Signatures of Mutational Processes in Human DNA Evolution

Home , Mutagenesis

bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

Hamid Hamidi1, Hamid Alinejad-Rokny2,3,4, Tim Coorens5, Rashesh Sanghvi5, Sarah J Lindsay5, Raheleh Rahbari5, Diako Ebrahimi*6

1Department of Mathematics & Statistics, University of Calgary, Calgary, AB T2N 4N1, Canada 2Systems Biology and Health Data Analytics Lab, The Graduate School of Biomedical Engineering, UNSW Sydney, Sydney, NSW, 2052, AU 3School of Computer Science and Engineering, UNSW Sydney, Sydney, NSW, 2052, AU 4Health Data Analytics Program, AI-enabled Processes (AIP) Research Centre, Macquarie University, Sydney, 2109, AU 5Wellcome Sanger Institute, Hinxton, Cambridgeshire, CB10 1SA, UK 6Biomedical Informatics Lab, Texas Biomedical Research Institute, San Antonio, TX, 78227, USA

*Corresponding author: [email protected]

Abstract The human genome contains over 100 million SNPs, most of which are C/T (G/A) variations. The type and sequence context of these SNPs are not random, suggesting that they are caused by distinct mutational processes. Deciphering the mutational signatures is a crucial step to discovering the molecular processes responsible for DNA variations across human populations, and potentially for causing genetic diseases. Our analyses of the 1000 Genomes Project SNPs and germline de novo mutations suggest that at least four mutational processes are responsible for human genetic variations. One process is European-specific and no longer active. The remaining three processes are currently active in all human populations. Two of the active processes co- occur and leave a single joint mutational signature in human nuclear DNA. The third active process is specific to mitochondrial DNA, and inflicts C-to-T mutations at mostly non-CG sites. We found neither evidence of APOBEC-induced cytosine deamination in the human germline, nor de novo mutation enrichment within certain regions of the human genome.

Introduction Variations observed in the human genome are signatures of contemporary and historic mutational processes. A systematic analysis of these genetic variations can help shed light on the underlying mutational processes and their potential roles as risk factors for diseases1. There is currently no consensus on the type and proportion of mutational signatures in human populations. Comparison of the 1000 Genomes Project2 SNPs in different populations revealed an elevated rate of TCC>TTC variations among Europeans3,4. However, analysis of the same dataset using hierarchical clustering showed three distinct patterns of mutation enrichment among populations5. Moreover, a recent study in which genomic coordinates of rare mutations were taken into consideration revealed 14 unique patterns of mutations in the TOPMed dataset6. Unexpectedly, few prior studies report signatures of germline mutations induced by proteins of our innate immune system known as APOBEC (Apolipoprotein B mRNA editing enzyme, catalytic polypeptide)7–9. Members of this enzyme family such as APOBEC3A and APOBEC3B are considered to be the primary sources of cancer mutational signatures known as SBS2 and SBS1310–12. A study of rare SNPs in the 1000 Genomes Project reported that APOBEC3A and APOBEC3B were likely responsible for 20% of the germline C>T and C>G mutations within TCW (W:T, A)8. Pinto et al. reached a similar conclusion but suggested that a cytoplasmic and antiviral member of the APOBEC3 family, APOBEC3G, was the source of germline mutations and primate evolution7. Other studies reported clusters of C>G mutations but ruled out a role for APOBEC3 because the context of mutations was not enriched for TCW13,14,15. The methodologies used in these studies often assume that each mutation type (e.g. TCC>TTC) is generated by only one mutational process. In other words, different mutational processes cannot generate the same mutation. Studies of cancer mutations suggest that this assumption is often incorrect. For example, CCC>CAC can have several different sources, such as defective homologous recombination-based DNA damage repair (SBS3), tobacco smoking (SBS4), or even exposure to aflatoxin (SBS24)11. An alternative approach used to identify mutational signatures without making this assumption is based on the deconvolution of mutational datasets10,11,16,17. Recent quantitative analysis approaches, based on non-negative matrix factorization (NMF)18 of complex tumor mutations, have been able to decipher signatures of many mutational processes in cancer10,16. In these studies, each mutational signature is defined by a measure of the relative frequency of 96 mutation types (nNn>nMn, where N is mutated to M and n: A, C, G, or T). These studies have led to invaluable findings about the mechanisms underlying mutations in cancer such as spontaneous deamination of 5-methylcytosine (SBS1), APOBEC mutagenesis (SBS2 and SBS13), and homologous recombination-based repair deficiency due to BRCA mutations (SBS3)10,11,19. NMF and similar mutation deconvolution approaches have also been used recently to gain insight into mutational processes driving human DNA evolution6,17. For example, Mathieson et al. used independent component analysis followed by NMF to deconvolute signatures of rare mutations. By analyzing two independent datasets, high coverage sequences of 300 individuals from the Simons Genome Diversity Project and phase 3 of the 1000 Genomes Project, the authors discovered four mutation signatures17. A study of multi-sibling families revealed that the spectrum of de novo germline mutations can be explained by only two signatures, SBS1 and SBS5, which are found ubiquitously in cancer20. SBS1 is referred to as ‘aging’ signature and is thought to be a signature of CG (a.k.a. CpG) bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

methylations followed by spontaneous deamination10,21. This signature is characterized by four C>T mutation peaks at CG sites. Signature SBS5, which is also associated with aging, has an unknown etiology and is characterized by C>T and T>C mutations, with the latter showing enrichment for ATA, ATG, and ATT motifs. In this study, we performed a systematic analysis of population and trios datasets using four independent approaches, namely, PCA (principal component analysis), NPCA (nonnegative PCA)22, NMF (non-negative matrix factorization), and HDP (hierarchical Dirichlet process)23–26 to identify signatures of present and past mutational processes in different human populations. Our analyses revealed that present-day de novo germline mutations can be described using a single mutational signature, which is population independent. We also identified a European-specific signature reported previously3–5,17,27,28 and a novel mutational signature that is specific to mitochondria DNA.

Results Deciphering mutational signatures 1000 Genomes Project Dataset We quantified the abundance of all 96 types of mutations (nNn>nMn) in the 1000 Genomes Project donors by comparing them to the most common human ancestral sequence and choosing the low frequency SNPs (<0.01) (Method). PCA applied to this raw mutation data matrix (2,504 x 96) showed four PCs (Supplementary Fig. S1). Fig. 1 shows the distribution of populations in PC1- PC2 and PC3-PC4 spaces. bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

Fig. 1 PCA applied to raw mutation data matrix (A) PC1-PC2 (B) PC3-PC4

Analysis of rare SNPs using NPCA revealed four significant principal components (i.e. mutational signatures), explaining ~70% of total variance (Supplementary Fig. S2). The first principal component is characterized by, mostly, C>T and T>C mutations (Fig. 2A). This signature (1000G- Sig1) has four main C>T peaks at CG motifs (ACG, CCG, GCG, and TCG) and one C>T peak at ACA. Additionally, it has three main T>C peaks at ATA, ATG, and ATT. The second mutational signature (1000G-Sig2) is represented by multiple C>T peaks, the top three being at motifs ACC, TCC, and TCT (Fig. 2A). The third mutational signature (1000G-Sig3) has five main peaks of C>A at ACA, C>T bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

at TCT, T>A at ATA, and T>G at GTG and GTT (Fig. 2A). The fourth mutation signature (1000G- Sig4) is represented by multiple C>G and C>T peaks mostly at non-CG sites and four main T>C peaks at ATA, ATG, ATT, and TTT (Fig. 2A). Analysis of the same mutation matrix using NMF also returned four mutational signatures (Supplementary Fig. S3). Each NMF mutational signature shows its unique set of peaks, which are identical to those of NPCA signatures after a background correction (Method). The cosine similarities between the mutational signatures obtained using NMF and their corresponding signature in NPCA were >0.82 (Supplementary Fig. S4). Analysis using the HDP method also returned the first three mutational signatures (1000G-Sigs 1-3) with cosine similarities >0.93 (Supplementary Fig. S5). This method was unable to detect 1000G-Sig4. Altogether, the results obtained from these four independent analyses confirmed the existence of at least three mutational signatures in the 1000 Genomes Project dataset. Fig. 2B shows that Africans have the lowest level of 1000G-Sig1 (p<0.0001 pairwise Mann- Whitney U test Bonferroni-corrected) and the highest level of 1000G-Sig4 (p<0.0001). The signature 1000G-Sig2 has the highest level in Europeans (p<0.0001). Compared to other three signatures, there is less between-population differences in the level of 1000G-Sig3. Nevertheless, in some individuals, except South Asians, the proportion of this mutational signature is over two- fold higher than average (Fig. 2B). bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

Fig. 2. Deconvolution of nuclear DNA mutational signatures in the 1000 Genomes Project dataset using NPCA. (A) Mutational signature 1-4; (B) Coefficient (i.e. weight) of mutational signatures in different populations. Significantly different populations with a Bonferroni- corrected (p<0.01) are denoted by different letters above the horizontal axis.

Trio Datasets Analysis of de novo mutations from three trio datasets20,29,30 using NPCA showed only a single mutational signature (DN-Sig, Fig 3). Same results were found using PCA, NMF, and HDP (Supplementary Figs. S6, S7, and S8). This de novo mutational signature is almost identical (cosine similarity = 0.997) to the mutational signature 1000G-Sig1.

Fig. 3. Deconvolution of nuclear DNA de novo mutations from two European trio datasets and one dataset containing samples with different ancestry using NPCA. (A) A single mutational signature in Icelanders dataset13,29; (B) A single mutational signature in a dataset containing donors from different populations30; (C) A single mutational signature in data from Rahbari et al.20

Analysis of mitochondrial DNA 1000 Genomes Project Dataset Typically, in de novo mutation studies, mitochondrial DNA and Y chromosome are excluded from the analyses13,20. To investigate if, compared to other chromosomes, chrY and mtDNA have a different mutation spectrum, we separated the 1000 Genomes Project mutation datasets by chromosome and computed cosine similarities between each two chromosomes and also between all chromosomes and mtDNA (Fig. 4A). The results indicate that all nuclear chromosomes have comparable mutation profiles; however, mtDNA is distinct. Analysis of mtDNA mutations using PCA, NPCA, and NMF returned a single mutational signature (Supplementary Figs. S9, S10, and S11). This signature (1000G-Sig-mt) is dominated by C>T and T>C mutations; however, compared to nuclear DNA, there is significantly less CG>TG and TG>CG mutations (Fig. 4B). To investigate if this is due to lower representation of CG and TG in mtDNA compared to nuclear DNA, we performed a motif representation analysis. As indicated, bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

compared to nuclear DNA, mtDNA has a significantly higher level of CG and lower level of TG (Fig. 4C).

Fig. 4. Analysis of mitochondrial DNA in the 1000 Genomes Project dataset using NPCA. (A) Heatmap of similarities between the mutation spectra of chromosomes and mitochondrial DNA; (B) A single mutational signature present in mtDNA (1000G-Sig-mt); (C) Represntation of CG and TG dinucleotides in chromosomes and mitochondrial DNA.

Trio Dataset The number of de novo mutations in mitochondrial DNA is very limited (average of ~18 mutations per donor). Using the available data (three European trios)20, it was not possible to perform a deconvolution analysis. However, a sum of mutations in this trio dataset shows a cosine similarity of 0.78 to the 1000G-Sig-mt (Supplementary Fig. S12)

Analysis of similarities between germline and somatic mutational signatures When compared, to the COSMIC v3 cancer mutational signatures11,31, each of the five 1000G Signatures (4 nuclear DNA signatures and one mtDNA signature) and de novo signature (DN-Sig). For each germline mutational signature, we observed a cosine similarity of >0.8 to at least one of the recently updated cancer mutational signatures (COSMIC v3) (Supplementary Table S1). The bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

1000G-Sigs 1-4 and 1000G-Sig-mt show the greatest correlations with cancer signatures SBS5, SBS7a, SBS51, SBS5, and SBS12, respectively (Supplementary Table S1). Using the expectation maximization (EM) algorithm (Method) we decomposed each of the 1000G-Sigs1,2,3,4,mt signatures into combinations of cancer signatures (COSMIC v3) (Fig. 5). Table 1 shows the optimum combinations after the removal of cancer signatures with <10% contribution (Method). The optimum combinations for 1000G-Sigs1,2,3,4,mt are SBS5+SBS1 (Supplementary Fig. S13), SBS11+SBS7a (Supplementary Fig. S14), SBS47+SBS43+SBS51 (Supplementary Fig. S15), SBS39+SBS16+SBS37 (Supplementary Fig. S16), and SBS12+SBS30 (Supplementary Fig. S17), respectively (Table 1).

Table 1. Combinations of cancer signatures (COSMIC v3) best describing the 1000G and de novo mutational signatures

COSMIC V3 combination Original Sigs Cosine Similarity 0.13*SBS1+0.87*SBS5 1000G-Sig1 1 0.36*SBS7a+0.64*SBS11 1000G-Sig2 0.9 0.26*SBS43+0.58*SBS47+0.16*SBS51 1000G-Sig3 0.83 0.29*SBS16+0.23*SBS37+0.48*SBS39 1000G-Sig4 0.75 0.68*SBS12+0.32*SBS30 1000G-Sig-mt 0.92 0.15*SBS1+0.85*SBS5 DN-Sig 0.99

Fig. 5. Comparison of germline and COSMIC cancer mutational signatures. The expectation maximization (EM) algorithm was used to determine the combinations of cancer reference signatures (COSMIC v3) that best model each of the five 1000G signatures (1-4 and MT) and DN- Sig. Color intensities in this heatmap represent the contribution of cancer signatures in each germline signature.

Table 1 shows high cosine similarities between 1000G-Sigs and combinations of cancer signatures. Therefore, we investigated if such high similarities could be achieved when distinct cancer mutational signatures (individually or in combination) are compared to each other. Our results showed that cosine similarities between cancer mutational signatures can be as high as 0.99 (Supplementary Table S2).

Enrichment analysis of de novo mutations within endogenous elements Studies suggest that processes such as CpG methylation32–34 and APOBEC-induced cytosine deamination35–37 play a role in silencing endogenous retroelements. To determine if such processes leave a signature on these elements, we performed an enrichment analysis. We quantified the percentage of mutations occurring within different repeat element (LINEs, SINEs, ERVs, and DNA) and non-repeat regions (e.g. genic regions) for all 1,548 Icelander samples. We first quantified the percentage length of each repeat element class in the human genome to use as a benchmark for mutation enrichment. If de novo mutations are enriched within a certain element class, we expect the percentage mutation to be significantly higher than the percentage length of that element. As indicated in Fig. 6, the percentage of de novo mutations in different regions of the human genome is proportional to the percentage length of the regions. In other words, de novo mutations are not enriched within certain repeat element classes (Pearson correlation coefficient = 0.998, p<2.2e-16). However, we note that de novo mutations are slightly depleted in the centroid satellite sequences.

Fig. 6. Analysis of de novo mutation enrichment within endogenous elements. Percentage of de novo mutation within different endogenous elements is shown in red. The total percentage length of each element is shown in blue. Ychr and mtDNA were excluded.

Discussion Identification of germline mutational signatures is the first step to discovering the molecular processes responsible for human DNA evolution, their differential triggers in different human bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

populations, and their potential roles in diseases such as cancer. By deconvoluting human genetic variation datasets, we uncovered the mutational signatures imprinted in the human genome by present and past mutational processes. Our analyses show only a single mutational signature (DN-Sig) in germline de novo mutation datasets20,29,30 This signature is population-independent and likely represents two processes occurring concurrently during gametogenesis and/or early embryogenesis. As a result, their mutational signatures are inseparable by deconvolution methods. One of these processes inflicts C>T within CG (a.k.a. CpG) dinucleotides38–44 and the other induces T>C within mostly ATA, ATG, and ATT trinucleotide motifs20. The former pattern is identical to the cancer mutational signature SBS110,11,31 and represents methylation-induced deamination of CpG cytosines21,46,47. The latter closely resembles SBS5 whose etiology remains unknown10,11. SBS1 and SBS5 have also been identified in normal cells48 and they show a clock- like behavior happening spontaneously throughout life10,11,45. Our data indicate that the CG depletion caused by de novo C>T mutations is not fully compensated for by de novo T>C mutations, which mostly occur at non-TG motifs. Specifically, at each birth, on average, ~10 CGs (a.k.a. CpGs) are removed from and ~4 CGs are introduced to the human nuclear DNA by C>T and T>C mutations, respectively. Assuming that mechanisms such as GC-biased gene conversion49 have negligible compensatory impacts, these data suggest that human genome will continue to lose ~6 CG dinucleotides per birth. De novo mutations are signatures of present-day mutational processes, thus cannot be used to investigate the mutational processes acting in the past. To overcome this limitation, we analyzed the human genetic variation data reported by the 1000 Genomes Project3,4,8,17 and identified four distinct mutational signatures in nuclear DNA, and one in mitochondria. The first signature (1000G-Sig1) is almost identical to the de novo mutational signature, implying that one can use the 1000 Genomes SNPs to infer de novo mutation signatures, without analyzing trio datasets. The proportion of this signature is slightly lower in Africans compared to other populations. The second signature (1000G-Sig2) is composed, almost exclusively, of highly context-specific C>T mutations (TCC>TTC, TCT>TTT, ACC>ATC, and CCC>CTC) and is significantly more prevalent in European populations. Previous studies have identified most of these trinucleotide motifs within which C>T changes are enriched in Europeans3–5,17,28. Our matrix deconvolution approaches precisely identified these mutation types and grouped them into a distinct mutational signature. Our analysis of present-day trios did not show this mutational signature, confirming that 1000G- Sig2 has an ancient source3,4,27,28. The signature 1000G-Sig2 can be modelled by a linear combination of two cancer signatures SBS7a and SBS11. The former is associated with skin cancer and exposure to UV light, and the latter resembles the mutations caused by alkylating agents. UV-mediated degradation of folate leading to defects in DNA synthesis were suggested to be a potential source of germline mutations in Europeans with light skins3. Mathieson et al. argued against this hypothesis based on two main lines of evidence17. First, ACC>ATC is not enriched in the genomes of other lightly pigmented populations such as Siberians and Northeast Asians. Importantly, our data also do not show an enrichment for 1000G-Sig2 in East Asians. Second, UV irradiation is known to induce CC>TT mutations, but this type of mutation is not enriched in West Eurasians17. With respect to alkylating agents, there is no evidence to suggest a historical exposure of Europeans to these chemicals, although it is suggested by Mathieson et al.17 Taken together, a high similarity between 1000G-Sig2 and SBS7a+SBS11 may not be sufficient evidence bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

to suggest exposure to UV radiation and alkylating agents have played a role in human genome diversification. All of the three cancer signatures SBS47, SBS43, and SBS51 predicted to be related to 1000G-Sig3 are currently deemed to be sequencing artefacts. Additionally, this signature contains all of the four peaks of GTN>GGN (NAC>NCC), which were first reported as a Japanese-specific signature4,5, but later found to be a batch effect50. These data suggest our analyses may have resolved a systematic sequencing error present in the 1000Genome Project dataset. The mutational signature 1000G-Sig4 may also be an artefact of the analysis. This signature was found by PCA, NPCA, and NMF, but not HDP. Additionally, compared to other signatures, it shows the lowest coefficients and model accuracy when reconstructed using a combination of cancer signatures. These suggest further analyses are needed to determine if this signature represents a real mutational process. Nevertheless, 1000G-Sig4 exhibits a T>C profile that resembles SBS16 (and SBS5), and a C>T profile that is highly depleted for CG>TG mutations. This is the opposite of de novo signature DN-Sig in which T>C mutations are always accompanied by C>T mutations at CG sites. Assuming that 1000G-Sig4 is not an artefact, these data suggest that T>C mutations might have happened independently of CG>TG mutations in the past, particularly in Africans with an elevated level of 1000G-Sig4, but these two mutation types co-occur in the present time. We found that compared to nuclear DNA, mtDNA has a distinct mutational signature featured by C>T changes at mostly non-CG sites and T>C changes at mostly non-TG sites. To investigate if the lower rates of CG>TG and TG>CG are due to the low abundance of CG and TG in the mtDNA compared to nuclear DNA, we performed a motif representation analysis. As indicated in, TG dinucleotide is ~3 folds less abundant in mtDNA, suggesting that lower frequency of TG in mtDNA may be a player in reduced TG>CG level. Nevertheless, further studies are needed to confirm this hypothesis. As opposed to TG, the representation of CG is significantly higher in mtDNA (~4 folds), implying that the low rate of CG>TG mutation in mtDNA is due to another mechanism. One hypothesis could be that compared to nuclear DNA, the rate of CG methylation (followed by spontaneous deamination) is lower in mtDNA. In support of this hypothesis, a recent study by Patil et. al. provides evidence that methylation of human mtDNA mostly occurs at non-CG sites51. Importantly, the majority of somatic mutations identified in mtDNA from tumor samples are also predominantly C>T and T>C changes52. But the sequence contexts of these somatic mutations are very different from germline de novo mutations. For example, contrary to de novo germline mutations, which are depleted for CG>TG mutations, these are the most frequent mtDNA mutations in cancer52. These data suggest that the etiologies of germline and somatic mtDNA mutations are different. Silencing of retroelements using mechanisms such as methylation and APOBEC-induced deamination are well documented32–37,53,54. Methylation-induced deamination of cytosines in these elements during gametogenesis, fertilization and/or in early embryogenesis can be a potential source of de novo mutations. Such targeted mechanisms would lead to a de novo mutation enrichment within retroelements. However, our systematic analysis of mutations in all main classes of endogenous elements did not show signs of mutation enrichment in any element classes. These data suggest that epigenetic silencing of retroelements is not likely a player in the observed pattern of de novo mutation. bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

Altogether, human polymorphism datasets contain at least three mutational signatures that are likely associated with biological processes. One of these signatures is currently active and appears to represent two co-occurring mutational processes, methylation-induced deamination and an unknown process. The second signature, which is elevated in the genome of Europeans, is no longer active3,4,27. The third signature is specific to mitochondrial DNA and represents mutational process(es) that inflict C>T and T>C changes but mostly at non-CG and non-TpG sites. Each of these three signatures (and also the other two signatures that might be artefacts) show unexpectedly high cosine similarities with combinations of COSMIC cancer signatures. This may indicate, at first glance, that germline mutations are simply a combination of two or more processes already identified in cancer. Our analyses show that cancer mutational signatures can also be modelled, with cosine similarities of up to 0.99, using combinations of other cancer signatures. This implies a high similarity between a germline signature and a combination of known cancer signatures cannot always be used to infer mutational processes.

Methods Data In this study we used four independent datasets: a) Phase 3 of the 1000 Genomes Project dataset, which includes whole genome SNPs of 2504 individuals2, b) whole-genome de novo mutation of 1,548 trios from Iceland29, and c) whole genome de novo mutation dataset of three multi-sibling families20 d) whole-genome de novo mutation dataset of 816 trios with different ancestry30. For each dataset, we performed a separate analysis using all mutations except those from mtDNA DNA and Y chromosome for which only had access to data from the first and third datasets. For the analysis of mutation distributions, we used the Repeatmasker dataset to determine the location of various DNA elements such as LINE, SINE, ERVs, etc55.

Analysis of 1000 Genomes SNPs We used the FASTA file of the most common ancestral sequence (MCA) (http://ftp.ensembl.org/pub/release-74/fasta/ancestral_alleles/) as a reference and compared each of the 2,504 individuals to this reference sequence using available VCF files (http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/). Using this method, we identified, for each individual and each polymorphic position, the type of mutations with respect to the ancestral sequence. For example, if a position in the ancestral FASTA file is T, and the studied individual has a G in the same position, there has been a T>G mutation in this position. Considering all possible 3’ and 5’ flanking nucleotides of a mutated base, there would be a total of 192 types of mutations (e.g. AAA>ACA, AAA>AGA, …, TTT>TGT). This number would be halved if we combine each mutation with its reverse complement (e.g. ACT>ATT and AGT>AAT). We identified the trinucleotide context of each of the mutated positions described above and created a matrix of 96 x 2,504. Each cell in this matrix represents the total number of one of the 96 type of mutations in one individual. In our analyses, we did not include the SNPs for which there was no information on the type of nucleotides in the MCA FASTA file. To avoid the potential effect of selection, previous studies have focused on rare variants17. Previous studies have also removed singletons, which are mutations occurring in only one sample out of the entire population and might represent cell line and/or somatic mutation artifacts17. We removed, from our analyses, all bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

singletons and all SNPs with an allelic frequency >1%. We used four independent analyses, namely PCA, NPCA, NMF, and HDP on the matrix of mutations, which included all chromosomes except chromosomes Y and mitochondria DNA. We performed separate analyses for mtDNA. For parsing the VCF and FASTA files we used the “pyvcf56” and “biopython57” libraries in Python. We noted that the number of all 96 mutation types were 2-4 folds higher in African populations compared to other populations (Supplementary Fig. S18). Therefore, we normalized the data such that within each donor, each mutation is presented as a percentage of total number of mutations. The NPCA, NMF, and HDP analyses were done on this normalized matrix of mutations. Analysis of de novo mutations in trios Using the reference parental alleles, the derived alleles of the offspring, and the human reference sequence GRCh38 for Icelandic trios and GRCh37 for 816 trios with different ancestry, we identified the type and sequence contexts of all de novo mutations and created a matrix of size (96 mutation type × number of individuals). We then performed NPCA, PCA, and NMF as described below on each dataset. We also analyzed the mixed-matrix of all datasets consisting of 2364 individuals using HDP.

Deconvolution methods PCA – The PCA analyses were performed using the ‘prcomp’ function in R58, which uses a singular value decomposition algorithm. NPCA –The NPCA analyses were performed using the ‘nsprcomp’ package22 in R58, and only the non-negativity constrained was used to perform PCA. We performed NPCA with 100 random initializations to avoid local minima solutions. NMF – Alexandrov et al. developed a six-step mutation deciphering method, which has a non- negative matrix factorization step to identify signatures of mutational processes in cancer10,16,59. We used these steps to identify mutational processes in de novo mutations from three trio datasets. We used both signature reproducibility and reconstruction error measures to determine the optimum number of mutational signatures. For the analysis of 1000 Genomes dataset, we used the elbow point of explained variance to identify the number of mutational processes60,61. We used the “NMF” package60 in R programming language58 with 100 random initialization (to avoid local minima) and “brunet” algorithm62 to perform NMF. We noted that the NMF signature 1 peaks appear as background in all other three signatures 2, 3, and 4. To remove this background, we subtracted NMF signatures 2, 3, and 4 from NMF signature 1, and then determined the cosine similarities between this background-corrected NMF signatures and NPCA signatures. HDP – To extract mutational signatures, we ran an algorithm based on the hierarchical Dirichlet process (HDP)23 (https://github.com/nicolaroberts/hdp) on the 96 trinucleotide counts. HDP was run with individuals or - where available - the population as the hierarchy, in twenty independent chains, for 40,000 iterations, with a burn-in of 20,000.

Signature reconstruction using COSMIC reference signatures - The resulting 1000G-Sigs were decomposed into optimal combinations of cancer signatures (COSMIC v3) using expectation bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

maximization (EM) algorithm24. First, the EM algorithm was performed on all COSMIC signatures. Afterward, EM algorithm was performed on signatures with contribution of more than 10% to avoid overfitting. The EM algorithm finds the optimal linear combination of chosen COSMIC signatures that reconstructs our 1000G-Sigs. The cosine similarity between reconstructed signatures and original signatures was computed.

Analysis of motif representation We observed that the mutation profile of mitochondria DNA was different from that of nuclear DNA. To investigate if this difference is due to the genetic makeup difference between nucleus and mitochondria, we quantified the ratio of observed frequency of each dinucleotide motif over its expected frequency. The observed frequency is, simply, the proportion of a given motif in the DNA. The expected frequency of a dinucleotide motif is calculated by multiplying the constituent mononucleotides63.

Analysis of de novo mutation enrichment within DNA elements We investigated the potential enrichment of mutations within major endogenous elements, LTR (ERV1, ERVL, ERVK), LINE (L1 and L2), SINE (Alu and MIR), DNA, and/or other minor classes. We used Repeatmasker to download the genomic coordinates of these elements, available at (http://www.repeatmasker.org/genomes/hg38/RepeatMasker-rm405- db20140131/hg38.fa.out.gz). We quantified the percentage of the length of each element class in the human reference genome, excluding chrY and mtDNA. We determined the mutation burdens by quantifying the number of substitutions within each element class listed in the repeat masker file, summed over all 1,548 probands. These numbers were divided by the total number of de novo mutations to calculate the percentage of mutations within each element class. We used the Pearson's product moment correlation coefficient in R to investigate the statistical difference.

Analysis of mitochondrial DNA de novo mutations Mutations were initially called using bcftools mpileup using the 1000Genomes_hs37d5 as reference. Variants with a quality of less than 20 and a depth of greater than 100 reads were excluded. Sequencing of these individuals is described in Rahbari et al.20

Acknowledgement The authors thank Mrs. Rachael Rodriguez for her scientific comments and proofreading the manuscript.

Funding sources D.E. is funded by The William and Ella Owens Medical Research Foundation. R.R. is funded by Cancer Research UK (C66259/A27114).

References

1. Shendure, J. & Akey, J. M. The origins, determinants, and consequences of human mutations. Science (80-. ). 349, 1478–1483 (2015). 2. Consortium, 1000 Genomes Project. A global reference for human genetic variation. Nature 526, 68 (2015). 3. Harris, K. Evidence for recent, population-specific evolution of the human mutation rate. Proc. Natl. Acad. Sci. 112, 3439–3444 (2015). 4. Harris, K. & Pritchard, J. K. Rapid evolution of the human mutation spectrum. Elife 6, e24284 (2017). 5. Aikens, R. C., Johnson, K. E. & Voight, B. F. Signals of variation in human mutation rate at multiple levels of sequence context. Mol. Biol. Evol. 36, 955–965 (2019). 6. Seplyarskiy, V. B. et al. Population sequencing data reveal a compendium of mutational processes in human germline. bioRxiv (2020). 7. Pinto, Y. et al. Clustered mutations in hominid genome evolution are consistent with APOBEC3G enzymatic activity. Genome Res. 26, 579–587 (2016). 8. Seplyarskiy, V. B., Andrianova, M. A. & Bazykin, G. A. APOBEC3A/B-induced mutagenesis is responsible for 20% of heritable mutations in the TpCpW context. Genome Res. 27, 175–184 (2017). 9. Lindley, R. A. & Hall, N. E. APOBEC and ADAR deaminases may cause many single nucleotide polymorphisms curated in the OMIM database. Mutat. Res. 810, 33–38 (2018). 10. Alexandrov, L. B. et al. Signatures of mutational processes in human cancer. Nature 500, 415 (2013). 11. Alexandrov, L. B. et al. The repertoire of mutational signatures in human cancer. Nature 578, 94–101 (2020). 12. Rebhandl, S., Huemer, M., Greil, R. & Geisberger, R. AID/APOBEC deaminases and cancer. Oncoscience 2, 320 (2015). 13. Jónsson, H. et al. Parental influence on human germline de novo mutations in 1,548 trios from Iceland. Nature 549, 519 (2017). 14. Francioli, L. C. et al. Genome-wide patterns and properties of de novo mutations in humans. Nat. Genet. 47, 822 (2015). 15. Besenbacher, S. et al. Multi-nucleotide de novo mutations in humans. PLoS Genet. 12, e1006315 (2016). 16. Alexandrov, L. B., Nik-Zainal, S., Wedge, D. C., Campbell, P. J. & Stratton, M. R. Deciphering signatures of mutational processes operative in human cancer. Cell Rep. 3, 246–259 (2013). 17. Mathieson, I. & Reich, D. Differences in the rare variant spectrum among human populations. PLoS Genet. 13, e1006581 (2017). 18. Lee, D. D. & Seung, H. S. Learning the parts of objects by non-negative matrix factorization. Nature 401, 788–791 (1999). 19. Behjati, S. et al. Genome sequencing of normal cells reveals developmental lineages and mutational processes. Nature 513, 422–425 (2014). 20. Rahbari, R. et al. Timing, rates and spectra of human germline mutation. Nat. Genet. 48, bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

126 (2016). 21. Pfeifer, G. P. Mutagenesis at methylated CpG sequences. in DNA methylation: basic mechanisms 259–281 (Springer, 2006). 22. Sigg, C. D. . J. M. B. Expectation-maximization for sparse and non-negative PCA. in Proceedings of the 25th international conference on Machine learning 960–967 (2008). doi:10.1145/1390156.1390277. 23. Roberts, N. D. Patterns of somatic genome rearrangement in human cancer. (Univ. Cambridge, Wellcome Trust Sanger Institute, 2018). doi:10.17863/CAM.22674. 24. Lee-Six, H. et al. The landscape of somatic mutation in normal colorectal epithelial cells. Nature 574, 532–537 (2019). 25. Yoshida, K. et al. Tobacco smoking and somatic mutations in human bronchial epithelium. Nature 578, 266–272 (2020). 26. Moore, L. et al. The mutational landscape of normal human endometrial epithelium. Nature 580, 640–646 (2020). 27. Speidel, L., Forest, M., Shi, S. & Myers, S. R. A method for genome-wide genealogy estimation for thousands of samples. Nat. Genet. 51, 1321–1329 (2019). 28. Carlson, J., DeWitt, W. S. & Harris, K. Inferring evolutionary dynamics of mutation rates through the lens of mutation spectrum variation. Curr. Opin. Genet. Dev. 62, 50–57 (2020). 29. Jónsson, H. et al. Whole genome characterization of sequence diversity of 15,220 Icelanders. Sci. data 4, 170115 (2017). 30. Goldmann, J. M. et al. Parent-of-origin-specific signatures of de novo mutations. Nat. Genet. 48, 935 (2016). 31. Tate, J. G. et al. COSMIC: the Catalogue Of Somatic Mutations In Cancer. Nucleic Acids Res. 47, D941–D947 (2018). 32. Lavie, L., Kitova, M., Maldener, E., Meese, E. & Mayer, J. CpG methylation directly regulates transcriptional activity of the human endogenous retrovirus family HERV- K(HML-2). J. Virol. 79, 876–883 (2005). 33. Leung, D. C. & Lorincz, M. C. Silencing of endogenous retroviruses: when and why do histone marks predominate? Trends Biochem. Sci. 37, 127–133 (2012). 34. Hurst, T. P. & Magiorkinis, G. Epigenetic Control of Human Endogenous Retrovirus Expression: Focus on Regulation of Long-Terminal Repeats (LTRs). Viruses 9, (2017). 35. Esnault, C. et al. APOBEC3G cytidine deaminase inhibits retrotransposition of endogenous retroviruses. Nature 433, 430–433 (2005). 36. Koito, A. & Ikeda, T. Intrinsic immunity against retrotransposons by APOBEC cytidine deaminases. Front. Microbiol. 4, 28 (2013). 37. Anwar, F., Davenport, M. P. & Ebrahimi, D. Footprint of APOBEC3 on the genome of human retroelements. J. Virol. 87, 8195–8204 (2013). 38. Narasimhan, V. M. et al. Estimating the human mutation rate from autozygous segments reveals population differences in human mutational processes. Nat. Commun. 8, 303 (2017). 39. Carlson, J. et al. Extremely rare variants reveal patterns of germline mutation rate heterogeneity in humans. Nat. Commun. 9, 3753 (2018). 40. Smith, T. C. A., Arndt, P. F. & Eyre-Walker, A. Large scale variation in the rate of germ-line bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

de novo mutation, base composition, divergence and diversity in humans. PLoS Genet. 14, e1007254 (2018). 41. Duncan, B. K. & Miller, J. H. Mutagenic deamination of cytosine residues in DNA. Nature 287, 560–561 (1980). 42. Ehrlich, M., Zhang, X.-Y. & Inamdar, N. M. Spontaneous deamination of cytosine and 5- methylcytosine residues in DNA and replacement of 5-methylcytosine residues with cytosine residues. Mutat. Res. Genet. Toxicol. 238, 277–286 (1990). 43. Tomso, D. J. & Bell, D. A. Sequence context at human single nucleotide polymorphisms: overrepresentation of CpG dinucleotide at polymorphic sites and suppression of variation in CpG islands. J. Mol. Biol. 327, 303–308 (2003). 44. Molaro, A. et al. Sperm methylation profiles reveal features of epigenetic inheritance and evolution in primates. Cell 146, 1029–1041 (2011). 45. Alexandrov, L. B. et al. Clock-like mutational processes in human somatic cells. Nat. Genet. 47, 1402–1407 (2015). 46. Holliday, R. & Grigg, G. W. DNA methylation and mutation. Mutat. Res. 285, 61–67 (1993). 47. Hernando-Herraez, I., Garcia-Perez, R., Sharp, A. J. & Marques-Bonet, T. DNA Methylation: Insights into Human Evolution. PLoS Genet. 11, e1005661 (2015). 48. Martincorena, I. et al. Somatic mutant clones colonize the human esophagus with age. Science (80-. ). 362, 911 LP – 917 (2018). 49. Glémin, S. et al. Quantification of GC-biased gene conversion in the human genome. Genome Res. 25, 1215–1228 (2015). 50. Anderson-Trocmé, L. et al. Legacy Data Confound Genomics Studies. Mol. Biol. Evol. 37, 2–10 (2020). 51. Patil, V. et al. Human mitochondrial DNA is extensively methylated in a non-CpG context. Nucleic Acids Res. 47, 10072–10085 (2019). 52. Ju, Y. S. et al. Origins and functional consequences of somatic mitochondrial DNA mutations in human cancer. Elife 3, (2014). 53. Schulz, W. A., Steinhoff, C. & Florl, A. R. Methylation of endogenous human retroelements in health and disease. in DNA Methylation: Development, Genetic Disease and Cancer 211–250 (Springer, 2006). 54. Suzuki, Y. et al. AgIn: measuring the landscape of CpG methylation of individual repetitive elements. Bioinformatics 32, 2911–2919 (2016). 55. Smit, A. F. A., Hubley, R. & Green, P. RepeatMasker Open-4.0. 2013-2015. (2015). 56. Casbon, J. PyVCF-A Variant Call Format Parser for Python. 2017. 57. Cock, P. J. A. et al. Biopython: freely available Python tools for computational molecular biology and bioinformatics. Bioinformatics 25, 1422–1423 (2009). 58. Team, R. C. R: A Language and Environment for Statistical Computing. (2019). 59. Bayati, M. et al. CANCERSIGN: a user-friendly and robust tool for identification and classification of mutational signatures and patterns in cancer genomes. Sci. Rep. 10, 1286 (2020). 60. Gaujoux, R. & Seoighe, C. A flexible R package for nonnegative matrix factorization. BMC Bioinformatics 11, 367 (2010). 61. Hutchins, L. N., Murphy, S. M., Singh, P. & Graber, J. H. Position-dependent motif bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

characterization using non-negative matrix factorization. Bioinformatics 24, 2684–2690 (2008). 62. Brunet, J.-P., Tamayo, P., Golub, T. R. & Mesirov, J. P. Metagenes and molecular pattern discovery using matrix factorization. Proc. Natl. Acad. Sci. 101, 4164–4169 (2004). 63. Ebrahimi, D., Anwar, F. & Davenport, M. P. APOBEC3 has not left an evolutionary footprint on the HIV-1 genome. J. Virol. 85, 9139–9146 (2011).

A C>A C>G C>T T>A T>C T>G PC1 0.3

0.2

0.1

0.0

PC2 0.6

0.3

0.0

0.25 PC3 PCA loading PCA 0.00

-0.25

-0.50 0.25 PC4

0.00

-0.25

-0.50 A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T

99.975 e c n a i r a v 99.950 d e n i a l p x e

f 99.925 o

n o i t r o p

o 99.900 r p

e v i t a l u 99.875 m u C

99.850

1 2 3 4 5 6 7 8 Number of components

Supplementary Figure S1. Analysis of the 1000 Genome Project dataset using PCA. (A) First four loadings plots. (B) Evaluation plot showing the %cumulative variance in PCA models with different number of components. bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

20 % Cumulative proportion of explained variance explained of proportion Cumulative %

0 12345678 Number of components

Supplementary Figure S2. Analysis of the 1000 Genome Project dataset using NPCA. Evaluation plot showing the %cumulative variance in NPCA models with different number of components.

A C>A C>G C>T T>A T>C T>G

0 6 5 4 3 2 1 0

4 % Mutation 3

0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

99.9

99.8

99.7 % Cumulative proportion of explained variance

1 2 3 4 5 6 7 8 Number of components

Supplementary Figure S3. Analysis of the 1000 Genome Project dataset using NMF. (A) Mutational signatures 1-4. (E) Evaluation plot showing the %cumulative variance in NMF models with different number of components.

C>A C>G C>T T>A T>C T>G

6 Cosine similarity with 1000G-Sig1 = 0.96

0 40 Cosine similarity with 1000G-Sig2 = 0.99 30

0 20 Cosine similarity with 1000G-Sig3 = 0.83

% Mutation 15

0 8 Cosine similarity with 1000G-Sig4 = 0.87 6

0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

Supplementary Figure S4. Analysis of the 1000 Genome Project dataset using NMF followed by the removal of a common background from signatures 2-4. NMF signatures 1-4 and their corresponding cosine similarities to the 1000G-Sigs 1-4 obtained using NPCA.

C>A C>G C>T T>A T>C T>G 5 Cosine similarity with 1000G-Sig1 = 1 4

Cosine similarity with 1000G-Sig2 = 0.93 15

% Mutation % 5

0 Cosine similarity with 1000G-Sig3 = 0.94 15

Supplementary Figure S5. Analysis of the 1000 Genome Project dataset using HDP. Three identified mutational signatures and their corresponding cosine similarities to the 1000G-Sigs 1- 3, obtained using NPCA, are shown.

A C>A C>G C>T T>A T>C T>G

0.3

0.2

Loading 0.1

0.0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT B 0.4

0.3

0.2 Loading 0.1

0.0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT C ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT 0.4

0.3

0.2 Loading 0.1

0.0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

D E F

60 75

40 40 50

20 20 25 % Cumulative proportion of explained variance explained of proportion Cumulative % variance explained of proportion Cumulative % variance explained of proportion Cumulative %

0 0 0

012 345 678 012 345 678 012 345 678 Number of components Number of components Number of components

Supplementary Figure S6. Analysis of three de novo mutation datasets using PCA. (A) Mutational signature obtained from the analysis of a dataset of 1,548 Icelander trios. (B) Mutational signature obtained from the analysis of a dataset of 816 trios from diverse populations (dbGap). (C) Mutational signature obtained from the analysis of a dataset of 12 European trios. (D-E) Evaluation plots showing the %cumulative variance in PCA models with different number of components for datasets A, B, and C, respectively.

A C>A C>G C>T T>A T>C T>G

2 % Mutation

0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT B ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT 6

2 % Mutation

0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT C 6

% Mutation 2

0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT Frobenius Reconstruction Error D 1.0 1.0e+05 0.9 0.8 7.5e+04 0.7 0.6 0.5 5.0e+04 0.4 0.3 2.5e+04 0.2

Signatuer reproducibility 0.1 0.0 0.0e+00 1 2 3 4 5 6 Frobenius Reconstruction Error E 1.0 0.9 3e+04 0.8 0.7 0.6 2e+04 0.5 0.4 0.3 1e+04 0.2 Signatuer reproducibility 0.1

1 2 3 4 5 6

F Frobenius Reconstruction Error 1.0 6e+02 0.9 0.8 0.7 0.6 4e+02 0.5 0.4 0.3 2e+02 0.2 Signatuer reproducibility 0.1 1 2 3 4 5 6 Number of mutational signatures

Supplementary Figure S7. Analysis of three de novo mutation datasets using NMF. (A) Mutational signature obtained from the analysis of a dataset of 1,548 Icelander trios. (B) Mutational signature obtained from the analysis of a dataset of 816 trios from diverse populations (dbGap). (C) Mutational signature obtained from the analysis of a dataset of 12 European trios. (D-E) Evaluation plots showing the signature reproducibilities and reconstruction errors in NMF models with different number of components for datasets A, B, and C, respectively. bioRxiv preprint doi: https://doi.org/10.1101/2021.01.09.426041; this version posted January 9, 2021. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license.

C>A C>G C>T T>A T>C T>G 6

2 % Mutation %

Supplementary Figure S8. Analysis of three de novo mutation datasets using HDP. One mutational signature was identified.

A C>A C>G C>T T>A T>C T>G 0.4

0.3

0.2 Loading 0.1

0.0 T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T T A A A A C C C C G G G G T T T

50 e c n a i 40 r a v

d e n i a l p x

e 30

f o

n o i t r o p o

r 20 p

e v i t a l u m u

C 10

0 1 2 3 4 5 6 7 8 Number of components

Supplementary Figure S9. Analysis of mtDNA mutations using PCA. (A) The mtDNA mutational signature obtained using PCA (B) Evaluation plot showing the %cumulative variance in PCA models with different number of components.

10 % Cumulative proportion of explained variance explained of proportion Cumulative %

012345678 Number of components

Supplementary Figure S10. Analysis of mtDNA mutations using NPCA. Evaluation plot showing the %cumulative variance in NPCA models with different number of components.

A C>A C>G C>T T>A T>C T>G

2 % Mutation

10 % Cumulative proportion of explained variance

0 1 2 3 4 5 678 Number of components

Supplementary Figure 11. Analysis of mtDNA mutations using NMF. (A) The mtDNA mutational signature obtained using NMF (B) Evaluation plot showing the %cumulative variance in NMF models with different number of components.

C>A C>G C>T T>A T>C T>G 30

Mutation 10

0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

Supplementary Figure S12. The mutation profile of de novo mutations in mitochondrial DNA. The cosine similarity between this profile and 1000G-mt-Sig is 0.78.

SBS1 C>A C>G C>T T>A T>C T>G

30 20 10 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

SBS5 C>A C>G C>T T>A T>C T>G

4 3 2

% Mutation % 1 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

0.13*SBS1+0.87*SBS5, Cosine Similarity = 1 C>A C>G C>T T>A T>C T>G 5 4 3 2

1000G-Sig1 C>A C>G C>T T>A T>C T>G 5 4 3 2

Supplementary Figure 13. Reconstruction of 1000G-Sig1 using COSMIC signatures. The best model of 1000G-Sig1 is 0.13*SBS1+0.78*SBS5.

SBS7a C>A C>G C>T T>A T>C T>G 30 20 10 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

SBS11 C>A C>G C>T T>A T>C T>G 16 12 8 4 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

0.36*SBS7a+0.64*SBS11, Cosine Similarity = 0.9 C>A C>G C>T T>A T>C T>G 20 15 10

% Mutation % 5 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

1000G-Sig2 C>A C>G C>T T>A T>C T>G 40 30 20

% Mutation % 10 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

Supplementary Figure S14. Reconstruction of 1000G-Sig2 using COSMIC signatures. The best model of 1000G-Sig2 is 0.36*SBS7a+0.64*SBS11.

SBS43 C>A C>G C>T T>A T>C T>G 25 20 15 10 % Mutation % 5 0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

SBS47 C>A C>G C>T T>A T>C T>G 16

% Mutation % 4

SBS51 C>A C>G C>T T>A T>C T>G

5 % Mutation %

0.26*SBS43+0.58*SBS47+0.16*SBS51, Cosine Similarity = 0.83 C>A C>G C>T T>A T>C T>G 10.0

7.5

5.0

% Mutation % 2.5

0.0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

1000G-Sig3 C>A C>G C>T T>A T>C T>G

% Mutation % 5

Supplementary Figure S15. Reconstruction of 1000G-Sig3 using COSMIC signatures. The best model of 1000G-Sig3 is 0.26*SBS43+0.58*SBS47+0.16*SBS51.

SBS16 C>A C>G C>T T>A T>C T>G 25 20 15 10 % Mutation % 5 0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

SBS37 C>A C>G C>T T>A T>C T>G

% Mutation % 2

SBS39 C>A C>G C>T T>A T>C T>G 5 4 3 2 % Mutation % 1 0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

0.29*SBS16+0.23*SBS37+0.48*SBS39, Cosine Similarity = 0.75 C>A C>G C>T T>A T>C T>G

7.5

5.0

% Mutation % 2.5

1000G-Sig4 C>A C>G C>T T>A T>C T>G

5 4 3 2 % Mutation % 1 0 ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT

Supplementary Figure S16. Reconstruction of 1000G-Sig4 using COSMIC signatures. The best model of 1000G-Sig4 is 0.29*SBS16+0.23*SBS37+0.48*SBS39.

SBS12 C>A C>G C>T T>A T>C T>G

7.5 5.0 2.5 % Mutation % 0.0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

SBS30 C>A C>G C>T T>A T>C T>G

5 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

0.68*SBS12+0.32*SBS30, Cosine Similarity = 0.92 C>A C>G C>T T>A T>C T>G 6

2 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

1000G-Sig-mt C>A C>G C>T T>A T>C T>G

6 4 2 % Mutation % 0 ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ACA ACC ACG ACT CCA CCC CCG CCT GCA GCC GCG GCT TCA TCC TCG TCT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT ATA ATC ATG ATT CTA CTC CTG CTT GTA GTC GTG GTT TTA TTC TTG TTT

Supplementary Figure S17. Reconstruction of 1000G-Sig-mt using COSMIC signatures. The best model of 1000G-Sig-mt is 0.68*SBS12+0.32*SBS30.

10000

7500 Population Ad Mixed American African East Asian European South Asian 5000 Number of mutation of Number

2500

ACA>AAA ACC>AAC ACG>AAG ACT>AAT CCA>CAA CCC>CAC CCG>CAG CCT>CAT GCA>GAA GCC>GAC GCG>GAG GCT>GAT TCA>TAA TCC>TAC TCG>TAG TCT>TAT ACA>AGA ACC>AGC ACG>AGG ACT>AGT CCA>CGA CCC>CGC CCG>CGG CCT>CGT GCA>GGA GCC>GGC GCG>GGG GCT>GGT TCA>TGA TCC>TGC TCG>TGG TCT>TGT ACA>ATA ACC>ATC ACG>ATG ACT>ATT CCA>CTA CCC>CTC CCG>CTG CCT>CTT GCA>GTA GCC>GTC GCG>GTG GCT>GTT TCA>TTA TCC>TTC TCG>TTG TCT>TTT ATA>AAA ATC>AAC ATG>AAG ATT>AAT CTA>CAA CTC>CAC CTG>CAG CTT>CAT GTA>GAA GTC>GAC GTG>GAG GTT>GAT TTA>TAA TTC>TAC TTG>TAG TTT>TAT ATA>ACA ATC>ACC ATG>ACG ATT>ACT CTA>CCA CTC>CCC CTG>CCG CTT>CCT GTA>GCA GTC>GCC GTG>GCG GTT>GCT TTA>TCA TTC>TCC TTG>TCG TTT>TCT ATA>AGA ATC>AGC ATG>AGG ATT>AGT CTA>CGA CTC>CGC CTG>CGG CTT>CGT GTA>GGA GTC>GGC GTG>GGG GTT>GGT TTA>TGA TTC>TGC TTG>TGG TTT>TGT

Supplementary Figure 18. Analysis of 1000G Project datasets. Number of mutations (changes relative to the sequence of most common human ancestor) is shown for different human populations.