A method to build extended sequence context models of point mutations and indels

Abstract

The mutation rate of a specific position in the human genome depends on the sequence context surrounding it. Modeling the mutation rate by estimating a rate for each possible k-mer, however, only works for small values of k since the data becomes too sparse for larger values of k. Here we propose a new method that solves this problem by grouping similar k-mers. We refer to the method as k-mer pattern partition and have implemented it in a software package called kmerPaPa. We use a large set of human de novo mutations to show that this new method leads to improved prediction of mutation rates and makes it possible to create models using wider sequence contexts than previous studies. As the first method of its kind, it does not only predict rates for point mutations but also insertions and deletions. We have additionally created a software package called Genovo that, given a k-mer pattern partition model, predicts the expected number of synonymous, missense, and other functional mutation types for each gene. Using this software, we show that the created mutation rate models increase the statistical power to detect genes containing disease-causing variants and to identify genes under strong selective constraint.

Regularized sequence-context mutational trees capture variation in mutation rates across the human genome

“…Here, we present Baymer, a Bayesian method to model mutation rate variation that computationally scales to large windows of nucleotide sequence context (S2 Text and S8 Table ), robustly manages sparse data through an efficient regularization strategy, and emits posterior probabilities that capture uncertainty in estimated probabilities. Consistent with previous studies [24][25][26], we show that expanded sequence context models in most current human datasets are overfit with classic empirical methods but considerably improve model performance when properly regularized. As a result, this method allows for renewed evaluation of experiments that originally were statistically limited to polymorphism probability models with small sequence context windows.…”

“…Next, we attempted to discover specific motifs that are enriched in the highest or lowest 1% of 9-mer polymorphism probabilities within each mutation type (S1 Text and S4 Table ). We recapitulate almost all previously reported motifs [19,23,25]. Consistent with previous reports, we identify a preponderance of repeat-rich motifs, which is perhaps due to the impact of slippage in introducing mutations [18].…”

“…While fewer than 2% of 9-mer edges meaningfully impact the final estimates, they still account for the most total absolute edges (7189 total edges > 0.95 PIP) and are enriched for larger effect sizes (S4C Fig) . This observation holds even after filters for data sparsity (S1 Text and Fig 3E). This implies a considerable impact on polymorphism probabilities in extended sequence contexts, consistent with previous work [19,[23][24][25]. This general trend is similarly consistent across mutation types (Fig 3F), although with a variable degree of impact, most notably with less additional variability estimated in wider CpG>T edges (S4D Fig) . We thus evaluated the overall improvement in likelihood by expanding window sizes up to 9-mers.…”

“…Previous work has sought to address these challenges, though methods introduced to date do not address all limitations simultaneously. Sparsity and scalability have been tackled through a deep-learning framework [ 24 ] as well as an IUPAC-motif-based clustering approach [ 25 ] which modeled polymorphism probabilities up through 9-mers. Another method explored polymorphism probabilities up through 7-mers using DNA shape covariates to reduce the parameter space [ 26 ].…”

Joint, multifaceted genomic analysis enables diagnosis of diverse, ultra-rare monogenic presentations

“…Different mutational processes lead to indel mutations, so Roulette values cannot necessarily be adapted to model this mutation type. 21 We approximated per-gene joint distributions of indel mutation rates and deleteriousness scores as follows. First, we considered all possible exonic indels of length ≤10nt for which precomputed CADD scores were available for download and all possible intronic insertions of length 1nt and deletions of length ≤4nt for which precomputed SpliceAI scores were available for download.…”