Analysis model

Our motivation is to quantify the impact of deamination at motifs, both known and hypothetical, across many different potential targets. Previous structural descriptions of APOBEC mutagenesis have indicated the role of a recognition pocket in the -2 and -1 positions on the single-stranded DNA substrate preceding the targeted Cytosine [59]. We therefore examined the effects of mutation at each of the 16 different NNC hypothetical hotspots, referred throughout as NNC motifs (N = any nucleotide) to NNT, consistent with the AID/APOBEC deamination mechanism in the absence of further repair. We measure a genome’s susceptibility to mutation by extending a method first introduced by Karlin [60] (see: “Calculating susceptibility” section in Methods for a detailed definition, including discussion of necessary controls). Briefly, the model we use assesses the impact of a particular hotspot or motif (for example, AGC) on a given gene coding sequence by first counting the number of observed motifs in both orientations, then counting the number of nonsynonymous (changing the amino acid) mutations on the sequence that would occur by mutating C to T in every observed motif (or G to A in the reverse orientation) (Fig 1A). In quantifying a coding sequence’s susceptibility to mutations in a particular motif, we assumed that nonsynonymous mutations have more potential to be deleterious. Given the complexity of nucleotide sequence evolution, there are many factors this simple model does not account for, such as the GC content of the genome, or the impact of hypothetical mutations other than from C to T. To address these considerations, we also evaluate several alternative models (see: “Alternative models and controls corroborate observed susceptibilities” section below).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Description of the model used to calculate susceptibility of a sequence to a particular motif.

A) The example sequence has 4 AGC motifs, indicated in purple (with two motifs occurring in the reverse orientation as GCT). B) If every AGC and GCT mutation were mutated to AGT and ACT (as indicated in orange) consistent with AID deamination, three out of four of these motifs cause nonsynonymous mutations changing the corresponding amino acid sequence. C) Susceptibility to the AGC motif is calculated by comparing the percentage of random sequences (maintaining protein sequence) with a lower nonsynonymous mutation fraction. D) Since only one random sequence out of four has a lower nonsynonymous mutation fraction than the baseline, this sequence’s percentile, or susceptibility is 0.25.

https://doi.org/10.1371/journal.pcbi.1005471.g001

To illustrate our method, we show an example in Fig 1. Here, the target sequence has 4 AGC motifs, with two on the forward strand, and two on the reverse strand (appearing as the reverse complement of AGC, GCT) of which 3 cause nonsynonymous mutations and 1 is synonymous when mutated from C/G to T/A. This sequence therefore has a nonsynonymous mutation fraction of 0.75, equivalent to the probability that a mutation at an AGC site causes an amino acid change (Fig 1B).

This fraction was then compared to a null model consisting of 1000 randomized sequences that preserve the amino acid sequence, but where synonymous codons are chosen randomly. For each randomized sequence in the null model, the nonsynonymous mutation fraction was calculated. Under the model, if a gene is susceptible to a particular NNC motif, then it would have a higher incidence of nonsynonymous mutations than the majority of randomized versions from the null model. We therefore refer to the percentile of randomized sequences having a lower nonsynonymous mutation fraction than the wild-type sequence as that gene’s susceptibility for a given motif (Fig 1C). It is in fact equivalent to a P-value for susceptibility. Thus, a particular motif is susceptible to mutation if the percentile is close to 1. Conversely, a gene with a low susceptibility, closer to 0, can be considered resistant to that same motif. As an example, if we are considering AGC to AGT mutations, if a gene has a nonsynonymous mutation fraction of 0.75 but in the null model, 1 out of 4 permutations have an even lower nonsynonymous mutation fraction, this gene’s sequence would be at the 25^th percentile for the nonsynonymous mutation fraction of the AGC motif (Fig 1D).

We next examined how mutational susceptibility of retrotransposons and viral genomes was affected by different APOBEC preferences. We defined a gene set as a group of genes with a common biological function, for example, a set of housekeeping genes, or the genes in a particular virus. Starting with the individual susceptibility measures for each gene, we calculated the average susceptibility for each gene set (Fig 2A). As an example, Fig 2B shows the susceptibility values for the Epstein-Barr Virus (EBV), a herpesvirus that is tropic to B-cells, potentially exposing it to AID, which has a strong preference for mutating AGC motifs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Computing average susceptibilities for each genome to a given motif.

A) The susceptibility scores in the EBV genome are calculated for each gene together with the average B) Distribution of percentiles in the EBV genome and C) in the human housekeeping gene set, to the AID motif AGC. The average, shown in a bold colored line, is low (close to 0) in a very resistant genome.

https://doi.org/10.1371/journal.pcbi.1005471.g002

We reasoned that under a null model, if susceptibility is equivalent to a P-value, then random percentiles would assume a uniform distribution from 0 to 1, and would have a median and mean of 0.5. In contrast, the average susceptibility of the EBV genome is approximately 0.2 and for the set of human housekeeping genes it is 0.4 (Fig 2C), suggesting human housekeeping genes are slightly less resistant to the impact of AGC deamination. As each coding sequence susceptibility score can be treated as a P-value, to assess the susceptibility of each gene set statistically we combined the individual P-values using an unweighted Z-transform approach, which is more powerful than the standard Fisher method [61]. We found this approach was capable of detecting significance even for small gene sets, such as HIV or Adeno-Associated Virus 2 (AAV2), which have only 7 and 8 genes respectively but show significant results for several motifs (S1 Table). In the example of Fig 2B, the susceptibility of the EBV genome to the AGC motif was statistically lower than a null median of 0.5 (P < 10⁻⁴⁹, S1 Table) and in general EBV displays statistical resistance to 13 of the 16 NNC motifs we examined (using a cutoff of P < 3.125 × 10⁻³ for significance, representing Bonferroni corrected P = 0.05). We applied the same z-score method to determine the statistical vulnerability of a gene set, defining it as the combination of the scores (1-susceptibility) to obtain a P-value (S2 Table).

Differences in susceptibility across genomes

To quantify the impact of deamination on various genomes we analyzed a diverse set of genomes, including viruses previously reported as AID or APOBEC targets, sets of housekeeping genes from mammalian (mouse and human) genomes and, as controls, genomes from viruses of non-vertebrate hosts that have evolved in the absence of AID/APOBECs (see Methods section “Data Sources”, and Table 2). For each of these gene sets we were interested in quantifying the impact of C>T mutations at the reported hotspots for AID/APOBEC (Table 1). Since we wanted to gain insight into why these particular hotspots evolved, we considered all motifs of the form NNC (N = any nucleotide), which includes 10 hotspots targeted by AID/APOBEC and 6 hypothetical hotspots such as GCC that have not evolved specificity within the AID/APOBEC family. We measured susceptibility values (as described above) for every gene within each gene set, calculating the mean susceptibility in each case. This was repeated for all 16 NNC motifs, thus obtaining 16 susceptibility values for each gene set.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of gene sets analyzed.

https://doi.org/10.1371/journal.pcbi.1005471.t002

We observed that one of these gene set clusters (indicated by the red box labeled “Resistant” in Fig 3) has high overall resistance (i.e., low susceptibility) to NNC motifs. This resistant cluster includes many viral genomes, such as Human Herpesvirus 1 [31], and hepatitis C Virus [35]. Curiously, this cluster includes the Epstein-Barr Virus, suggesting that this virus may have evolved resistance to mutations by both APOBECs and AID. Although to our knowledge no studies have demonstrated a direct mechanism of AID restricting EBV experimentally, the virus is B-cell tropic and would presumably be exposed to AID deamination in germinal centers and to an extent in extrafollicular compartments [62]. Gene sets of host (mouse and human) housekeeping genes also show overall resistance to mutations in NNC motifs. Housekeeping genes will, by definition, be co-expressed with AID and those APOBEC genes wherever they are expressed. Although little is known about the mechanisms by which our host genomes avoid APOBEC mutations, the evidence from AID in B-cells shows that mutations indeed occur genome-wide but are repaired, although the repair process is imperfect [56]. Our analysis here suggests that this mutational behavior (perhaps including APOBEC) may have created selective pressures on the genome to minimize this potentially hazardous activity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Gene sets cluster by hypothetical NNC motif susceptibility.

The clustering suggests that the gene sets we examined segregate clearly into AID/APOBEC-resistant (boxed in red), and AID/APOBEC-vulnerable clusters (boxed in blue). Many of the known hotspots for AID and APOBEC also cluster together (boxed in orange).

https://doi.org/10.1371/journal.pcbi.1005471.g003

The cluster of gene sets in the blue box of Fig 3 and labeled as Vulnerable had higher mean susceptibility scores across all NNC motifs. The cluster includes a human endogenous retrovirus (HERV), the human (LINE-1) and mouse retrotransposon (LINE-1 and MusD) coding regions, as well as the control genomes, which include viruses of invertebrate hosts (oyster, plants) that would not have co-evolved with AID or APOBEC. Surprisingly, there are several virus gene sets that are targeted by APOBEC and that we did not initially expect to be vulnerable, including HIV, Simian Foamy Virus (SFV), Hepatitis B Virus, Murid Herpesvirus 4 and Human Papillomavirus 16. HIV’s apparent vulnerability is surprising since HIV is a well-known target of human APOBEC3G. However, as described in the Introduction, HIV encodes vif, an APOBEC3G inhibitor, which may obviate the need for sequence-level avoidance. Other unknown defense mechanisms for these targeted viruses may also explain their apparent vulnerability. Mouse APOBEC3 is incapable of impairing MHV68 (Murid Herpesvirus 4) viral replication, for example, whereas several human APOBEC3s are capable of restriction, suggesting an alternate, unknown anti-mAPOBEC defense may be present [63]. Furthermore, recent work on HIV has shown that APOBEC3G may in fact help the virus mutate and create escape variants [64], which may in turn explain its relatively high level of susceptibility. A similar explanation may hold for human papillomavirus (we analyzed HPV16), which recent studies have shown to be extensively mutated in vivo and in vitro [28]. In light of these phenomena, our results suggest that even if a virus targeted by APOBEC is found to be susceptible to APOBEC hotspots, the virus itself may contain an alternate method of escape such as vif, subversion of APOBEC for evolutionary purposes, or other factors.

Considering now the hierarchical clustering results for the 16 NNC motifs (vertical dendrogram of Fig 3) we observed that many existing (or “canonical”) AID/APOBEC hotspots, including 7 of the 10 AID/APOBEC hotspots are contained within the same cluster (right-most 7 motifs in Fig 3, boxed in orange). This cluster contains 3 of the 4 canonical AID hotspots AGC, AAC, and TGC (only excluding TAC), as well as 3 of the canonical human APOBEC3B NTC hotspots (ATC, GTC, and TTC, only excluding CTC), and the mouse APOBEC3 hotspot TCC, but excludes the human APOBEC3G hotspot CCC. In light of the contrast observed, the existence of the two clusters, resistant and vulnerable, suggests that AID/APOBEC hotspots cause high mutational damage to exogenous viruses and native transposable elements. At the same time, we have observed that many of the genes targeted (intentionally in the cases of viruses and unintentionally in the case of host housekeeping genes) show resilience to the impact of these mutations at hotspots. This resilience may have evolved as an avoidance strategy to minimize mutations if the hotspots that APOBEC evolved do not cause high collateral damage to the host genome.

Alternative models and controls corroborate observed susceptibilities

GC bias is an important evolutionary constraint in many viruses, and may play important APOBEC-independent roles such as protection from insertions [65]. Although we demonstrate later that the GC content of the gene sets considered correlates strongly with susceptibility (see next section “GC content of genomes predicts deaminase hotspot susceptibility”), we sought to quantify the extent to which native GC content contributed to the observed susceptibilities. To that end, for each of the gene sets considered, we generated new susceptibility scores that controlled for the GC content by defining a new null model that used the GC content of the gene set being analyzed to select codons (see Methods). Using the same unweighted z-score calculation as described above to determine statistical significance, we observed that for the high GC-content EBV genome, 9 out of the 16 NNC trinucleotide contexts that we examined were still statistically resistant (compared to 13 out of 16 in the uncorrected case). For the vulnerable OsHV genome, 6 out of 16 NNC trinucleotides were statistically vulnerable (compared to 10 out of 16 in the uncorrected case). We confirmed these trends again by recalculating susceptibilities for each of the gene sets, now using the GC-corrected null model (S2 and S3 Tables). We visualize these trends with another heatmap similar to Fig 3 which maintains the ordering of the columns and rows for ease of comparison (S1B Fig). Although the trends in susceptibility are similar to the non-GC-corrected case, the differences between vulnerable and resistant gene sets are reduced, suggesting that GC content explains a great deal, though not all, of the observed resistance to APOBECs.

Additionally, there is a question of whether the total number of nonsynonymous mutations should be minimized, or whether it is more appropriate to examine the nonsynonymous mutation fraction, which normalizes the number of motifs. There are two possible scenarios to consider for the frequency of mutation. In one scenario, a gene may only be exposed to the mutagen a few times during its activity. This is observed in B-cells, where AID is exposed to a gene possibly only once per cell cycle [8]. In this scenario, because the number of mutations is limiting, the probability that a mutagenic event causes an amino acid change is reduced by minimizing the fraction of nonsynonymous mutations, i.e. the number of nonsynonymous mutations divided by the total number of motifs. If a mutagen is exposed to a sequence only rarely, as in the case of AID, a coding sequence could minimize the probability of each mutagen changing its sequence by having many hotspots in the silent position. In the second scenario, mutations are not rate-limiting and the absolute number of hotspots causing nonsynonymous mutations would need to be minimized rather than the fraction. We found that tests conducted under the second scenario show similar results, albeit with slightly weaker trends (see section below “GC content of genomes predicts deaminase hotspot susceptibility“). Thus, our results are clearest in the first scenario (fraction of nonsynonymous mutations) assuming it is more indicative of susceptibility to mutagenesis. We conclude that examining the fraction is more instructive in determining trends (S2A Fig).

Additionally, as a control, we applied this model to different mutation profiles rather than the standard C to T. We chose every type of transition mutation (NNA to NNG, NNT to NNC, and NNG to NNA, which is complementary to NNC to NNT). In the cases of NNA to NNG and NNT to NNC (S2B and S2C Fig), we observe trends that are the opposite of our findings as described in the main text. Averaged across all 16 hypothetical NNA -> NNG hotspots, the susceptibility is lower in the oyster herpes virus by 0.18 (P = 0.011, 2 tailed t-test) relative to the housekeeping genes, though still prevalent but not as strong for the L1 elements (difference = 0.10, P = 0.24). Similarly, with all 16 NNT to NNC T to C mutations, L1 (difference = 0.12, P = 0.048) and oyster herpes viruses (difference = 0.198, P = 0.0063) show lower susceptibility to hotspots relative to human housekeeping genes. This suggests that mutagenesis at NNC to NNT hotspots is preferable for host genomes compared to these other patterns of mutation. We applied the same analysis as we did in Fig 3, to the hypothetical mutations of NNG to NNA (S2D Fig). Because we consider motifs on both strands, NNG to NNA is equivalent to CNN to TNN. For NNG to NNA we observed some similarity to the NNC to NNT case. Again there appear to be groups of resistant and vulnerable gene sets and these overlap somewhat with the NNC to NNT case, although the clusters are less clearly separated. If the average susceptibility of each gene set (rows in Figs 3 and S2D) were the key feature separating the vulnerable and resistant clusters, then we would expect similar clusters for both NNC to NNT and CNN to TNN. However, the clusters in each case are not the same suggesting that hotspot (column-) specific features that are unique to each cluster, are also important.

Although considering all 12 possible types of mutation is beyond the scope of this study, we did consider the three transition mutations other than NNC to NNT (S2B–S2D Fig) and one transversion (NNC to NNG, S2E Fig) that is directly comparable to Fig 3. Concerning the latter case, we found some similarities to the NNC to NNT transition case in that all control and retrotransposon gene sets clustered together in a predominantly vulnerable group. Clearly, some similarity is to be expected since most mutations at the third position of a codon are synonymous, and the NNC contexts are shared between the two cases.

GC content of genomes predicts deaminase hotspot susceptibility

When we analyzed susceptibilities using the total number of nonsynonymous replacements rather than the fraction of replacements, we observed greater variability across the 16 NNC motifs (S2A Fig). In spite of this, at the highest level the same two clusters of vulnerable and resistant gene sets emerge. However, under this assumption many of the gene sets in the resistant cluster, such as the mammalian housekeeping genes and the Epstein-Barr virus genome here show high vulnerability to GC rich motifs. These motifs include the APOBEC3G hotspot CCC, and the hypothetical hotspots GCC and GGC, all three of which cluster together. These gene sets simultaneously show resistance to the AT rich motifs ATC, TTC, TAC, and AAC, which are also clustered together. Varying GC content of the gene sets may explain these discrepancies. A gene with a very high GC content would contain more GC-rich motifs and fewer AT-rich motifs (and therefore a higher raw nonsynonymous mutation count) by chance. However, it is surprising that the trends are similar to our observed susceptibilities (Fig 3), since our definition of susceptibility examines only the nonsynonymous mutation fraction instead of raw count, “normalizing” for the GC content. Our analysis from here on thus examines the model of susceptibility that calculates the percentile of nonsynonymous mutation fraction compared to random sequences.

It is not immediately clear how GC content will also affect the nonsynonymous mutation fraction that is used to calculate susceptibility (Fig 3). However, a gene set with varying GC content may adopt a codon usage that inherently favors APOBEC hotspots at positions that cause synonymous mutations. We therefore compared the GC content of our vulnerable and resistant gene sets and found a strong statistical difference, with resistant gene sets having much higher GC content (Fig 4A, 2-tailed t-test, P <1.0⁻⁵). Furthermore, we observed that the correlation of susceptibility with GC content was strong for some motifs and not others. For many of the 16 motifs we examined, a higher GC content correlated strongly with low susceptibility to that motif as demonstrated by very negative correlation coefficients (Fig 4B). For example, the AT-rich hotspot TTC showed a strong correlation (Fig 4C), but for the GC-rich hotspot CCC the trend is not as apparent (Fig 4D). There is zero correlation between susceptibility of the motif GGC (which is not an observed AID/APOBEC hotspot) and GC content. Finally, to further corroborate these trends we have assessed the correlations between GC content and the alternate definition of motif susceptibility which considers the absolute number of nonsynonymous mutations rather the fraction. We observe again that a low GC content confers stronger susceptibility (negative correlation) to motifs of the WNC (W = A/T, N = any nucleotide) motif (Fig 4E), consistent with our findings to be discussed below (see section: “Retrotransposons and viruses show great discrepancy in resilience to biological deaminase hotspots.”). Our analysis suggests that GC content is strongly predictive of existing APOBEC hotspot susceptibility, with the exception of the human APOBEC3G hotspot CCC.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Low GC content is correlated to susceptibility to some NNC motifs.

A) Boxplots show the distribution of GC content of APOBEC-resistant (left) and APOBEC-vulnerable (right) gene sets. B) Plot showing the correlation coefficient (Y-axis) between our gene sets’ susceptibility to a particular motif on the bottom axis to its GC content. C) The correlation of a gene set’s mean susceptibility to the hotspot TTC, a GC-rich hotspot, to its GC content (red = resistant, green = vulnerable based on our clustering observations). D) Correlation of a gene set’s GC content to mean susceptibility for the hotspot CCC. E) Correlation between GC content of all gene sets and a measure of susceptibility based on the absolute number of nonsynonymous mutations at motifs. Susceptibility in this case is negatively correlated with motifs beginning with A and T, and more positively correlated with motifs beginning with G and C.

https://doi.org/10.1371/journal.pcbi.1005471.g004

Differences in mutational disruption of retrotransposons and viruses arising from canonical deaminase hotspots

The results shown in Fig 3 suggested that the current APOBEC hotspot preferences may have evolved to mutate particular targets while minimizing self-damage. To explore this further, we asked which of the 16 possible NNC motifs were best in discriminating host against parasitic (viral, transposable element) genomes. We proceeded by performing all pairwise comparisons for each gene set in the resistant cluster to each gene set in the vulnerable cluster, noting the differences in the susceptibility scores for each of the 16 NNC motifs. Since we observed 10 resistant gene sets and 12 vulnerable gene sets (Fig 3), this resulted in 120 individual comparisons. For each comparison the 16 NNC motifs were sorted by the difference in susceptibility of the resistant to the vulnerable genome (resistant—vulnerable). As an example, when we compared human housekeeping genes (resistant) to L1 retrotransposon elements (vulnerable) we discovered that the housekeeping gene sets had an average susceptibility to the motif ATC of 0.39, whereas the average susceptibility for the L1 elements was 0.872 (S4 Table). Thus, with a difference of -0.482, the APOBEC3A/B hotspot ATC shows the strongest contrast out of any of the 16 motifs and is ranked the highest, whereas the motif GAC has a difference of +0.321 and is ranked the lowest. The rankings suggest that ATC, rather than GAC, might evolve to be a preferable motif due to its capability for damaging L1 ORFs more so than impacting host genes.

By analyzing all 16 motifs across 120 comparisons, we found that the rankings of human APOBEC3B (NTC, N = any nucleotide) and AID (WRC, W = A/T, R = A/G) hotspots in particular, were significantly lower than expected by chance (see Methods, bootstrapped P = 0.0071 and P = 0.0002 respectively). We further calculated, for each motif, the average difference across all pairwise comparisons of vulnerable and resistant gene sets, again ranking each motif by the overall difference in susceptibility. For 9 out of the 16 motifs we examined, the susceptibility of the vulnerable genes was strongly lower than that of the resistant gene sets to a similar degree, a difference that was highly significant (t-test, P < 2.2 × 10⁻¹⁶ for all 9 motifs, Fig 5A columns shown in orange box). Curiously, these 9 motifs which demonstrate the greatest difference in mean susceptibilities also include all 8 motifs defined by the motif WNC (W = A/T). Although WNC itself is not necessarily a canonical preference seen in an existing cytidine deaminase, this set of 9 significant motifs still contains considerable overlap with existing cytidine deaminases, including all AID hotspots (WRC, where R = A/G), three of the four possible APOBEC3A/B hotspots (ATC, TTC and GTC), the human APOBEC3C hotspot (TTC) and the murine APOBEC3 hotspot (TCC).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. The naturally observed cytidine deaminase hotspots better distinguish vulnerable and resistant gene sets than hypothetical NNC motifs.

A) Boxplot distribution of differences in average susceptibilities between APOBEC-vulnerable and APOBEC-resistant gene sets. Canonical AID hotspots are in red, and APOBEC3A/B hotspots are in blue. The degenerate motifs WRC and NTC, are labeled in white. The top cluster of motifs (boxed in orange) are enriched in these motifs. B, C) Violin plots of separate gene families’ susceptiblities, clustered (left family in red, right in green), with protruding black lines representing the mean for that gene set. Motifs are sorted by the greatest difference in mean susceptibilities, with many of the motifs showing the greatest difference having the motif NTC or WWC (W = A/T).

https://doi.org/10.1371/journal.pcbi.1005471.g005

As an alternative method to confirm the significantly lower susceptibility of these 9 motifs (WNC and GTC), we compared the average rank of the 16 NNC motifs across all of our 120 pairwise comparisons, noting that the average rank of the 9 motifs were all statistically lower than expected by chance, whereas the average ranks of the other 7 motifs were higher than this (S5 Table, BH-corrected P < 10⁻⁴, one-sample MWU). Additionally, although it was not possible to include degenerate motifs such as WRC and NTC in our original analysis due to the non-independence with other observed motifs such as ATC, we did calculate susceptibility scores for these two motifs separately, as described next.

The 9 motifs we identified here as having lower susceptibility (Fig 5A) includes all 7 motifs that we previously indicated as distinguishing APOBEC-resistant and vulnerable gene sets based on the clustering of motif susceptibility (Fig 3, orange box). These 7 motifs cluster primarily as a result of their biological ability to distinguish vulnerable and resistant gene sets. Additionally, the two degenerate motifs of respectively AID (WRC) and hAPOBEC3B (NTC) showed large differences between vulnerable and resistant motifs as well (shown in white in Fig 5A). Both results suggest that the motifs that most strongly distinguish APOBEC-resistant and vulnerable genomes tend to be the existing hotspots, rather than hypothetical hotspots not associated with AID or APOBEC.

We observed that broadly, the vulnerable cluster (Fig 3, blue box) contains two qualitatively distinct families of genes, notably viruses and transposable element clusters. Similarly, the resistant cluster includes endogenous host genes such as the housekeeping gene set, and also viruses such as EBV and HSV-1, that have likely co-evolved with the APOBECs and may have acquired low APOBEC susceptibility as a consequence of this co-evolution. We assumed that the preferred APOBEC hotspots would be those that maximally damage their intended targets in intrinsic immunity, i.e. retrotransposons and vulnerable viruses, while minimizing damage to the host genome. With this assumption in mind, we looked more closely at two particular comparisons relevant to APOBEC evolution in humans. We first examined human housekeeping genes against human L1 elements, as the L1 elements are a major biologically significant target for APOBECs [39, 66]. Secondly, we compared two viruses: the Epstein-Barr virus (Human Herpesvirus 4) from the resistant cluster, against the Ostreid Herpesvirus, which we used as a control genome since its host is invertebrate and therefore not expressing any APOBEC genes, and is categorized in the vulnerable cluster.

In the first comparison (Human housekeeping vs Human L1) we found that the top 3 motifs in terms of susceptibility differences are all APOBEC3B TC hotspots: ATC, GTC, and TTC (Fig 5B, S4 Table). Interestingly, the observed difference does not arise because human housekeeping genes are particularly resistant to these motifs, since they have mean susceptibilities of 0.39, 0.51, and 0.43 respectively, which is close to that expected by chance (0.5). However, retrotransposons are particularly susceptible to these hotspots, with susceptibilities ranging from 0.87 to 0.91 for these three hotspots, where 1 represents maximum susceptibility. These high values suggest that human APOBEC3B hotspots may have evolved due to a particularly high capability of inducing nonsynonymous mutations in L1 elements, rather than low susceptibility to these hotspots in the host genome. We verified this effect for the AID hotspot WRC on some human non-housekeeping genes as well, by comparing the susceptibilities of our human housekeeping genes to a set of B-cell specific genes [67]. We used the list of B-cell specific genes listed in this study to obtain their coding sequences from the UCSC genome browser and calculated their susceptibilities, using the longest sequence if multiple isoforms were available for a single gene. For the B-cell specific genes, the susceptibilities to the 4 WRC hotspots were even lower than housekeeping genes: (AGC: difference = 0.10, P = 0.0075), (AAC: difference = 0.1, P = 0.000038), (TAC: difference = 0.12, P = 8.8 × 10⁻⁹), (TTC: difference = 0.07, P = 0.00313), suggesting that B-cell genes are under even more pressure to avoid deleterious mutations at AID hotspots than other genes.

In the second case (Fig 5C, S6 Table), we compared the APOBEC-resistant Epstein-Barr virus (EBV) to the Ostreid Herpesvirus (OsHV). The most highly ranked motifs in terms of susceptibility differences again included many A/T-rich motifs, with 8 out of the 9 top motifs again having the motif WNC. Furthermore, the two motifs showing the strongest difference are the hotspots ATC and TTC (collectively the motif WTC), which are human APOBEC3A/APOBEC3B hotspots. Many of the AID hotspots that EBV might be expected to evolve resistance to (given that EBV is a B-cell tropic virus) also show strong contrast, with the hotspots TAC and AAC ranked 4^th and 5^th and TGC and AGC being ranked 8^th and 9^th respectively. These differences, in contrast to the comparison between housekeeping genes and L1 elements, arise not only due to the high susceptibility of OSHV to motifs such as ATC (0.82) and TTC (0.83), but also due to EBV’s strong resistance, i.e. low susceptibility (0.21 for ATC, 0.277 for TTC). These low susceptibilities to APOBEC hotspots in EBV, together with our observation that AID/APOBEC hotspots again show the strongest contrast between resistant and vulnerable viruses, suggest that viral genomes such as EBV have adopted resistance to this particular form of mutagenesis.

In general, we note that motifs of the form WNC, and more specifically, WTC (W = A/T) appear to show the largest difference in vulnerabilities between housekeeping genes and retrotransposons, as well as between resistant virus and vulnerable genes. While WNC itself is not necessarily a canonical AID or APOBEC hotspot, it is a superset of the well-characterized AID hotspot WRC [1]. Furthermore, WTC overlaps with the hotspot NTC seen in human APOBEC3A/APOBEC3B. Thus, the fact that many of the motifs with the largest differences correspond to existing AID/APOBEC hotspots supports a hypothesis that the ability of AID/APOBECs to act effectively against viruses and retrotransposons while avoiding damage to the host is an evolutionary advantage for the host.

GC content and codon bias of genomes contribute to hotspot susceptibility

Many factors play an important role in shaping virus nucleotide evolution [68, 69]. We next describe how we investigated in particular the effects of codon bias, GC content, CpG dinucleotide frequency, selective pressure, and AID hotspot targeting of a hypothetical or actual virus sequence to validate our model.

We constructed new gene sets of 250 genes of completely random open reading frames that are 1500 nucleotides long, beginning with the methionine codon, ending with a stop codon, and with random non-stop codons in between. For the random codons, each nucleotide was chosen randomly, with either 40%, 50%, or 60% GC content. We calculated the susceptibility scores of these random sequences to validate susceptibility independent of evolutionary effects. Under uniform distribution of C and G (50% GC content), meaning each of the four bases is equally likely to be chosen, we observe that the median susceptibility to the canonical hotspots NTC (hA3B), CCC (hA3G), and WRC (AID) is 0.5 (S3A Fig). With random sequences of varying GC content, however, the susceptibility changes. Randomized sequences of low (40%) GC content showed high susceptibility to the hotspots NTC and WRC, and similarly, high GC content sequences showed low susceptibility (60%). We reaffirmed this by creating new randomized sequences of the resistant, high-GC content EBV and vulnerable, low-GC content OsHV, replacing each codon in each sequence with a synonymous one weighted by the GC parameter (a similar weighing description is given in the Methods under “Calculating Susceptibility”, though here it is applied to the baseline sequences rather than the null model) and observed similar trends, suggesting a strong role of wildtype GC content in explaining APOBEC resistance. We note, however, that this effect is not present for the hotspot CCC. We corroborated these effects by looking at the effect of GC content on codons for existing protein sequences. To determine the effects of native GC content on determining APOBEC hotspot resistance, we additionally calculated susceptibility under a GC-controlled null model (see Methods), weighing the GC parameter to be the mean GC content of the entire gene set. This meant that each gene sequence would now be compared to random synonymous sequences with a chosen GC content. Using this model, we found that changing the GC content of the random sequences affects the observed susceptibilities, reaffirming the importance of GC content in determining susceptibility. If a baseline (wild type) coding sequence is being compared to random sequences with higher GC content, the baseline sequence will tend to show higher susceptibilities to NTC and WRC hotspots (suggesting that lower GC content is more vulnerable to deamination). Despite these trends, the vulnerable OsHV sequence still has higher susceptibilities to NTC and WRC than EBV when GC is corrected (S3B Fig), although are not significantly so (NTC P-val: 0.18, WRC P-val: 0.34, BH corrected t-test). L1 elements, however, have their apparent high susceptibilities to NTC (the hotspot that hAPOBEC3B has which restricts them) neutralized when corrected for in this manner. Thus, high impact of these hotspots on L1 may be a beneficial effect of their inherent low GC content. There is little change in the susceptibilities of HIV to the NTC and WRC degenerate motifs (NTC P-val: 0.93, WRC P-val: 0.82) when we apply this correction, suggesting that while low GC genomes can be resistant to AID/APOBEC hotspots, there can be genomes for which their GC content does not contribute strongly to their vulnerability.

Additionally, we obtained codon usage tables of the coding sequences of OsHV, EBV, and our human housekeeping genes and L1 elements using Emboss cusp (http://www.bioinformatics.nl/cgi-bin/emboss/cusp). When we compared OsHV to EBV, we regenerated sequences in each genome using the codon usage table of the other and calculated the susceptibility of the new sequence. We observed that OsHV, when shuffled with codons with probability proportional to the codon usage bias of EBV, adopts a susceptibility much closer to that of EBV, and vice versa (S3C Fig). We observe similar effects when we exchange the codon biases of HK and L1 elements (two mammalian gene sets that we assume would be differentially targeted by APOBECs) and calculate susceptibilities (S3C Fig). We conclude that the nature of APOBEC hotspot susceptibility is also influenced by codon bias for both virus and vertebrate genes. However, note that the results described above using the GC-corrected model in effect creates a model with codon bias according to wildtype GC content. Thus, it is most likely that our results here using codon bias changes (S3C Fig) are correlated with those using GC-content changes (S3B Fig).

Additionally, codons within the same amino acid can be classified as “intra-codon” or “inter-codon” depending on whether they are a NNC with C to T at the silent position or not. For example, AGC is an intra-codon for serine, since AGC to AGT does not change its amino acid, but AGA would be an inter-codon. Statistical tests of all intra- and inter-codons comparing the vulnerable OsHV and resistant EBV show that OsHV has statistically fewer intra-codons than EBV (difference = 0.182, Fisher Test P < 2.2 × 10⁻¹⁶, 2-tailed Fisher test), providing additional evidence for the role of codon bias in hotspot susceptibility. Finally, our findings on codon bias were corroborated by the trinucleotide distribution of resistant and vulnerable gene sets (“Clustering algorithm”, under Methods).

Selective pressure was also modeled by taking the original EBV sequence, a resistant genome, and inducing mutations at different positions with probability weighed to either favor nonsynonyous or synonymous mutations, according to a parameter we use to approximate the dN/dS ratio. dN/dS represents the ratio of the number of observed mutations at nonsynonymous sites per site, to the number of mutations at synonymous sites per site. Specifically, mutations occurring in the first two positions of a random codon almost always represent nonsynonymous mutations and mutations in the third position are typically (but not always) synonymous. Therefore, we mutated positions along the EBV sequence with probability weighed by a parameter. This parameter is the ratio of mutation at the first two positions to the third, and acts under the assumption that each site is equally likely to be mutated and that this ratio of mutations is correlated to dN/dS. For example, if this parameter is equal to 2, then a mutation at each position of the codon is equally likely, since 2/3 of the mutations would be nonsynonymous, affecting one of the first two positions, and the remaining mutations would affect the third nucleotide, which is often synonymous. The corresponding value for the dN/dS ratio would be closer to 1, as there are twice as many nonsynonymous sites as synonymous. Under these assumptions, dN/dS is roughly half of our selective pressure parameter. For values of this parameter being 0.25 (red), 1 (green), 4 (blue), and 8 (purple) and for up to 500 mutations over time we calculated susceptibilities for all 16 NNC motifs (S3D Fig). Over time, averaged across all these motifs, the low susceptibility changes to a higher susceptibility, which may be explained by a change in GC content to neutral (or 50%) as each of the 12 mutation types (from one base to another) are equally likely. If dN/dS is high, however, the change is smaller. Indeed, the susceptibilities between the observed EBV sequences and EBV after 500 mutations when our selective pressure parameter = 8 (approx. dN/dS = 4) is not statistically significant except for the motifs GCC and GGC (GCC: P = 0.0004694, GGC: P = 0.0053; G2-tail t-test, BH corrected). The difference however between EBV after 500 mutations under high selective pressure (parameter = 8, approx. dN/dS = 4) and low (parameter = 0.25, approx. dN/dS = 0.125) is significant for many motifs including AAC (P = 0.0019), ACC (P = 0.0018), ATC (P = 0.0018), CGC (P = 0.0019), GAC (P = 0.025), TAC (P = 0.00039), TCC (P = 0.0377), and TTC (P = 0.0018). Based on our simulations, high selective pressure appears to play a modest role in maintaining APOBEC hotspot resistance. We note that many mammalian herpesviruses display resistance (Fig 3), and if APOBEC deamination is deleterious to survival of the virus, then there would be evolutionary pressure on the virus to change in such an environment. However, this may be mediated simply by changing the GC content of the virus.

Finally, there are other models of APOBEC/AID activity at hotspots. One of these is the S5F model of Yaari et al. [70] which inferred the mutability and substitution profile during somatic hypermutation among all 1024 possible 5-mers that targeted silent positions and which are therefore assumed to be independent of the effects of selection. Although somatic hypermutation includes not only AID activity, but also non-canonical Base Excision Repair and Mismatch Repair, the model is arguably the most comprehensive available for somatic hypermutation. To investigate the impacts of this activity on a gene’s susceptibility over time, we used this model and induce mutations at each position to the vulnerable OsHV and HIV genomes, resistant EBV genome, and random sequences, with probabilities proportional to the S5F model. This model includes the effects of base excision and mismatch repair which induce mutations beyond that which occur at hotspots, so we also simulate mutations using the same frequencies but only at C to T. We induce a number of mutations equal to up to 60% of the sequence length (to account for different gene lengths). Our results show that when we consider susceptibility to the AID hotspot WRC, AID targeting decreases susceptibility to WRC motifs (S4A Fig), highlighting the possibility that mutations at non-hotspot sites can decrease the susceptibility of a gene. In the case where we apply S5F, but only mutating C to T (and thus also G to A on the opposite strand), we mutate up to 60% of C and G positions instead per gene. Under this mutational model, that ignores the DNA repair mechanisms that act downstream of AID, we observe the opposite trend, namely that susceptibilities to the WRC motif instead increase as mutations accumulate (S4B Fig).

To extend this analysis beyond AID and somatic hypermutation, we also modeled the activity of two other APOBECs. We performed a simple simulation of APOBEC mutation at TC (hAPOBEC3B) and CCC (hAPOBEC3G) hotspots and then calculated the changes in susceptibility to those motifs accordingly. When we mutate the resistant EBV sequence at the hotspot NTC, susceptibility increases to a high level (S4C Fig). Interestingly, we did not see a change in susceptibility to the other hotspot CCC (S4D Fig) for all genomes that we examined. This suggests that regardless of the number of hotspots that are at nonsynonymous or synonymous positions, susceptibility to hotspots can be altered by GC content. The fact that CCC susceptibility, a hotspot which appears neutral to GC content (S3A Fig), does not change by this mechanism lends support for this idea.

Finally, the CpG dinucleotide has biological significance as a marker for DNA methylation and therefore gene activation and inactivation, especially in mammals [71]. We calculate another susceptibility measure that properly controls for this frequency (See: “Controlling for CpG dinucleotide content” under Methods). As shown in S5 Fig, when we consider this, most of the gene sets clustered similarly to the uncorrected method (Fig 3). As far as clustering of the motifs is concerned, we obtained one cluster of 6 motifs (rightmost dendrogram of S5 Fig), containing all 4 ANC motifs including two AID hotspots, AGC and AAC, which strongly overlaps with the cluster of motifs obtained without correcting for CpG (as described above for Fig 3, which contains 3 of the 4 ANC motifs, 3 of 4 TNC motifs, and GTC). These results suggest that the impact of these motifs is still similar when measuring susceptibility controlled for CpG. The gene set clusters also remain the same as the clusters without controlling for CpG content, with the exception of two formerly vulnerable gene sets, MHV68 and HERV which now cluster within the resistant gene sets. These gene sets, respectively a mammalian virus and set of retrotransposons, are co-expressed in environments alongside APOBECs, and in the case of HERV, is strongly edited by them [24]. Since accounting for CpG does not appear to affect the results very strongly, we do not control for this motif in the analyses, although this observation suggests that adapting CpG usage can be a mechanism of reducing effective susceptibility for a subset of viruses, particularly HERV.

Altogether our results suggest that resistance to APOBEC hotspots is a natural consequence of GC content, and provides an explanation of how, under selective pressure, a coding sequence may adopt resistance to APOBECs. These findings are consistent with the results described above (Fig 4) where we found a strong correlation between a genome’s GC content and its motif susceptibility. High GC content, for example, is a feature of several herpesviruses, and a previous study proposes that GC content may play a previous protective role against retrotransposon insertion in these viruses [65]. While we are not suggesting that gene sequences always adopt or evolve their GC content solely for the purposes of APOBEC resistance, we do show that GC content adjustment may confer additional defense against the consequences of restriction by most APOBECs, with the notable exception of the hAPOBEC3G hotspot CCC.

Calculating susceptibility

We describe a model that calculates the susceptibility, U_H, to motif H, from NNC to NNT, a wildtype coding sequence (denoted as s_wt), that contains L codons as follows. Given s_wt, we generate a vector: where c_i is the ith codon in s_wt. We generate a set W, of N (in practice 1000) random sequences: W = (s₁, s₂, … s_N) where each sequence s_j, in W is a sequence generated from s_wt by randomizing codon usage as follows. For each sequence s_j, for each codon c_i in C_L we replace the codon with another synonymous codon chosen either with uniform probability or a biased probability derived from a GC-bias parameter (see next paragraph). For each sequence s_j, the k occurrences of (real or hypothetical) hotspot motif H are counted. Then for each sequence s_j we associate with it a vector H_k = (h₁, h₂, … h_k), where h_t is the position of the t_th occurrence of hotspot motif H on sequence s_j, in both orientations. We define a function f(h_t) which is 1 if changing the nucleotide at position h_t from C to T (or G to A), causes an amino acid change, and is 0 otherwise. Then the nonsynonymous mutation fraction of sequence s_j is:

The susceptibility, U_H of wildtype sequence s_wt to the hotspot H, is calculated as:

Where R(l) = 1 if M(S_l) < M(s_wt), (M(s_wt) being the nonsynonymous mutation rate of the wild type sequence), and 0 otherwise.

Additionally, we may generate random sequences taking into account GC content, by assigning a weight to each synonymous codon. Given a GC content (fraction) of p, and a set of synonymous codons for a particular amino acid (e.g. GAG and GAA for Glu), a weight is derived for each codon based on the GC content of the codon. Within each codon, we assume that G and C nucleotides are chosen according to GC content with probability p, whereas A and T nucleotides are chosen with probability 1-p. The codon weight is the product of the three corresponding probabilities. Thus, for Glu, the codon GAG would be assigned a weight of p×(1-p)×p = (1-p)p², whereas the codon GAA would be assigned the weight p(1-p)². Codons are then selected with probability equal to a normalized vector of these weights. Note that if p = 0.5, each synonymous codon is selected with equal probability.

The effect of GC content on susceptibility

As described herein, susceptibility is impacted in many cases by GC content (Fig 4). We evaluated the larger gene sets on a gene-by-gene basis to assess the impact of GC content on susceptibility. These large gene sets include the human and mouse housekeeping gene sets and several virus genomes (S6 Fig). Although not every hotspot we examined demonstrated a strong relationship, the TTC motif had the strongest mean correlation among the 16 NNC motifs we examined, in the human and mouse housekeeping gene sets (Spearman correlation = -0.728 and -0.486 respectively, S6A and S6B Fig). For many of the viral genomes we analyzed, the GC content fell within a very narrow range and/or had very few genes, leading to weak correlations. We therefore grouped the genomes of EBV, HSV-1, Murid Herpesvirus 2, and the Ostreid herpesvirus, as well as many non-herpesviruses with 10 or fewer genes including HIV, Adenoviruses, and others to obtain a wider range of GC content across many genomes. By analyzing all the genes together we observed a trend similar to that for the housekeeping genes, where GC content of the gene correlates with the motif susceptibility (-0.493 for the TTC hotspot, S6C Fig). Because GC content influences susceptibility in some cases, but not others (Fig 4B), we chose in the first instance to keep our analysis neutral and we generated our null model without any constraints on GC, uniformly choosing each synonymous codon at random. We subsequently compared our susceptibility results to the GC content of the corresponding gene set to determine the extent to which GC content determines susceptibility (Fig 4B).

Controlling for CpG dinucleotide content

To control for CpG dinucleotide frequency we are interested in calculating susceptibility for each NNC motif conditioned on the CpG frequency in the original sequence. It is usually impractical to compute this conditional probability empirically. Controlling for this condition might be achieved by considering only those randomly generated sequences that have, by chance, produced the same number of CpG motifs as the original sequence. However, in practice the fraction of random sequences that exactly match the CpG motif count is often very small, particularly if the CpG content is already skewed, making it unlikely that such exact matches are produced by the random model. This may lead to very small numbers, making an approximation preferable. Thus, we approximated the joint distribution of CpG dinucleotide frequency and susceptibility, assuming the joint distribution is a bivariate normal distribution. The corrected susceptibility is simply the percentile of the approximated Gaussian conditional on the CpG frequency in the analyzed sequence. Correcting for CpG frequency made little difference to our results.

Data sources

We obtained protein-coding sequences from a variety of biological sources, and broadly speaking our data sources are categorized into three groups: Known or suspected AID/APOBEC targets including viruses and repeat elements; non-targeted viruses; and putative unintentional or collateral targets of AID/APOBEC which include host genes. Known virus targets that we analyzed included HSV-1, AAV-2, HERV (specifically the HCML-ARV retrovirus) and SFV. Since papillomaviruses are highly diverse we selected a well- characterized strain with medical relevance, which was HPV16 due to its link to cervical cancer. We also analyzed EBV, due to its shared B-cell tropism with AID. MHV68, a close relative of the other herpesviruses we studied, is not suspected to be restricted by mAPOBEC3 although multiple human APOBECs were capable of restricting the virus [63], so to test the possibility that mAPOBEC3 hotspots may have impacted murid herpesvirus coding sequence vulnerability or vice versa we included both Murid herpesvirus 2 and MHV68. As controls, we included several viruses unlikely to be restriction targets, due to not being vertebrate or mammalian viruses, which consist of the Ostreid herpesvirus, as well as two plant viruses, the cherry and strawberry viruses (Table 2). These controls were evaluated by substituting other potential controls consistent with the same rationale, i.e. they are not restricted or exposed to AID/APOBEC. We also analyzed a set of potential mutational targets in the mouse genome, namely a data set that is enriched in potentially oncogenic AID off-target genes [56] and sets of housekeeping genes from the human and mouse genomes (see below). Altogether we examined a total of 22 distinct gene sets. We wanted to avoid including an overwhelming number of controls, but at the same time we checked whether those we did include were unusual. Therefore, additional control viruses including plant viruses were obtained and individually verified by including these viruses in the clustering analysis (Fig 3) and determining whether they clustered consistently with the viruses originally used as controls. In each case these clustered with our previous control viruses, demonstrating the validity of the original choices (S7 Fig). Additional controls used to validate our current control set included: Tobacco mosaic virus (NC_001367), Grapevine leafroll associated virus 1 and 3 (NC_016509, NC_004667), Citrus tristeza Virus (NC_001661), Carrot yellow leaf Virus (NC_013007), and Fly C virus (NC_001834). Housekeeping genes for the vertebrate zebrafish, which lacks APOBEC3, and the invertebrate C. elegans were obtained from a RT-PCR array (http://www.sabiosciences.com/rt_pcr_product/HTML/PAZF-000Z.html) and a study of C. elegans housekeeping genes respectively [2, 80]. The list of human housekeeping genes were identified in a previous study [81]. In cases where a gene had multiple splice variants, we used the longest sequence available (UCSC Genes track/ knownGenes table via the UCSC genome browser query tool). For the set of mouse housekeeping genes, we used a previously published list which assigned probabilities of being housekeeping to each gene [82]. We used a threshold probability of 0.75 in order to obtain a gene set of comparable size to our human housekeeping gene set. Again, the longest sequence was used in case one gene had several coding sequences. L1 sequences for both mouse and human were downloaded from the UCSC genome browser database under the RepeatMasker table from the mm10 and hg19 databases respectively. Sequences were filtered with a minimum of 5500 base pairs, to ensure only full length elements, which are canonically around 6000 bp long. ORFs were extracted using EMBOSS getorf, with a minimum length of 1000 nucleotides finding all sequences between start and stop codons. These ORFs were separated based on their family and all ORFs were subsequently separated by family. In total 14 families were identified, including members from the L1PA and L1PB families identified in [52]. For each family, sequences were clustered. Representative sequences were processed to reduce statistical redundancies between repetitive sequences using cd-hit with default parameters [83]. The final L1 set was obtained by merging the representative sequences from each cluster.

Clustering algorithm

The data being clustered was a matrix of the average susceptibility scores, which range from 0 to 1, of all 16 NNC motifs in the columns, and gene sets (e.g. L1, Housekeeping genes, etc.) in the rows. Thus, both rows and columns were clustered using standard hierarchical agglomerative clustering and average linkage, using the Euclidean distance metric as the distances between each row and column. These clusters were computed and displayed using the heatmap.2 function in the R library gplots. We further validated these clusters using a k-means approach, showing that when we tested across a range of values for k, 2 clusters represents an optimal separation of the gene sets, indicated by the largest decline in intra-cluster sum of squares (S8A Fig). As an alternative approach, we used Principal Component Analysis (PCA). Here, one principal component accounts for a large amount of the overall trend in motif susceptibility (S8B Fig), and the clusters of gene sets are consistent for the different methods and separate primarily along the first principal component (S8C Fig).

Codon bias is likely correlated with the more general measure of trinucleotide (NNN) frequency. We confirmed this by calculating the trinucleotide frequency usage of all of our gene sets and clustering them similarly (S9 Fig). The occurrence of the trinucleotide AAA was skewed for the L1 elements for human and mouse, so it was excluded as an outlier. We observed the same clusters as before in Fig 3, except that Hepatitis B virus, and HERV, which under susceptibility are grouped with the vulnerable clusters that include L1 elements among others, now appears in the resistant cluster that includes housekeeping genes and mammalian viruses.

Statistical sampling of hotspot ranks

To assess statistically whether the rankings of certain groups of the 16 possible NNC (for example, the 4 human APOBEC3B NTC hotspots of ATC, CTC, GTC, and TTC) motifs were significantly high or low, we performed a pairwise comparison between each of the 10 resistant and the 12 vulnerable genomes of Fig 3. For each comparison, we ranked the motifs by the difference in average susceptibilities between the resistant and the vulnerable gene set (resistant-vulnerable), noting the ranks of the 4 motifs in each comparison. We assessed significance by sampling one resistant and one corresponding vulnerable gene set chosen at random, until one of each resistant gene set was sampled, and ensuring no gene set was sampled more than once. In each comparison we noted the ranks of the 4 motifs, summed them and compare the sum to a random null model selection of ranks, taken 4 at a time uniformly from the numbers 1 through 16 without replacement. For example, if examining the statistical significance of the motif NTC, the ranks of the hotspots ATC, CTC, GTC, and TTC, for a particular resistant vs vulnerable gene set, were 1, 3, 5, and 9, these would be compared to a random selection of numbers 1–16 which could be 3, 8, 9, 15. The sums of these ranks would be 18 and 35 respectively. We then summed the ranks of the 4 hotspots across all our comparisons and compared this to the sum of null model ranks. Repeated 10000 times, our bootstrapped P-value was the fraction of times the sum of observed ranks found to be greater than the sum of the random ranks. If this fraction (equivalent to a P-value) were greater than 0.05, this demonstrated that the observed ranks were not significant.