arising from J. G. Monroe et al. Nature https://doi.org/10.1038/s41586-021-04269-6 (2022)

Although mutation rates vary within genomes, suggestions1,2 that more selectively important DNA has a lower mutation rate are contentious not least because unbiased estimation of the mutation rate is challenging3. Monroe et al.4 (hereafter Monroe) also report that in Arabidopsis more important sequences have lower mutation rates and, while overlooking similar claims1, suggest that this challenges “a long-standing paradigm regarding the randomness of mutation”4. We find, however, that their mutation calling has abundant sequencing and analysis artefacts explaining why their data are not congruent with well-evidenced mutational profiles. As the key trends associated with sequence importance are consistent with well-described mutation-calling artefacts and are not resilient to reanalysis using the higher-quality components of their data, we conclude that their claims are not robustly substantiated.

In principle, identifying new mutations is simple: one sequences genomes of close relatives and identifies new differences between them. There are, however, multiple pitfalls. For example, incorrect mapping of short reads to the genome can result in erroneous, and commonly clustered, mutation calls. Further, as the rate of sequencing errors is orders of magnitude higher than the rate of mutation, these errors must be excluded. Robust rules for mutation calling, such as requiring multiple independent sequence reads from both strands supporting the same mutation, can obviate many issues.

We expect higher than normal error rates in Monroe as, to identify somatic mutations, they relaxed their previous5 stringency in mutation calling (Supplementary Methods). To assay the impact of this, we compared their mutation calls to those generated by a conventional pipe. We find only 3.7% (n = 160) concordance with Monroe’s 4,322 filtered putative mutations, 61% of which are uncallable. We term the 96.3% of Monroe’s mutations that fail conventional filters as low quality (LQ). Their previous data5 using a more stringent pipe (we dub this Weng data) agree with our analysis: 94.2% agreement vis-à-vis mutations ‘confidently’ called, only 1.4% uncallable. Prima facie, most of Monroe’s mutation calls thus may well be unsafe. This is supported by analysis of the profile of LQ mutations as this is different to the higher-quality calls: they are of a qualitatively different type (intergenic, intronic and so on) compared with high-quality datasets (that mutually agree; Extended Data Fig. 1a,b) and have a different mononucleotide mutational profile (Extended Data Fig. 1c).

Deviations of this magnitude are unlikely to be accounted for by a somatic versus germline difference. Instead, Monroe’s data are different largely because they are enriched for sequencing and analysis artefacts. We consider two artefact fingerprints. First, in Illumina sequencing6,7, a base can be erroneously replaced by a bleeding one within8,9 (for example, AAAGAAA appears as AAAAAAA) or in the vicinity of6,7 (for example, AAAAC appears as AAAAA) homopolymeric runs. Second, failure to eliminate poor-quality reads and mapping artefacts will overreport clustering of putative mutations.

More than half (54%) in the LQ data are bleeding-type putative mutations within 5 base pairs of A/T homopolymers (Extended Data Fig. 1d), compared with 24% in the high-quality (HQ) Weng set. Within genic regions (exonic + intronic), the proportion is 67.5% in the LQ data, 91% of which are intronic. Depending on the protocol, the error is biased to bleed-over of either AT or GC residues, but not both6. It is then notable that the bias is particular to A/T homomeric runs with 1.5% in HQ Weng and 1.2% in LQ near GC runs (Extended Data Fig. 1d). This AT bias artefact is reported for related Illumina machines6.

Of the 2,247 bleed mutations near homopolymeric runs, 1,149 (51.1%) are immediately next to or within the runs. Of the remaining 1,098, at least 648 cluster with other bleed errors (for example, AAAAACACACA is read as AAAAAAAAAAA giving three putative mutations). These bear the hallmarks of artefacts: typically only one strand is affected, all of the putative mutations are seen in the same read and their rate decays as a function of distance from the true end of the run. As also expected from the profile of sequencing errors6, the probability of a mutation being called increases with the length of the homopolymer: in introns, regression of log10(putative mutations per base pair of homopolymeric sequence) predicted by run length, slope = 0.27, Pearson’s r2 = 0.87, P = 0.0006, degrees of freedom (df) = 6.

For Monroe’s somatic mutations (recalled from Weng’s vcf data) unassociated with homomeric runs (46% of their mutations), most are clustered (2 mutations within 10 base pairs of each, 27.5% of all mutations) or unexpectedly common (>10 samples, 8.7% of all mutations), indicative of mis-mapping issues. Only 2.5% in the Weng HQ data are clustered. Many of Monroe’s putative mutations are associated with more than one error: about 34% are associated with A/T homomeric runs and in a tight cluster. As centromeres are prone to mapping errors10, mis-mapping probably explains why 40.9% of LQ mutations are centromeric (see, for example, Extended Data Fig. 1e) compared with 27.9% in Weng HQ.

We do not suppose these to be all of the errors. Whereas Monroe call 773,141 mutations using our sequence11, using the same HaplotypeCaller-GVCF calling method12, with default parameters and without any further filtering, we identify only 31,486 raw indels and 72,516 raw single nucleotide polymorphisms (all but 17 of which are unsafe). This gross excess of mutations in Monroe is due to an analysis error on their part (see correction from Monroe et al.13).

Analysis and sequencing errors explain many of the core claims of Monroe. They report that the mutation rate alters markedly around transcription start sites (TSSs) and stop sites (TTSs), arguing that this provides evidence that gene bodies are mutationally protected. We simulated random errors associated with A/T homopolymeric runs and derived a distribution that is a near-perfect match to their somatic data (Fig. 1a).

Fig. 1: Core claims of Monroe et al. are error artefacts.
figure 1

a, Top row: the profile of errors (misascribed as mutations) expected around TSS and TTS attributable to the errors associated with A/T homopolymeric runs. We simulated 2 million errors associated with A/T homopolymeric runs and used the Monroe script to generate the left plot. The data are an exact match to their somatic data calls (reproduced here as the top right plot, figure reused from Monroe et al.4 under Creative Commons Attribution 4.0 International License). Bottom row: germline mutations in TEs (left panel). Observed mutations in germline unmasked. bp, base pair. b, Essential genes do not have a low mutation rate. The second claim of Monroe to substantiate that mutation is lower in more functionally important sequences is that essential genes have the lowest mutation rates (their Fig. 3c). To substantiate this, they seem to have used many thousands of unfiltered calls. We repeat the analysis using filtered data, either Weng or Monroe, including indels. In neither case is there significant heterogeneity (we provide χ2 values for all comparisons, df = 3, but for Monroe CDS numbers are so small that these calculations are not valid and presented for completeness alone). Tests are one-sided in the sense that we call significance only if there is heterogeneity not if they are more similar than expected. Tests are two-sided in the sense that we ask about deviation from null in any direction. Unification of the two datasets does not alter conclusions: CDS, χ2 = 3.3; intron, χ2 = 5.76, P > 0.05 for all without multi-test correction. Error bars are s.e.m. across gene samples for which sample sizes are: essential (n = 719), morphological (n = 861), cellular or biochemical (n = 297) and environmental (n = 522) for CDS analysis (Monroe CDS and Weng CDS), and essential (n = 671), morphological (n = 789), cellular or biochemical (n = 270) and environmental (n = 452) for the intron analysis (Monroe intron and Weng intron). For representation of the underlying data, see Extended Data Fig. 2.

Full size image

Monroe also claimed a low mutation rate in essential genes as evidence for mutational protection for more important sequences (their Fig. 3c). However, to do this, they included orders of magnitude more putative mutations than in their filtered datasets. We repeated their analysis using the Weng data and the Monroe data (their filtering). In neither, nor in the merged dataset, is there heterogeneity in the mutation rate between gene classes in coding sequence (CDS) or intron (Fig. 1b and Extended Data Fig. 2). Indeed, in the best data (Weng), essential genes have the highest mutation rate per base pair of CDS (Fig. 1b and Extended Data Fig. 2).

Artefacts explain other heterogeneities in the mutation rate. Weng’s data report a plausible intron to CDS per base pair ratio of about 0.91:1 (paired t-test on normalized dinucleotide mutation rates, P = 0.58, df = 95), whereas Monroe’s data report an unprecedented 5.2 to 1 ratio (paired t, P = 3.5 × 10−8, df = 95). This comparison is especially informative as it controls for transcription-associated mutational effects. Much of this higher intronic rate in Monroe’s data is attributable to homopolymeric run artefacts as CDS has fewer, and less error prone, runs (Supplementary Results).

The artefacts are also evident in the profile of mutations called (Extended Data Fig. 1f). Counts of the 96 dinucleotide mutations from the Monroe and Weng data are discordant (χ2 = 1,516, P < 2 × 10−16, df = 95). The most common dinucleotide mutations in the Monroe data end AA or TT as the resolved mutational event, with G/C mutated to the neighbouring A/T being especially discrepant (Extended Data Fig. 1f). The mutational events terminating AA/TT contribute 34.4% of the relative normalized mutations in the Monroe set but only 21.6% in Weng’s.

The Monroe data also incorrectly predict the mutational equilibrium frequency of AA/TT dinucleotides compared to observed frequencies. Using the 16 × 16 normalized mutational matrix for the Monroe and the Weng data individually, we predict mutational equilibrium dinucleotide content14 and compare with intergenic dinucleotide content. The Weng data are not influenced by AA/TT calls (P = 0.38), whereas in the Monroe data AA/TT are over-called outliers (P = 0.003; Extended Data Fig. 1g).

This neighbour base matching affecting both A and T in Monroe’s data is an expected bleed artefact with no biological basis. By contrast, we expect CpG>TpG mutations to be common given well-described methylated CpG hyperinstability15. In Weng’s data, but much less so Monroe’s, this is the case (Extended Data Fig. 1f).

Although Monroe’s claim that the mutation rates are lower at more functionally important sites seems to be highly influenced by artefacts, nonetheless, the mutation rate is not uniform. In some part this is because transposable elements (TEs) have high mutation rates and TEs are rare in gene bodies. In Arabidopsis, cytosine methylation-mediated TE suppression16 should lead to C instability. In the Weng data, 65% of mutations in TE are CG, CHH or CHG versus 51% in intergenic non-TE (for example, 5.2:1 ratio of CpG>TpG per CG, TE to intergenic non-TE). In (robust) germline data there is a higher mutation rate in TEs than elsewhere, including the best comparator, non-TE intergenic sequence: TE versus non-TE intergenic sequence, mean ratio per dinucleotide = 3.93 (paired t-test on normalized dinucleotide rates, P < 3 × 10−7, df = 95). This TE enrichment largely explains why in germline data mutation rates are higher 5′ of TSS and 3′ of TTS, TEs being enriched outside transcribed domains (Fig. 1a).

Although TE mutational enrichment is seen in Monroe HQ data (Extended Data Fig. 1a,b), it is not seen in the Monroe data in toto (paired t-test on normalized dinucleotide rates, P = 0.9). Given this and the failure to capture well-described methyl C instability15, although we do not doubt that epigenetic marks such as methylation can affect mutation, the correlations evidenced by Monroe between various marks and mutation rate variation should be treated with the same caution as their claim that mutation is rarer in more important sequences.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.