Mutational processes in cancer are heterogeneous, which can bias dN/dS estimates of selective pressures. dN/dS overcomes this issue by dividing observed mutation counts by what is expected under neutral evolution using null models. These null models must account for mutational biases that are often specific to cancer types and genomic regions.

To ensure our dN/dS calculations are robust and reproducible, we applied two different methods to account for mutational biases. The first approach uses a previously established parametric mutational model (dNdScv) that explicitly estimates the background mutational bias of each gene in its calculation of dN/dS (Martincorena et al., 2017). The second approach uses a permutation-based, non-parametric (parameter-free) estimation of dN/dS. In this approach, every observed mutation is permuted while preserving the gene, patient samples, specific base change (e.g. A>T) and its tri-nucleotide context. Note that permutations do not preserve the codon position of a mutation and thus can change its protein coding effect (nonsynonymous vs. synonymous). The permutations are then tallied for both nonsynonymous d_N^(permuted) and synonymous d_S^(permuted) substitutions (Figure 2—figure supplement 1) and used as expected proportional values for the observed number of nonsynonymous d_N^(observed) (or simply d_N) and synonymous d_S^(observed) (d_S) mutations in the absence of selection. The unbiased effects of selection on a gene, dN/dS, is then:dNdS=dN(observed)/dN(permuted)dS(observed)/dS(permuted)

For all cancer types and patient samples, p-values and confidence intervals are determined by bootstrapping patient samples. Note that this permutation procedure will account for gene and tumor-level mutational biases (e.g. neighboring bases [Alexandrov and Stratton, 2014], transcription-coupled repair, S phase timing [Haradhvala et al., 2016], mutator phenotypes) and their covariation. We confirmed that this approach accurately measures selection even in the presence of simulated mutational biases (Materials and methods, Figure 2—figure supplement 2A). In addition, this approach also reliably measures the absence of selection (dN/dS = 1) in weakly expressed genes (Figure 2—figure supplement 2C).

We find that both the parametric and non-parametric approaches identify similar patterns of selection (Figure 2A). Since parametric mutational models can become very complex in cancer (exceeding 5000 parameters in some cases; Martincorena et al., 2017; Zapata et al., 2018), we elected to use the non-parametric approach, which makes fewer assumptions about underlying mutational processes, in subsequent calculations of dN/dS.

We estimated dN/dS patterns in both driver and passenger gene sets across 10,288 tumors from TCGA aggregated over 33 cancer types (Ellrott et al., 2018) (Materials and methods). Since TCGA is composed of whole-exome data, which limits our ability to assess mutations in non-coding regions, we elected to use the total number of protein-coding mutations as our proxy for the mutational burden of tumors. To quantify the extent that selection attenuates as the mutational burden increases, we stratified tumors into bins based on their total number of substitutions on a log-scale. For each bin of tumors, we pooled all of the variants together and estimated dN/dS jointly. Consistent with the inefficient selection model, whereby selection fails to eliminate deleterious mutations in high mutational burden tumors, we observe pervasive selection against passengers exclusively in tumors with low mutational burdens (dN/dS ~ 0.56 in tumors with ≤3 substitutions, while dN/dS ~ 0.93 in tumors with >10 substitutions, Figure 2A). We observed little negative selection in passenger genes when aggregating tumors across all mutational burdens (dN/dS ~ 0.93), which is broadly similar to previous estimates (Martincorena et al., 2017; Weghorn and Sunyaev, 2017; Zapata et al., 2018; Ostrow et al., 2014).

We confirmed that negative selection on passengers is specific to low mutational burden tumors and not biased by small sample sizes (Figure 2—figure supplement 2B). We randomly sampled passengers from high mutational burden tumors (>10 substitutions) 1000 times using the same bin sizes in Figure 2A and calculated dN/dS. Within the smallest bin size (N=168 somatic nucleotide variant [SNVs]), negative selection on passengers sampled from high mutational burden tumors was absent (average dN/dS ~ 0.96) compared to observed dN/dS in low mutational burden tumors (dN/dS ~ 0.56; p<2.2^–16). In fact, only 1.7% of randomly sampled sets of sites had similar signals of negative selection (dN/dS < 0.56).

Also consistent with the inefficient selection model, drivers exhibit a similar but opposing trend of attenuated selection at elevated mutational burdens (dN/dS ~ 2.7 when in tumors with ≤3 substitutions and dN/dS gradually declines to ~1.16 in tumors with >100 substitutions). This pattern is not specific to drivers that are oncogenes or tumor suppressors (Figure 2—figure supplement 3). While the attenuation of selection against passengers in higher mutational burden tumors is a novel discovery, this pattern among drivers has been reported previously (Martincorena et al., 2017). Furthermore, we confirmed that these patterns are robust to the choices that we made in our analysis pipeline. These include the: (i) effects of germline SNP contamination (Figure 2—figure supplement 4), (ii) choice of driver gene set (Bailey et al., 2018, IntOGen Gonzalez-Perez et al., 2013, and COSMIC Tate et al., 2019; Forbes et al., 2008, Figure 2—figure supplement 5), (iii) differences in tumor purity and thresholding (Figure 2—figure supplement 6), and (iv) null model of mutagenesis (dNdScv, Figure 2A and Figure 2—figure supplement 7; Martincorena et al., 2017) (Materials and methods).

If negative selection is more pronounced in low mutational burden tumors, then the nonsynonymous mutations observed should also be less functionally consequential. By annotating the functional effect of all missense mutations using PolyPhen2 (Adzhubei et al., 2010; Figure 2B), we indeed find that observed nonsynonymous passengers are less damaging in low mutational burden cancers. Similarly, driver mutations become less functionally consequential as mutational burden increases, as expected for mutations experiencing inefficient positive selection (Figure 2B). Together these two trends provide additional and orthogonal evidence that selective forces on nonsynonymous mutations are more efficient in low mutational burden cancers.

Since all mutational types experience Hill-Robertson interference, attenuated selection should also persist in CNAs. We used two previously published statistics to quantify selection in CNAs: breakpoint frequency (Korbel et al., 2007) and fractional overlap (Zack et al., 2013). For both measures, we compare the number of CNAs that either terminate (breakpoint frequency) within or partially overlap (fractional overlap) Exonic regions of the genome relative to non-coding (Intergenic and Intronic) regions (dE/dI, see Materials and methods). Like dN/dS, dE/dI is expected to be <1 in genomic regions experiencing negative selection, >1 in regions experiencing positive selection (e.g. driver genes), and ~1 when selection is absent or inefficient (Figure 2—figure supplement 8). Using dE/dI, we observe attenuating selection in both driver and passenger CNAs as the total number of CNAs increases for both breakpoint frequency (Figure 2C) and fractional overlap (Figure 2—figure supplement 9). While CNAs of all lengths experience attenuated selection, CNAs longer than the average gene length (>100 KB) experience greater selective pressures in drivers. Collectively, these results strongly support the inefficient selection model and argue that the observed patterns must be due to a universal force in tumor evolution. We find that selection consistently attenuates in both drivers and passengers across all cancers as mutational burden increases.

We next tested alternative hypotheses to the inefficient selection model. We considered the possibility that selection is strong only during normal tissue development, but absent after cells have transformed to malignancy. This would disproportionately affect low mutational burden tumors, as a greater proportion of their mutations arise prior to tumor transformation. If true, then attenuated selection should be absent in subclonal mutations, which must arise during tumor growth. However, selection clearly attenuates with increasing mutational burden for the subset of likely subclonal mutations with variant allele frequency (VAF) below 20% (Figure 2D and Figure 2—figure supplement 10). Although selection attenuates in drivers and passengers in both subclonal and clonal mutations, selection is weaker in both drivers and passengers with lower VAFs. Weaker efficiency of selection among less frequent variants is expected under a range of population genetic models (Messer, 2009) and especially so in rapidly expanding, spatially constrained cancers (Sottoriva et al., 2015). In addition, heterozygous mutations, to the extent they are only partially dominant (López et al., 2020), are also expected to exhibit lower VAFs and experience weaker selection.

Next, we considered and rejected the possibility that attenuated selection is limited to particular types of genes. We first annotated our observed mutations by different functional categories and Gene Ontology (GO) terms (Harris et al., 2004) and find that negative selection is not specific to any particular gene functional category expected to be under constraint, and specifically not limited to essential or housekeeping genes – a key prediction of the ‘weak selection’ model (Martincorena et al., 2017; Figure 2—figure supplement 11, p<0.05, Wilcoxon signed-rank test).

Finally, we found that these patterns of attenuated selection persist across cancer subtypes for both SNVs and CNAs. We calculated dN/dS in tumors grouped by nine broad anatomical sub-categories (e.g. neuronal) and 33 subtype classifications (Grossman et al., 2016; Figure 2E–F). We find that patterns of attenuated selection in SNVs persists in the broad and specific (drivers p=3.8 × 10^–5, passengers p=1.7 × 10^–2, Wilcoxon signed-rank test; Figure 2—figure supplement 12) classification schemes. Furthermore, dE/dI measurements of CNAs exhibit similar patterns of selection in broad (Figure 2—figure supplement 13) and specific subtypes (Figure 2F; drivers p<0.05 and passengers p<0.05).

Collectively, these results suggest that tumors with elevated mutational burdens carry a substantial deleterious load. Since nonsynonymous mutations are thought to be primarily deleterious by inducing protein misfolding (Drummond and Wilke, 2008; Lobkovsky et al., 2010), we tested whether an increase in the number of passenger mutations in tumors would lead to elevated protein folding stress, and, in turn, drive the upregulation of heat shock and protein degradation (McGrail et al., 2020) pathways in cancer (Santagata et al., 2011). Indeed, gene expression of HSP90, Chaperonins, and the Proteasome does increase across the whole range of SNV (weighted R² of 0.84, 0.78, and 0.78, respectively) and CNA burdens (weighted R² of 0.83, 0.88, and 0.85, respectively) (Figure 2G and Figure 2—figure supplement 14A). This trend persists across cancer types for SNVs and CNAs (Figure 2—figure supplement 14D-E). Importantly, expression of these gene sets increases across the whole range of mutational burdens, even after the dN/dS of passengers approaches 1. This result presents additional evidence that passengers continue to impart a substantial cost to cancer cells, even in high mutational burden tumors.

We next tested whether Hill-Robertson interference – a process where selection becomes inefficient due to interference between linked mutations with competing fitness effects – alone can generate these patterns of attenuated selection. Specifically, we modeled tumor progression as a simple evolutionary process with advantageous drivers and deleterious passengers. We then used approximate Bayesian computation (ABC) to compare these simulations to observed data and infer the mean fitness effects of drivers and passengers.

Our previously developed evolutionary simulations model a well-mixed population of tumor cells that can randomly acquire advantageous drivers and deleterious passengers during cell division (McFarland et al., 2013). The product of the individual fitness effects of these mutations determines the relative birth and death rate of each cell, which in turn dictates the population size N of the tumor. If the population size of a tumor progresses to malignancy (N>1,000,000) within a human lifetime (≤100 years), the accrued mutations and patient age are recorded. The mutation rate of each simulated tumor is randomly sampled from a broad range (10^–12–10^–7 mutations · nucleotide^–1 · generation^–1, Materials and methods). Although this model ignores a great deal of known tumor biology, we believe it constitutes the simplest evolutionary model that could possibly recapitulate observed selection for drivers and against passengers. Our question is not whether this model is correct in all details but rather whether even such a simple model can generate quantitatively similar patterns as observed in the data with sensible values of mutation rates and selection coefficients.

Figure 3A illustrates the ABC procedure. To compare our model to observed data, we simulated an exponential distribution of fitness effects (DFEs) with mean fitness values that spanned a broad range (10^–2–10⁰ for driver and 10^–4–10^–2 for passengers, Materials and methods). We summarized observed and simulated data using statistics that capture three relationships: (i) the dependence of driver and passenger dN/dS rates on mutational burden, (ii) the rate of cancer age incidence (SEERs database National Cancer Institute, 2007), and (iii) the distribution of mutational burdens (summary statistics of (ii) and (iii) were based on theoretical parametric models Frank, 2007, Materials and methods, Figure 3—figure supplements 1–2). We then inferred the posterior probability distribution of mean driver fitness benefit and mean passenger fitness cost using a rejection algorithm that we validated using leave-one-out cross validation (CV) (Materials and methods, Figure 3—figure supplement 3).

Using this approach, the maximum likelihood estimate (MLE) of mean driver fitness benefit is 53% (Figure 3B), while the MLE of passenger mean fitness cost is 1.03% (Figure 3C). Simulations with these MLE values agree well with all observed data (Figure 3D–F, Pearson’s r=0.988 for combined driver/passenger dN/dS).

While Hill-Robertson interference alone explains dN/dS rates in the passengers well, the simulations most consistent with observed data still exhibited consistently higher dN/dS rates in drivers (Figure 3D). We tested whether positive selection on synonymous mutations within driver genes could explain this discrepancy. Indeed, we find that a model incorporating synonymous drivers agrees modestly better with observed statistics (3.5-fold relative likelihood, ABC posterior probability). The best-fitting model predicts that ~6% of synonymous mutations within driver genes experience positive selection, which is consistent with previous estimates for human oncogenes (Supek et al., 2014) (Materials and methods, Figure 3D and Figure 3—figure supplement 4). Furthermore, we observe additional evidence of selection and codon bias in synonymous drivers exclusive to low mutational burdens (TCGA samples, Materials and methods, Figure 3—figure supplement 4).

We note that although deleterious passengers are necessary to explain attenuation of negative selection with mutational burden in passengers, alternative explanations could also contribute to attenuation of positive selection in drivers. Specifically, high mutational burden tumors are more likely to contain mutations in pan-cancer driver gene sets which might not directly contribute to tumorigenesis in specific tumors, and thus might not be under direct positive selection in all tumors. Similarly, additional driver mutations might not directly contribute to tumor fitness beyond a certain number of driver mutations (e.g. 5-hit model). Nonetheless, it’s important to note that Hill-Robertson interference is capable of reproducing all the features of the data (steep attenuation of negative selection in passengers and gradual attenuation of positive selection in drivers).

Overall, our results indicate that rapid adaptation through natural selection – acting on entire genomes, rather than individual mutations – is pervasive in all tumors, including those with elevated mutational burdens. Given the quantity of drivers and passengers observed in a typical cancer (TCGA), our model implies that cancer cells are in total ~90% fitter than normal tissues (119% total benefit of drivers, 46% total cost of passengers). A median of five drivers each of which has a mean benefit of ~19% accumulate per tumor in these simulations – also consistent with estimates from age incidence curves (National Cancer Institute, 2007), known hallmarks of cancer (Hanahan and Weinberg, 2000), and estimates of the selective benefit of individual drivers (Dai et al., 2007). Lastly, the mutation rates of tumors that could progress to cancer in our model also recapitulate observed mutation rates in human cancer (Camps et al., 2007) (median 3.7×10^–9, 95% interval 1.1×10^–10–8.2×10^–8, Figure 3—figure supplement 5).

Most notably, under our modeling assumptions, all passengers together confer a fitness cost of ~46% per tumor. While this collective burden appears large, the individual fitness effects of accumulated passengers in these simulations (mean 0.8%) are similar to observed fitness costs in cancer cell lines (1–3%) (Williams et al., 2008) and the human germline (0.5%) (Cassa et al., 2017). Note that in our model, these passengers accumulated primarily via Muller’s ratchet, while only ~5% accumulated via hitchhiking inferred using population genetics theory (McFarland et al., 2013) and MLE fitness effects, Materials and methods, Figure 3—figure supplement 6. These results suggest that Hill-Robertson interference is a plausible model for the empirical patterns of attenuated selection with mutational burden observed in the data.

Since TCGA is composed of whole-exome data, which limits our ability to accurately assess mutations in non-coding regions, we elected to use the total number of protein-coding mutations (i.e. missense, nonsense, and synonymous mutations) as our proxy for the mutational burden of tumors. This allows us to focus on the highest quality set of mutations that we have, which can impact the power to detect selection (Figure 2—figure supplement 15). We note that this high-quality set mutations does not have evidence of germline contamination by common SNPs (MAF > 5%) from 1000 Genomes Project (1000 Genomes Project Consortium et al., 2015) (v2015 Aug) using ANNOVAR (Wang et al., 2010) to annotate mutations in TCGA (Figure 2—figure supplement 4). For all analyses calculating dN/dS in tumors stratified by their mutational burden, all variants within each bin of tumors were pooled together and dN/dS was calculated jointly on each bin of tumors. Counts of the number of mutations use to estimate dN/dS in each mutational burden can be found in Figure 2—figure supplement 16.

We assume that for any particular tumor, mutation rates are constant across a gene for a particular tri-nucleotide context and base change (e.g. C>G). Our procedure is inspired by constrained marginal models (or ‘edge switching’ in network analysis), whereby the marginal distributions of observations aggregated over known confounding variables are preserved under permutation to create a null distribution. In our application of this strategy, the marginal distributions of mutations (across tri-nucleotide context, base change, gene, and tumor) remain preserved – as they would be in a constrained marginal model; however, we exhaustively consider every acceptable permutation of the data. Because our approach is highly constrained, these permutations are exhaustively computable (median 36 alternatives per mutation). Thus, resampling is unnecessary.

Our null model presumes that all mutations of type i, defined by a tri-nucleotide context and base change, arise with probability M_igt within each gene g and tumor t. For each gene, we tally the total quantity of nonsynonymous mutations N_ig and synonymous mutations S_ig. Suppose selection enriches or depletes nonsynonymous mutations within a gene and tumor by a rate ω_gt. The expected number of nonsynonymous and synonymous mutations within a particular tumor and gene are EdN=ω∑iMigtNig and EdS=∑iMigtSig in the absence of selective pressures on synonymous mutations. As with the main text, d_N and d_N^(observed) are used interchangeably. Although M_igt is unknown, dN/dS statistics attempt to infer selection nonetheless by noting that:E[dN]E[dS]=ωgt∑iMigtNig∑iMigtSig=ωgt<Migt,Nigt><Migt,Sigt>=ωgtρMN||Mgt||||Ngt||ρMS||Mgt||||Sgt||=ωgtρMN||Ngt||ρMS||Sgt||

Note that ρAB=A,B>(‖A‖‖B‖), where ‖A‖=<A,A> is the Pearson product-moment correlation coefficient. When ρ_MN ≈ ρ_MS,EdN/NiEdS/Si≈ω

That is, dN/dS is approximately equal to the selective pressures on nonsynonymous mutations when the accessible nonsynonymous and synonymous loci are properly accounted and when the correlation between mutational processes and nonsynonymous loci are roughly equivalent to the correlation between mutational processes and synonymous loci. Traditionally, this assumption was used to calculate dN/dS. To improve resolution of dN/dS, researchers have attempted to account for these correlations using sophisticated parametric models of M_igt. An alternative statistical approach, however, is to treat these correlations as nuisance parameters.

Constrained marginal models permute observed data in all possible manners that preserve the underlying covariance structure of the data (e.g. ρ_MN, ρ_MS). In our particular case of this method, we note that by definition, dNpermuted=∑i(dNobservediNi+dSobservediNi) . Thus:E[dNpermuted]E[dSpermuted]=∑i(ωMigtNigt2+MigtNigtSigt)∑i(ωMigtNigtSigt+MigtSigt2)=ωρMN‖M‖‖N‖2+ρMN‖M‖‖N‖‖S‖ωρMS‖M‖‖S‖‖N‖+ρMS‖M‖‖S‖2=ρMN‖N‖ρMS‖S‖

Hence, by dividing the observed mutations by all permutations, we eliminate the covariance of mutational processes with available loci and, thus, measure ω_gt directly for any particular gene-tumor combination without mutational bias.

Unfortunately, because of the log-sum inequality, mutational bias can arise once cohorts of genes and cohorts of tumor samples are binned. This problem is common to all dN/dS measures and is a consequence of the correlation of mutational biases with selection (i.e. Migt,ω>) – not the correlation of mutational biases with one another, as these covariances are already accounted for in a constrained marginal model. For example, if tri-nucleotide biases covary linearly with gene-level biases, and are independent of tumor-level biases, then a parametric estimate of M_igt may deconstruct M_igt into Migt=fi,g,t,ρig , where ρig is the covariation of tri-nucleotide mutational biases with gene-level biases. Nonetheless, Migt,ω>∝<ρig,ω> will still be ignored. Indeed, this covariation of mutational processes with selective forces is the focus of our current study: selection and genome-wide mutation rate are correlated (i.e. ∑tMigtω≠0) because of Hill-Robertson interference. Hence, the level at which observed d_N values d_S are binned necessarily ignores covariation between mutational processes and selection (in addition to any variation of ω_gt within the bin). Another example of this binning challenge arises when positive and negative selection act on different regions of the same gene, which gene-level dN/dS binning can misinterpret as neutral evolution.

To confirm that our null model can accurately estimate dN/dS even in the presence of extreme tri-nucleotide mutational biases, we simulated artificial data where different COSMIC signatures (Tate et al., 2019Forbes et al., 2008; Forbes et al., 2008) (SBS Signatures 1–9, v3) contribute to all of the mutations. Permuted d_N and d_S tallies for each mutational context were simulated by randomly sampling 1000 genes with the same mutational context. The fraction of permuted d_N and d_S tallies for each mutational context was used as weighted probabilities to derive observed d_N and d_S tallies. To simulate negative selection, d_N counts were randomly removed from each context at a rate 1 − ω_gt (e.g. a simulated ‘true’ dN/dS of 0.8 in a cohort of samples indicates a 20% chance of nonsynonymous mutations being removed in the samples). These simulated (true) rates were then compared to observed and permuted d_N and d_S tallies according to the dN/dS metric that we used throughout this study:dNdS=dN(observed)/dN(permuted)dS(observed)/dS(permuted)

We confirmed that this approach accurately measures selection in the presence of simulated mutational biases (Figure 2—figure supplement 2).

Lastly, we note that binning nonsynonymous and synonymous mutations at the genome-wide level (e.g. drivers and passengers) provided the most robust estimates of dN/dS when bootstrapping observed tumor samples. Statistical power is insufficient when binning at the individual gene level. Bootstrapping also demonstrated that log-transformation of dN/dS values increases statistical power, and thus was generally applied to dN/dS analyses in this study.

For comparison, we also calculated dN/dS using dNdScv (Campbell and Martincorena, 2017) – a previously published parametric null model of mutagenesis in cancer (Martincorena et al., 2017). To compare both methods, dNdScv ran globally and separately on samples stratified by the total number of substitutions using the following parameters: max_coding_muts_per_sample = Inf max_muts_per_gene_per_sample = Inf.

Global dN/dS values of all nonsynonymous mutations (w_all, reported by dNdScv) were used. This model reproduced our non-parametric dN/dS trends (Figure 2A) and was used to infer patterns of selection in synonymous mutations (Figure 3—figure supplement 4). We note that stratifying tumors in TCGA into 20 bins of equal sample size (as was done in Martincorena et al., 2017), rather than evenly spaced bins, averages out a significant proportion of the negative selection observed in passengers, since low mutation burden tumors reside within the tail-end of the distribution (Figure 2—figure supplement 7).

For all analysis using SNVs, unless explicitly stated, a comprehensive list of 299 pan-cancer driver genes derived from 26 computational tools was used to catalog driver genes (Bailey et al., 2018). Other pan-cancer driver gene sets tested were derived from COSMIC’s Driver Gene Census (Tate et al., 2019; Forbes et al., 2008) (downloaded on October 2016) and IntOGen’s Cancer Drivers Database (Gonzalez-Perez et al., 2013) (v2014.12) which contained 602 and 459 number of driver genes, respectively.

Many driver genes are associated with only particular tumor subtypes. To compare patterns of selection across cancer subtypes without increasing or decreasing the size of the list for each subtype, we chose to use a single set of driver genes for most analyses. This may understate the degree of positive selection in driver genes as mutations in these genes may be passengers in some tumor subtypes. In Figure 2—figure supplement 5, we investigate patterns of selection using the top 100 driver genes identified for each tumor type and observe decreased signatures of positive selection overall in driver genes. Nevertheless, the patterns of attenuated selection in drivers and passengers remain. While tissue-type specific driver genes certainly exist, our results suggest that our statistical power to detect drivers still remains too limited to justify subdividing analyses by tumor type in many cases.

For all CNA analysis, GISTIC 2.0 Mermel et al., 2011 was used to identify a set of genomic regions enriched for copy number gains and copy number losses using recommended settings with a confidence threshold of 0.9. CNAs used to identify these peaks were downloaded from the NIH Genomic Data Commons (GDC) (Grossman et al., 2016) in the TCGA cohort. For each amplification peak, the closest gene was annotated as a putative oncogene, and similarly the closest gene to each deletion peak was annotated as a putative tumor suppressor. The top 100 amplification peaks (oncogenes) and deletion peaks (tumor suppressors) were classified as drivers for each of the 32 tumor types. Thirty-four percent of identified driver genes appear in more than one tumor type, while 2.6% of identified driver genes appear in more than five tumor types.

For both SNV and CNA analysis, passengers were defined as mutations that did not reside within driver genes. The vast majority of mutations are passengers, and their relative totals for both SNVs and CNAs are depicted in Figure 2—figure supplement 16.

VAFs were calculated per site as the number of mutant read counts divided by the total number of read counts. VAFs were adjusted for purity using calls made by ABSOLUTE (Grossman et al., 2016; Carter et al., 2012), collected from GDC. A VAF threshold of 0.2 was used to define ‘subclonal’ (<0.2) vs. ‘clonal’ (>0.2) SNVs. Different VAF thresholds were considered (Figure 2—figure supplement 10) and the choice of ‘clonal’ thresholding did not impact the conclusions of this study.

PolyPhen2 annotations in the MC3 SNP calls were used (Adzhubei et al., 2010). Only missense mutations that were categorized as either ‘benign’, ‘probably damaging’, or ‘possibly damaging’ were used. The fraction of pathogenic missense mutations was calculated as the number of pathogenic mutations categorized as either ‘probably damaging’ or ‘possibly damaging’ divided by the total number of categorized mutations.

To test for patterns of selection in functionally related genes, we annotated all mutations by different functional categories and GO terms (Harris et al., 2004). Oncogenes and tumor suppressors were annotated from a curated set of 99 high confidence cancer genes (Kumar et al., 2015). Essential genes were collected from a genome-wide CRISPR screen that identified genes required for proliferation and survival in a human cancer cell line (Wang et al., 2015). Housekeeping genes were defined as genes with an exon that is expressed in all tissues at any non-zero level, and exhibits a uniform expression level across tissues (Eisenberg and Levanon, 2015). Interacting proteins were downloaded from the mentha database in April 2019 (Calderone and Cesareni, 2012).

To identify highly expressed genes, median transcripts per million (TPM) in 54 tissue types (v7 release) were downloaded from the Genotype-Tissue Expression (GTEx) project (GTEx Consortium, 2020; Carithers and Moore, 2015). Tissues that contained high expression in most genes, specifically testes, were removed. Only genes that had TPM counts above zero in any of the 53 remaining tissues were used. TPM counts were averaged across all tissues. Highly expressed genes were defined as the top 1000 genes expressed across all tissues.

To test for signals of negative selection in other functional groups, we annotated mutations by candidate GO terms according to biological processes: Transcription Regulation (GO Term ID: 0140110), Translation Regulation (GO Term ID: 0045182), and Chromosome Segregation (GO Term ID: 0007059).

All CNAs were downloaded from the COSMIC database on June 2015 (Tate et al., 2019; Forbes et al., 2008). Mitochondrial CNAs were discarded from analysis, as copy number changes are difficult to infer. Gene annotations and the locations of telomeres and centromeres were downloaded from the UCSC Genome Browser (hg19). Telomeric and centromeric regions were masked from all measurements of dE/dI. Because the selection patterns of non-focal CNAs – alterations with at least one terminus in a telomere or centromeric region – were not noticeably different from long (>100 kb) focal CNAs, these two alteration classes were aggregated for analysis. Notably, we observed positive selection for both amplifications and deletions within oncogenes, and for both deletions and amplifications within tumor suppressors. For this reason, we did not distinguish between gains and losses, nor oncogenes and tumor suppressors in published analyses: any CNA that overlapped an oncogene or tumor suppressor in any region (for any fraction of the CNA) was classified as a driver. Mutational burden was defined simply as the total number of CNAs within a sample. Pan-cancer CNAs from cBioPortal (August 2018) were also analyzed, however consistent purity and ploidy estimates could not be obtained by using either ABSOLUTE (Carter et al., 2012) or TITAN (Ha et al., 2015), so this data was not used for published analyses of CNAs.

dE/dI was calculated using a ‘breakpoint frequency’ metric and a ‘fractional overlap’ metric. For both metrics, the dE/dI of a particular gene set i (e.g. driver or passenger genes) is defined by a genomic track T_i,g, which is one for every annotated region g of the track and zero elsewhere. Only non-centromeric and non-telomeric regions are considered in the mappable human genome G. Each CNA C_g,m is defined by its position on the genome g and the mutational burden m of the tumor harboring the mutation. For ‘breakpoint frequency’ C_m,i is one at the position of both termini of the CNA and zero elsewhere. For ‘fractional overlap’ C_m,i is 1/L, where L is the length of the CNA, for every region of the genome spanned by the CNA and zero elsewhere. For a particular range of mutational burdens M, dE/dI was defined as:dEdIi,M=∑mM∑gGTi,gCm,g∑gGTi,g

We note that calculation is accelerated by >×100 by commuting T_i,g with the outer summation (Σ_m^M). Lastly, we randomly permuted the start and stop positions of each CNA, while preserving its length, to derive a set of neutral CNAs not experiencing selection. This permutation analysis finds that dE/dI for both breakpoint frequency and fractional overlap is ~1 in the absence of selection (Figure 2—figure supplement 8).

Tumor purity estimates from the ABSOLUTE algorithm (Carter et al., 2012) were downloaded from the GDC on May 2020. To evaluate the effects of tumor purity on patterns of selection, tumors below increasing thresholds of tumor purity were removed from the analysis, and dN/dS was calculated on tumors stratified by mutational burden bins (as described above.)

Gene expression data was downloaded from the COSMIC database on September 2019. Genes used to identify different protein folding pathways were downloaded from Kampinga et al., 2009, genes involved in protein degradation pathways were identified from Tanaka, 2009. The median gene expression of all genes in each protein folding pathway was used. Patients were binned by the total number of substitutions (using MC3 SNP calls from TCGA) and CNAs, and the average gene expression of each bin was calculated.

All tumor subtypes in were grouped into nine sub-categories, based on broad, predominantly anatomical features. Anatomical features (i.e. organ and systems of organs), rather than histological features or inferred cell-of-origin, were used as groupings because we believe that the fitness effects of mutations should be predominantly defined by the environment of the tumor. Nevertheless, we observed attenuated selection in both drivers and passengers in many broad histologically defined classifications (e.g. adenocarcinomas and sarcomas). For all cancer grouping analysis (broad and subtype), tumors were stratified into bins by the total number of substitutions (d_N+d_S) on a log-scale. Since tumor subtypes vary in their range of mutational burdens, (e.g. KIRC cancer subtypes only have tumors with <100 substitutions), dN/dS values in the lowest and highest mutational burden bin for each cancer subtype are shown.

Specific cancer subtype categories were taken directly from the NCI GDC (Grossman et al., 2016). Because CNAs were downloaded from COSMIC, CNA datasets were not classified with this same ontology. Supplementary file 2 details how CNA classifications were mapped on GDC categories (and sometimes more broadly defined groups). All subtypes with >200 samples were used in our CNA subtype analyses (Figure 2—figure supplement 13).

Somatic cells in our populations are modeled as individual cells that can stochastically divide and die in a first-order (memoryless) Gillespie algorithm. This model was developed and described previously (McFarland et al., 2014). During division, cells can acquire advantageous drivers with rate µT_drivers and deleterious passengers with rate µT_passengers – these values specify the mean of Poisson-distributed pseudo-random number (PRN) generators that prescribe the number of drivers and passengers conferred during division (e.g. the number of drivers per division n_d = Poisson[n_d = k; λ = µT_drivers] = λ^k e^−k/k!). The DFEs conferred by each driver and each passenger are exponentially distributed PRNs with probability densities P(s_i = x; s_drivers)=Exp[−x/s_drivers]/s_drivers and P(s_i = x; s_passengers) = −Exp[−x/s_passengers]/s_passengers, respectively. Simulations with other exponential-family DFEs do not qualitatively differ from these exponential distributions (McFarland et al., 2013). The aggregate absolute cellular fitness is f=∏iall mutations(1+si) in our multiplicative epistasis model and Δf=si/1+νf with ν=1 in our diminishing-returns epistasis model, where Δf is the change in cellular fitness with each mutation (Arjan et al., 1999). The rate of cell birth is inversely proportional to cellular fitness, while the rate of cell death D(N;N0)=Log[1+N(e−1)N0] increases with the population size of the tumor N. With these birth and death processes, mean population size abides by a Gompertzian growth law in the absence of additional mutations, which is scaled by the mean cellular fitness E[N(<f > )]=log[1 +<f >/ N ⁰] (derived from master equation McFarland et al., 2013). While, programmatically, mutations exclusively affect the birth rate and the constraints on growth exclusively affect the death rate, we previously demonstrated that birth and death rates are generally nearly balanced such that dynamics are not affected by this design choice.

Because somatic cells do not recombine during cell division, dominance coefficients were not explicitly modeled. Thus in diploid cancers, our selection coefficients estimate the mean heterozygous effect of drivers and passenger (i.e. hs). Similarly, loss of heterozygosity (LOH) events (gene losses, gene conversions, mitotic recombination, etc.) are not explicitly modeled either; however, these events can be viewed as additional mutations that may be either adaptive drivers or deleterious passengers in the model. As sequencing data improves, we believe that it will be informative to explicitly model dominance coefficients, tumor ploidy, and LOH events.

Simulations progressed until tumor extinction (N=0 cells), malignant transformation (N=10⁶ cells), or until ~100 years had passed (18,500 generations). Only fixed mutations (present in the most recent common ancestor) within clinically detectable growths were analyzed in our ABC pipeline. The behavior of this model has been described previously (McFarland et al., 2013; McFarland et al., 2014) and the most relevant assumptions of this model and their effects on the conclusions of this study are described in Supplementary file 1.

Cells in our populations are fully described by their accrued mutations, and birth and death times. Birth and death events were modeled using an implementation of the next reaction (Gibson and Bruck, 2000), a Gillespie algorithm that orders events using a heap queue. Generation time in our model was defined as the inverse of the mean birth rate of the population: 1/<B(d, p)>. While all mutation events occurred during cell division, if mutations were to occur per unit of time (rather than per generation), rapidly growing tumors would acquire drivers at a slightly slower rate as generation times decline over time. This effect, however, is negligible compared to the variation in waiting times conferred by the variation in mutation rates (division times merely double, while mutation rates vary by 100,000-fold).

This simple evolutionary model is defined by five parameters µT_drivers, µT_passengers, s_drivers, s_passengers, and N⁰. The target size of drivers is defined as the approximate number of nonsynonymous mutations in the Bailey driver screen T_drivers = (# of driver genes)·(mean driver length)·(fraction of SNVs that are nonsynonymous)=300 genes · 1298 loci/gene · 0.737 nonsynonymous loci/ loci = 286,886 nonsynonymous loci. The target size of passengers was simply the remaining loci in the protein coding genome, T_passengers = 20,451,136 nonsynonymous loci. The mutation rate was constant throughout each tumor simulation and randomly sampled from a uniform distribution in log-space that ranged from 10^–12 to 10^–7 mutations·loci^–1·generation^–1. While tumors were initiated from this broad range, malignancies (N>10⁶ cells) were almost always restricted to mutation rates between 10^–10 and 10^–8 (Figure 3—figure supplement 5), as tumors with mutation rates drawn below this range almost never progressed to cancer within 100 years and tumors with mutation rates drawn above this range went extinct through natural selection.

The likelihood that tumors progress to cancer in the presence of deleterious passengers depends heavily on the initial population size N⁰ of the tumor. This dependence was studied previously (McFarland et al., 2014), where it was demonstrated that reasonable evolutionary simulations (those that progress to cancer >10% of the time, but <90% of the time) are restricted to a four-dimensional manifold N* within the five-dimensional phase space of parameters. For this reason, N⁰=N*(s_drivers, s_passengers, µT_drivers, µT_passengers) was determined by the other four parameters. To first order, this manifold is T_passengerss_paassengers/ (T_drivers s_drivers) (Weghorn and Sunyaev, 2017), however a more precise estimate (Equation S8 of McFarland et al., 2014) incorporating more precise estimates of Muller’s ratchet and the effects of hitchhiking on both driver and passenger accumulation rates, which does not exist in closed form was used. Additionally, at very low values of s_drivers, progression to cancer is limited by time, not by the accumulation of deleterious passengers. Hence, we assigned N⁰ such that:N0=MaxN0[Pcancer(N0/N∗)=0.5,tcancer¯(N0/N∗)=18,500generations]

Here, P_cancer and t_cancer – the likelihood and waiting time to cancer – are defined byEquation S8 and S12 respectively in McFarland et al., 2014. N⁰ was determined from these equations using Brent’s method. Figure 3—figure supplement 2 depicts the values of N⁰, which ranged from 1 to 100 for all simulations.

In tumors that progress to malignancy (N=10⁶), only fixed nonsynonymous mutations (present in all simulated cells) were recorded. We also recorded (i) the fitness effect of these mutations, (ii) the mean population fitness, (iii) the number of generations until malignancy, and (iv) the mutation rate. These two values were used to generate the number of synonymous drivers and passengers, where P(d_s = k)=Poisson[k; λ = µT_{drivers/passengers}/r t_MRCA] defines the number of synonymous drivers/passengers conferred, t_MRCA represents the number of division until the most recent common ancestor arose in the simulation, r=2.795 represents the ratio of nonsynonymous to synonymous loci within the genome, weighted by the genome-wide tri-nucleotide somatic mutation rate, and the Poisson PRN generator was defined above. In simulations where synonymous drivers could arise, a fraction of the recorded nonsynonymous mutations (ranging from 0% to 20%) were simply re-labeled as synonymous drivers (as opposed to nonsynonymous drivers). This was done, again, by Poisson sampling in proportion to the desired fraction for each cancer simulation.

20×20 combinations of s_drivers and s_passengers parameters were simulated (Figure 3—figure supplements 1–2). Simulations were repeated until 10,000 cancers at each parameter combination were obtained or until 10 million tumor populations were simulated. While we attempted to initiate tumors at a population size where the probability of progression to cancer was 50%, some parameter combinations still did not yield 10,000 cancers after 10 million attempts (i.e. P_cancer < 0.1%). These combinations were predominately at low values of s_drivers, which were far from the MLE estimate of s_drivers and represent unrealistic evolutionary scenarios: drivers cannot be weakly beneficial, relegated to only 300 genes, and still overcome deleterious passengers within 100 years. These simulations are annotated as ‘progression impossible’. Simulation parameter sweeps were performed for both the multiplicative and diminishing returns epistasis models. Twenty fractions of synonymous drivers were also generated (ranging from 0% to 20%). These fractions were generated by simply re-labeling the driver mutations which conferred fitness (generated during the simulation) as synonymous, instead of nonsynonymous.

For both simulated and observed data, we summarized dN/dS rates vs. mutational burden for drivers and for passengers by decade-sized bins: (0, 10], (10, 100], (100, 1,000]. Mutational burden for simulations was defined as the total number of substitutions (d_N+d_S) – exactly as it was defined for observed data. For simulated data, dN/dS = d_N/(d_S · r). Like observed data, dN/dS rates attenuated toward 1 for both drivers and passengers for all values of s_drivers and s_passengers.

Mutational burdens (MB) for simulated and observed data were summarized with the parameters of a negative binomial distribution, where P(MB=k;n,p)=(k+n−1n−1)pn(1−p)k . This distribution has been used previously to summarize the MB of human tumors (Turajlic et al., 2012) and exactly defines the expected number of mutations at transformation in a multi-stage model of tumorigenesis (Frank, 2007) when n drivers are needed for transformation and the probability that any mutation be a driver is 1 – p (Michor et al., 2005). Both n and p were used to summarize MB. These quantities were determined by maximum likelihood optimization of the probability mass function above over the support of mutational burdens of [1, 1,000] substitutions. The Han-Powell quasi-Newton least-squares method was used for optimization.

Age-dependent cancer incidence rates (CI) were summarized with the parameters of a gamma distribution, where PCI≤t;k,θ=1Γkγk,tθ . Here, γs,x=∫0xts-1e-tdt is the lower incomplete gamma function and Γ(k) = γ(k, ∞) is the regular gamma function. Similar to our summarization of mutational burdens, this distribution is a generalization of the exact waiting time to transformation expected from a multi-stage model of tumorigenesis when tumors arise at a uniform rate over time, require k drivers for transformation, and wait an average time of θ between drivers (Michor et al., 2005). This cumulative distribution function was fit to observed incidence rates for all patients above 20 years of age using the least squares numerical optimization defined above (all cancer sites combined, both sexes, all races, 2012–2016; Howlader, 2013). Patients under 20 years of age were excluded because cancers in these patients generally arise from germline predispositions to cancer, which are (i) not directly modeled by our simulations, (ii) not detected as somatic mutations, and (iii) result in age incidence curves that do not agree with a gamma distribution (Frank, 2007). Because all cancer simulations are initiated at t=0 (instead of uniformly in time, as is presumed in the multi-stage model), the simulated data was fit using the probability density function of this distribution (instantaneous derivative) using maximum likelihood and the optimization algorithm described above. The cumulative distribution, then, represents the expected age incidence cancer incidence rate when simulations begin at uniformly distributed moments in time and, thus, was used to generate Figure 3D. Only the shape parameter k was used in ABC (and θ was ignored), as this parameter only specifies the dimensionality of time (simulation time was measured in cellular generations, not years) and all values of θ in our simulations are equivalent under a gauge transformation. Additionally, we do not expect the exact times of incidence to be particularly informative as the time of transformation is generally somewhat earlier than the time of detection.

Like many Bayesian analyses, the main steps of an ABC analysis scheme are: (i) formulating a model, (ii) fitting the model to data (parameter estimation), and (iii) improving the model by checking its fit (posterior-predictive checks), and (iv) comparing this model to other models (Csilléry et al., 2012; Gelman, 2004).

The nine summary statistics described above were used to compare simulations to observed data. Agreement was summarized with a log-Euclidean distance, as all summary statistics resided on the domain [0, ∞) and log-transformation of the summary statistics minimized heteroscedasticity of the simulated data relative to a square-root or no transformation. Variance of the summary statistics was not normalized. ABC was performed using the ‘abc’ R package (Csilléry et al., 2012).

The rejection method (feedforward neural net) and tolerance (0.5) were chosen based on their capacity to minimize prediction error of the simulated data using leave-one-out CV (Figure 3—figure supplement 3). Ten-thousand instances of the neural network, which was restricted to a single layer, were initiated and the median prediction of these networks were used. These parameters were used for both model comparison and parameter inference. For parameter inferencing, the s_drivers and s_passengers prior values were log-transformed.

For the synonymous driver model, the base model (without synonymous drivers) was simply the lowest quantity of synonymous drivers (0%) in the parameter sweep of synonymous driver quantities (Figure 3—figure supplement 3B). The posterior probability mass of this value 0.017 was used as the one-sided p-value for the null hypothesis that these two models are equally predictive. Although the synonymous driver model agreed with the observed data slightly better, s_drivers and s_passengers parameters could not be inferred from the data because the potential for synonymous drivers destroys the utility of a dN/dS statistics, which is predicated on the notion that synonymous mutations are neutral. Virtually any value of dN/dS is attainable when the right combinations of selective pressures on nonsynonymous and synonymous are paired (Figure 3—figure supplement 3C).

All code for empirical analysis and generation of summary statistics are publicly available under the open-source MIT License at https://github.com/petrov-lab/cancer-HRI. (Tilk, 2022a copy archived at swh:1:rev:5d67f4a946e2d80efdc71c2ef689266678d8ff75). Code for simulations of tumor growth with advantageous drivers and deleterious passengers is also available at https://github.com/mirnylab/pdSim, (Tilk, 2022b copy archived at swh:1:rev:f08bd75aabf7213e253baf26d219c374a745c8d4).