\text{C}}_1$ (inherited + de novo). Symbols P_U and C_L denote the parental upper bound and candidate de novo allele lower bound respectively. Each point corresponds to an individual read aligned to ${\text{C}}_0$ or ${\text{C}}_1$. Red-outlined points highlight two observations: a read from ${\text{F}}_1$, exceeding C_L, showing overlap with ${\text{C}}_1$, and a read from ${\text{C}}_1$ that falls below P_U, and overlaps with ${\text{F}}_1$. In ${\text{C}}_1$ there are 19 reads, of which 18 exceed P_U, contributing to the de novo coverage.