Tumour mutational burden is overestimated by target cancer gene panels

Abstract

Background

Tumour mutational burden (TMB) has emerged as a predictive marker for responsiveness to immune checkpoint inhibitors (ICI) in multiple tumour types. It can be calculated from somatic mutations detected from whole exome or targeted panel sequencing data. As mutations are unevenly distributed across the cancer genome, the clinical implications from TMB calculated using different genomic regions are not clear.

Methods

Pan-cancer data of 10,179 samples were collected from The Cancer Genome Atlas cohort and 6,831 cancer patients with either ICI or non-ICI treatment outcomes were derived from published papers. TMB was calculated as the count of non-synonymous mutations and normalised by the size of genomic regions. Dirichlet method, linear regression and Poisson calibration models are used to unify TMB from different gene panels.

Results

We found that panels based on cancer genes usually overestimate TMB compared to whole exome, potentially leading to misclassification of patients to receive ICI. The overestimation is caused by positive selection for mutations in cancer genes and cannot be completely addressed by the removal of mutational hotspots. We compared different approaches to address this discrepancy and developed a generalised statistical model capable of interconverting TMB derived from whole exome and different panel sequencing data, enabling TMB correction for patient stratification for ICI treatment. We show that in a cohort of lung cancer patients treated with ICI, when using a TMB cutoff of 10 mut/Mb, our corrected TMB outperforms the original panel-based TMB.

Conclusion

Cancer gene-based panels usually overestimate TMB, and these findings will be valuable for unifying TMB calculations across cancer gene panels in clinical practice.

1. Introduction

Tumour mutational burden (TMB) is defined as the number of non-synonymous mutations within coding regions across the genome. It has been considered a promising predictor to stratify patients for immune checkpoint blockade therapy. The proposed underlying principle is that higher TMB production is likely to generate more neoantigens that the human immune system can recognise, thereby eliciting an anti-tumour immune response.¹ As the gold standard, TMB is measured by somatic mutations from whole-exome sequencing (WES) data. However, due to the high cost, sample availability, as well as long turnaround times, panel sequencing-based estimates of TMB are more commonly used in clinical practice.² The implementation of panel TMB as the proxy of WES TMB has been extensively evaluated.3, 4, 5, 6 Panel TMB shows a good correlation with WES TMB, however, the accuracy heavily depends on the size of the genomic region covered by the panel, with a size of between 1.5 and 3 Mb being optimal.⁷ Budczies et al. proposed a mathematical law showing that panel TMB decreased inversely proportional to the square root of the panel size and the square root of the TMB level, and they also suggested measures to control the limitation by panel design and TMB classification scheme.⁸ There is also great variation across panels from different laboratories for TMB estimation, and the variability within and between panel TMB values increases with the WES TMB values.⁹ From a real-world study, Stenzinger et al. investigated bridging factors to transform panel TMB to WES TMB values, and compared the WES TMB cutoff point to projected panel TMB cutoff point and estimate the potential misclassification. Nevertheless, there remains a lack of a flexible approach to harmonise these cutoff points.¹⁰

Generally, genes that are selected in panel sequencing designs are cancer-associated genes with the principal aim to identify the driver or actionable mutations. Many of these mutations would undergo positive selection during cancer development.¹¹ On the other hand, mutational processes in cancer genomes are substantially affected by local determinants, such as gene expression, replication timing, tri-nucleotide context and CpG content.¹² The regions selected for panel design might show great heterogeneity for these factors compared to exome-wide regions. Therefore, although cancer gene panel-derived TMB correlates well with WES TMB, how these mutational processes present in cancer genes affect final TMB estimation is not well established.

In this study, by analysing 10,179 cancer samples across 33 cancer types from The Cancer Genome Atlas (TCGA) cohort, we find that cancer gene-based panels usually overestimate TMB calculation even when known hotspot driver mutations are accounted for. We implement a statistical model that enables interconversion of TMB between any panel or exome, thereby providing a method to unify and normalise TMB estimation regardless of the genomic regions sequenced. We demonstrate the importance of TMB normalisation for TMB-based immune checkpoint inhibitors (ICI) treatment decisions in clinical settings.

2. Materials and methods

2.1. Data collection

Pan-cancer data of 10,179 samples with mutation annotation across 33 cancer types from TCGA were downloaded from the UCSC XENA browser (https://xenabrowser.net/datapages/). Dataset with clinical information and ICI treatment were obtained from Samstein et al.¹³ and Zehir et al.¹⁴ The source of all data is summarised in Supplementary Table 1, and the demographic characteristics of patients for ICI hazard ratio analysis are shown in Supplementary Table 2.

2.2. Definition of exome and panel region

As the data from the pan-cancer study were based on various versions of whole exome region capture kits, we restrict the analysis to only coding regions to avoid any mutation calling bias. The bed file of all coding regions was obtained from the UCSC table browser using the UCSC knownGene table. For MSK-IMPACT (MSK) and FoundationOne CDx (F1CDX) panels, gene lists were obtained from Yaeger et al.¹⁵ and FDA Technical Information (https://www.accessdata.fda.gov/cdrh_docs/pdf17/P170019S006C.pdf) respectively. Simulated cancer panels were designed by randomly selecting coding regions drawn from COSMIC cancer gene census. Simulated random panels are based on randomly selected regions drawn from genome-wide coding regions.

2.3. TMB calculation

TMB was measured as the count of non-synonymous mutations (including missense, nonsense, nonstop, splice site and indel), then normalised by the region size of the panel and reported as mutations per megabase (mut/Mb). For TMB calculation with and without hotspot mutations, hotspot mutations were obtained from the Cancer Hotspots database.^16,17

2.4. Measure of positive selection

The degree of positive selection was measured by the dNdScv R package and represented as a non-synonymous to synonymous mutations (dN/dS) ratio.¹⁸ Briefly, this package uses trinucleotide context-dependant substitution models to avoid common mutation biases affecting dN/dS, and the dN/dS ratio is quantified at the level of different regions from various panels.

2.5. Binomial regression model

Binomial regression model is estimated to predict mutation probability for each site within exome regions. Four explanatory variables, including CpG position, tri-nucleotide context, gene expression and replication timing, are adopted.

Let Y_s = (Y_s,1, . . ., Y_s,n) be the observed sequence sites for sample s, s=1, . . ., S, where n denotes the total number of positions sequenced in the exome. A binomial regression model was estimated on the mutations from the S patients using k explanatory variables. Thus, coefficient vector estimates ${\widehat{\beta }}_1$, . . ., ${\widehat{\beta }}_k$ are available.

2.6. Dirichlet method

The basic idea of the Dirichlet method is to borrow prior knowledge from an existing dataset. In this study, we weighted the predicted probabilities from the dataset to optimise the TMB evaluation.

Let k = 1, . . ., K denote the combinations of explanatory variables observed on the exome. Then let $n_k$ denote the number of positions for each combination. From the cohort binomial regression model, the mutation probability can be predicted for each combination, denoting this by $p_k$. Similarly, we estimate a binomial regression model on the panel data of sample M and predict the mutation probability for each combination, denoting this by $p_k^M$. Then, we obtain the new weighted predictions

$$p_k^{*}\left(w\right)=w{p}_k^M+\left(1-w\right)p_k$$

where 0 ≤$w$ ≤1. For combinations $k$ where $p_k^M$cannot be estimated, we set $p_k^{*}\left(w\right)$ = $p_k$.

Then, we estimate $T_M$ by:

$${\widehat{T}}_M\left(w\right)=\sum _{k=1}^K{n}_k{p}_k^{*}\left(w\right)$$

The optimal choice of $w$ depends on the size of the panel, how similar the properties of the panel are to the whole exome, and the similarity of the new sample $M$ to the observed cohort. A cross-validation scheme is used to identify the optimal $w$ in the following algorithm.

For $s$ = 1, . . ., $S$:

• (1)Set sample $s$ as the test sample and set the whole cohort except sample $s$ as the training sample.
• (2)Estimate $p_k$ on the whole exome data from the training sample and $p_k^s$ on the panel data from the test sample.
• (3)Compute $p_k^{*}\left(w\right)$=$w{p}_k^s+\left(1-w\right)p_k$ for all values of $w$.
• (4)Compute ${\widehat{T}}_s\left(w\right)$for all values of $w$.
• (5)Compute the loss function $l_s\left(w\right)=\left|{\widehat{T}}_s\left(w\right)-T_s\right|/T_s$ for all values of $w$, where $T_s$ is the true total mutation burden in sample $s$.

Finally, compute $L\left(w\right)=\sum _{s=1}^S{l}_s\left(w\right)$. The value for $w$ with the smallest $L\left(w\right)$ is optimal.

2.7. Linear regression model

Linear regression was employed to model the relationship between whole coding regions derived TMB and panel derived all mutations for each cancer type:

$$\mathrm{lm}(y\sim x)$$

where, y is the whole exome TMB, and x is panel-derived mutations normalised by the region size of the panel. All mutations from the panel were used as we found this model performed better than using non-synonymous+indel mutations alone.

2.8. Poisson calibration models

Given the true total mutation count $x_i^t$ obtained in sample $i$ by exome sequencing (denoted by superscript $t$), as well as the total sequence length $n^t$ of the target region, we modelled counts $x_i^t$ as using a Poisson model with overdispersion parameter $\theta$, as well as platform-specific calibration parameters $\left({\beta }_0^q,{\beta }_1^q\right)$ for targeted panel $q$ as follows:

$$x_i^t\sim \text{Poisson}\left(n^t{\mu }_i\right)$$

$${\mu }_i=\text{exp}\left({\beta }_0^q+{\beta }_1^q{r}_i^q\right)$$

$$\mathrm{E}\left\{x_i^t\right|r_i^q\}=n^t{\mu }_i$$

$$\text{Var}\left\{x_i^t\right|r_i^q\}=n^t{\mu }_i\theta .$$

The covariate $r_i^q$ in this calibration model is the log mutation rate observed in sample $i$ by targeted panel $q$, and it is in turn given by

$$r_i^q=\text{log}\left(\frac{x_i^q+\lambda }{n^q+2\lambda }\right),$$

where $x_i^q$ is the observed mutation count in sample $i$ with targeted panel $q$, and $n^q$ is the total sequence length of the targeted panel. $\lambda >0$ is a regularisation parameter that ensures that the logarithm is always defined. The covariate $r_i^q$ in the Poisson model is provided in log scale because our Poisson calibration model uses the log link function. After having trained the model and obtained the parameter estimates $\left({\widehat{\beta }}_0^q,{\widehat{\beta }}_1^q\right)$, we calibrate a new targeted log mutation rate ${\tilde{r}}_i^q$ observed on panel $q$and calculate the calibrated total mutation burden ${\widehat{b}}_i$ using the predicted mutation count ${\tilde{x}}_i^t$ by

$${\widehat{b}}_i=\frac{1}{n^t}\mathrm{E}\left\{{\tilde{x}}_i^t\right|{\tilde{r}}_i^q\}=\text{exp}\left({\widehat{\beta }}_0^q+{\widehat{\beta }}_1^q{\tilde{r}}_i^q\right).$$

Building on this base Poisson model, we also developed a two-layer Poisson calibration model that analyses the mutation type-specific counts: the number of C>A, C>T, C>G, T>A, T>C, and T>G non-stranded single-nucleotide variants (SNVs), as well as indels. The C>X, T>X, and indel counts are, in turn, constrained by the total number of C's and G's in the target genome, the total number of T's and A's, and the total number of bases, respectively. We denote the specific mutation count by exome sequencing (denoted by superscript $t$) as $x_i^{t,s}$ for mutation type $s$ in sample $i$, the number of bases that can give rise to the specific mutation type as $n^{t,s}$, and the specific log mutation rate as $r_i^{q,s}$. In the first layer of the calibration model, we fit a similar model as the above for each mutation type $s$ using $x_i^{t,s}$, $n^{t,s}$, and $r_i^{q,s}$ as follows:

$$\mathrm{E}\left\{x_{\mathit{i}}^{\mathit{t},\mathit{s}}|r_{\mathit{i}}^{\mathit{q},\mathit{s}}\right\}=n^{\mathit{t},\mathit{s}}\text{exp}({\beta }_{\mathit{0}}^{\mathit{q},\mathit{s}}+{\beta }_{\mathit{1}}^{\mathit{q},\mathit{s}}{r}_{\mathit{i}}^{\mathit{q},\mathit{s}})$$

The second layer then models the total mutation count $x_i^t$ by combining the specific mutation counts $x_i^{t,s}$:

$$\mathrm{E}\left\{x_i^t\right|x_i^{t,s}\}={\gamma }_o^q+\sum _s{\gamma }_s^q{x}_i^{t,s}.$$

After training the two-layer model, we obtain first-layer parameter estimates $({\widehat{\beta }}_0^{q,1},\dots ,{\widehat{\beta }}_0^{q,S},{\widehat{\beta }}_1^{q,1},\dots ,{\widehat{\beta }}_1^{q,S})$ and second-layer parameter estimates $\left({\widehat{\gamma }}_0^q,\dots ,{\widehat{\gamma }}_S^q\right)$ for panel $q$. Using these parameters, we calibrate a new specific mutation count ${\tilde{x}}_i^{q,s}$ by

$${\widehat{b}}_i=\frac{1}{n^t}\mathrm{E}\left\{{\tilde{x}}_i^t\right|{\tilde{r}}_i^{q,1}\dots {\tilde{r}}_i^{q,S}\}=\frac{1}{n^t}\left({\widehat{\gamma }}_0^q+\sum _s{\widehat{\gamma }}_s^q{n}^{t,s}\text{exp}\left({\widehat{\beta }}_0^{q,s}+{\widehat{\beta }}_1^{q,s}{\tilde{r}}_i^{q,s}\right)\right)$$

$${\tilde{r}}_i^{q,s}=\text{log}\left(\frac{{\tilde{x}}_i^{q,s}+\lambda }{n^{q,s}+2\lambda }\right).$$

2.9. Evaluation of models

Deviation from true TMB for panel and root mean squared error (RMSE) were evaluated by leave-one-out cross-validation. We evaluated each calibration model by training the model on a subset of samples (training data) and evaluating the model predictions on held-out samples. For analyses within one cancer type, we used leave-one-out cross-validation. Additionally, in order to assess the generalizability of models across cancer types, we also performed cross-validation by leaving one cancer type out in each round, training the model on samples from the other cancer types, and evaluating the model on samples from the held-out cancer type.

For each held-out sample $s$, we predicted the TMB, denoted ${\widehat{T}}_s$, using a calibration model, and the deviation of sample $s$ is calculated as:

$$d_s=\frac{\left|{\widehat{T}}_s-T_s\right|}{T_s}$$

where $T_s$ is the true TMB calculated from the whole exome for sample $s$. The deviation is computed as:

$$D=\sum _{s=1}^S{d}_{\mathit{s}}$$

where $S$ is the total number of samples in the held-out set.

RMSE is calculated as:

$$\text{RMSE}=\sqrt{\frac{{\sum }_{s=1}^S{\left(\widehat{T_s}-T_s\right)}^2}{S}},$$

and normalized RMSE is calculated as:

$$\text{NRMSE}=\frac{\text{RMSE}}{\frac{1}{S}{\sum }_{s=1}^S{T}_s}.$$

2.10. Hazard ratio calculation

A cohort of patients with either ICI or non-ICI outcome^13,14 was used to evaluate the effect of ICI treatment on patient survival across different TMB cutoff points. Cox proportional hazard regression analysis was used to compute the hazard ratio of ICI treatment for overall survival.

2.11. Configuration of web portal and R package for TMB correction

The web portal is created by Shiny package in R and can be found at: https://cancergenomics-explore.shinyapps.io/shiny_tmb/. Briefly, we configure the two-layer Poisson regression in the portal that can model the relationship between any two genomic regions based on TCGA mutation data. The user can inter-convert the observed TMB to any other panel or genomic regions of interest. The scripts and tutorials for R the package are at https://github.com/jasonwong-lab/TMB/tree/master.

3. Results

3.1. Cancer gene panels overestimate tumour mutational burden

In order to compare TMB derived from common cancer panels with that are derived from WES data, we first mimic MSK and F1CDX panels, which are largely comprised of cancer-associated genes.^3,4 For each TCGA sample, we calculated the TMB based on non-synonymous mutations that overlapped with the coding regions of the genes in the panel. We found that panel-derived TMB is significantly inflated compared to WES TMB (P < 0.001, paired t-test, Fig. 1A). After removing hotspot mutations, it is still significantly higher than the TMB derived from randomly selected coding regions of similar size (∼1.2 Mb) (P < 0.05 and P < 0.001 for MSK and F1CDX, respectively, paired t-test, Fig. 1A). To further validate this, we simulated three additional panels (with size of ∼1.5Mb) based on COSMIC Cancer Gene Census (n = 723 genes) and observed similar TMB overestimation (Fig. 1B). We next examined the ratio of dN/dS, which represents the degree of positive selection for mutations across different panels. Strong positive selection (dN/dS > 1) was observed for mutations in cancer gene-based panels, while there is less positive selection for mutations within WES and a panel based on randomly selected genes (Fig. 1C). As the distribution of mutations across coding regions is not even and enriched in cancer genes, we next checked the potential overestimation based on the FDA-approved 10 mut/Mb cutoff. A large fraction of patients is overestimated by cancer panels across different cancer types (0–17% for MSK and 0–27% for F1CDX), which is only slightly alleviated by removing known cancer driver mutation hotspots (Fig. 1D). By contrast, a panel based on randomly selected genes shows no evidence of overestimation (Fig. 1D).

Fig. 1.

Cancer gene based panels overestimate TMB. (A) The distribution of median difference between TMB derived from various panels and WES TMB across different cancer types. TMB calculated with hotspots (With HS) and without hotspots (No HS) are indicated. (B) The distribution of median difference between simulated cancer panels TMB and WES TMB across different cancer types. TMB calculated with hotspots (With HS) and without hotspots (No HS) are indicated. (C) dN/dS ratio for mutations within MSK, F1CDX, panel based on randomly selected cancer genes, panel based on randomly selected genes, as well as whole exome across cancer types. (D) Fraction of overestimated patients across cancer types with TMB being measured by different panels (high TMB ≥10 mut/Mb). Panels with * indicates the removal of hotspots when calculating TMB. P-value is calculated by two-tailed Student's t-test. * 〈0.05, *** <0.001, n.s. 〉 0.05. ACC, adrenocortical carcinoma; BLCA, bladder urothelial carcinoma; BRCA, breast invasive carcinoma; CESC, cervical squamous cell carcinoma and endocervical adenocarcinoma; CHOL, cholangiocarcinoma; COAD, colon adenocarcinoma; DLBC, lymphoid neoplasm diffuse large B-cell lymphoma; dN/dS, non-synonymous to synonymous mutations; ESCA, oesophageal carcinoma; HNSC, head and neck squamous cell carcinoma; F1CDX, FoundationOne CDx panel; GBM, glioblastoma multiforme; KICH, kidney chromophobe; KIRC, kidney renal clear cell carcinoma; KIRP, kidney renal papillary cell carcinoma; LAML, acute myeloid leukaemia; LGG, brain lower grade glioma; LIHC, liver hepatocellular carcinoma; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; MESO, mesothelioma; MSK, MSK-IMPACT panel; Mut/Mb, mutations per megabase; OV, ovarian serous cystadenocarcinoma; PAAD, pancreatic adenocarcinoma; PCPG, pheochromocytoma and paraganglioma; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; SARC, sarcoma; SKCM, skin cutaneous melanoma; STAD stomach adenocarcinoma; TGCT, testicular germ cell tumors; THCA, thyroid carcinoma; THYM, thymoma; TMB, tumour mutation burden; UCEC, uterine corpus endometrial carcinoma; UCS, uterine carcinosarcoma; UVM, uveal melanoma; WES, whole exome sequencing.

3.2. Optimisation of TMB estimation by modelling regional variations in mutation rate

To address the discrepancy in TMB estimation between WES and cancer gene panels, we investigate different approaches to optimise TMB estimation. As several well-established local genomic and epigenomic factors can affect local mutation prevalence,¹⁹ we also evaluate the influence of these factors on TMB derived from target region panels.

We first develop a site-specific binomial regression model to predict mutation probability for each position by considering several local determinants as explanatory variables, including CpG position, tri-nucleotide context, gene expression, and replication timing from colorectal cancer (CRC) whole-exome sequencing data. As a result, the binomial regression model was estimated on > 1.7 billion observations.

In order to improve TMB estimation derived from panels designed by target regions, we adopt a Dirichlet method, which combines the predicted mutation probabilities from the cohort and a new input sample. The basic idea for the Dirichlet method is to borrow information from an existed cohort and then weigh the information for the sample from the cohort to obtain the best TMB prediction. Due to the property of the Dirichlet approach that borrows prior knowledge from a cohort, its performance is heavily dependant on the size of the panel, the similarity of properties between the panel and whole exome, as well the similarity of the new sample to the observed cohort. The procedures of this method are detailed in the method section.

Binomial regression models are estimated for microsatellite stable (MSS) and microsatellite instability-high (MSI-H) colorectal cancer samples separately as they represent cancer samples that show great variation in genomic characteristics and mutational burden. A cross-validation scheme is adopted to find the best weight to adjust the predicted probabilities for the new sample, and the predicted WES TMB is calculated by the optimised probabilities. The results showed that WES TMB prediction by the model improved on the direct calculation of the observed TMB for both MSI-H and MSS (Table 1, Supplementary Fig. 1A and B). In particular, the MSI-H model performed better than the MSS model (NRMSE 0.14 versus 0.33, Table 1, Supplementary Fig. 1A and B), suggesting that low mutation counts are more difficult to model.

Table 1.

Loss of function and NRMSE for uncalibrated vs. calibrated TMB estimates by the Dirichlet method or linear regression.

		Uncalibrated	Dirichlet	Linear
Loss of function	CRC MSS (n = 425)	175.1838	128.3783	107.4269
	CRC MSI-H (n = 74)	10.2787	9.6432	10.0073
NRMSE	CRC MSS (n = 425)	0.4924	0.3347	0.3073
	CRC MSI-H (n = 74)	0.1539	0.1429	0.1365

Abbreviations: CRC, colorectal cancer; MSI-H, microsatellite instable; MSS, microsatellite stable; NRMSE, normalised root mean squared error.

3.3. Generalised regression models for TMB estimation

One issue with the Dirichlet method is that it is difficult to generalise to all cancer types as cancer type-specific epigenetic data is required, and in practice, the cancer type may not always be known. Consequently, we sought to test a simple linear regression model to predict WES TMB based on panel sequencing data, from which the constant parameter that scales the panel size to the size of the whole exome is estimated. Based on this simple regression model, the predicted WES TMB was calculated and compared to the observed WES TMB and panel TMB. The predicted WES TMB was found to compare favourably to the Dirichlet method (Fig. 2A and B, Table 1).

Fig. 2.

Comparison of observed WES TMB, observed panel TMB and predicted WES TMB. TMB for MSS (A) and MSI (B) colorectal cancers, of which the predicted WES TMB is calculated from a linear calibration model. TMB for MSS (C) and MSI (D) colorectal cancers, of which the predicted WES TMB is calculated from a two-layer Poisson calibration model. NRMSE comparing observed and predicted WES TMB using different models based on training with all samples (E) or cross-validation with each cancer type held out (F). CRC, colorectal cancer; MSI, microsatellite instable; MSS, microsatellite stable; NRMSE, normalised root mean square error; TMB, tumour mutation burden; WES, whole exome sequencing.

To improve on the linear model, we further developed Poisson calibration models. We settled on a two-layer Poisson calibration model that estimates the specific mutation burden of each mutation type (e.g. C>A, C>T, C>G, T>A, T>C and T>G) and combines these specific burdens into an overall TMB (Fig. 2C and D). By doing so, we can account for the propensity of different cancer types to accumulate different types of mutations, owing to differences in their underlying mutational processes. Indeed, our two-layer Poisson calibration model achieved the lowest normalised root-mean squared error (NRMSE), across all targeted panels (Fig. 2E). In order to assess whether this performance is due to overfitting, we further performed cross-validation by training the model on samples from all but one cancer type, and evaluating the calibration error on samples of the held-out cancer type (Fig. 2F). Accordingly, our two-layer Poisson model can generalise across cancer types and calibrate TMB accurately regardless of cancer type and the underlying mutational processes.

We note here that some targeted panels, such as the DFCI panel, may have target region sequence content that closely resembles the exome, and for these panels, TMB calibration may not be necessary (Fig. 2F).

3.4. Evaluation of two-layer Poisson calibration model to harmonise TMB estimation

To make our two-layer Poisson calibration model readily accessible, we developed a web portal (https://cancergenomics-explore.shinyapps.io/shiny_tmb/) and R package (https://github.com/jasonwong-lab/TMB). Screenshots from web portal for two different modules are shown in Supplementary Fig. 2.

Adjusted TMB by this model shows significantly less deviation from WES TMB (Fig. 3A) and a smaller RMSE (Fig. 3B) across cancer types. We then applied our model to a previously published dataset (n = 1662) with MSK panel-based mutations and ICI treatment outcomes available.¹³ As expected, the panel TMB is higher than the predicted TMB across all the cancers we examined (Fig. 3C, Supplementary Table 3), and a substantial fraction of patients have an overestimation of TMB when using the cutoff of 10 mut/Mb, including 17.7% (62/350) of lung cancers (Fig. 3D).

Fig. 3.

TMB evaluation is optimised by the two-layer Poisson model. (A) Comparison of absolute TMB deviations that are calculated by panel TMB vs. predicted TMB across cancer types. P-value is calculated by paired Student's t-test. (B) Comparison of RMSE that are calculated by panel TMB vs. predicted TMB across cancer types. P-value is calculated by paired Student's t-test. (C) Distribution of panel TMB, panel TMB with hotspots removed and predicted TMB for different cancer types. The significance level of difference is indicated in Supplementary Table 3. (D) Profile of overestimated samples across different cancer types by using the cutoff value of 10 mut/Mb. The number of overestimated samples for each cancer type is indicated. (E) Original TMB derived from MSK-410 panel and predicted TMB for WES and other panels. ACC, adrenocortical carcinoma; BLCA, bladder urothelial carcinoma; BRCA, breast invasive carcinoma; CESC, cervical squamous cell carcinoma and endocervical adenocarcinoma; CHOL, cholangiocarcinoma; COAD, colon adenocarcinoma; DLBC, lymphoid neoplasm diffuse large B-cell lymphoma; ESCA, oesophageal carcinoma; HNSC, head and neck squamous cell carcinoma; GBM, glioblastoma multiforme; KICH, kidney chromophobe; KIRC, kidney renal clear cell carcinoma; KIRP, kidney renal papillary cell carcinoma; LAML, acute myeloid leukaemia; LGG, brain lower grade glioma; LIHC, liver hepatocellular carcinoma; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; MESO, mesothelioma; MSK, MSK-IMPACT panel; Mut/Mb, mutations per megabase; OV, ovarian serous cystadenocarcinoma; PAAD, pancreatic adenocarcinoma; PCPG, pheochromocytoma and paraganglioma; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; pWES, predicted whole exome sequencing; RMSE, root mean squared error; SARC, sarcoma; SKCM, skin cutaneous melanoma; STAD, stomach adenocarcinoma; TGCT, testicular germ cell tumors; THCA, thyroid carcinoma; THYM, thymoma; TMB, tumour mutation burden; UCEC, uterine corpus endometrial carcinoma; UCS, uterine carcinosarcoma; UVM, uveal melanoma.

To compare TMB values that are derived from different panels, we implement a function in the web portal that can inter-convert TMB between panels or genomic regions. We selected a sample that was tested by MSK-410 panel and then predicted the TMB value for different panels or genomic regions based on the model. Using mutations from TCGA samples as models, we show that the predicted TMB show high concordance with the observed TMB across all panel conversions (R2 > 0.98, Supplementary Fig. 3). In some cases, there can be substantial variation between an uncorrected TMB and the predicted TMB. In the example shown, the TMB derived from test panel (MSK-410) was over 10 mut/Mb, while the predicted WES TMB was less than 5 mut/Mb and all other panels were less than 10 mut/Mb (Fig. 3E). This means that for this case if the 10 mut/Mb cutoff for pembrolizumab based on the F1CDX panel approved by the FDA was being used, our calibration model would correctly reclassify this patient as TMB-low and, therefore not ideal for treatment with pembrolizumab.

3.5. Clinical implication of TMB threshold derived from different genomic regions

To further explore the clinical implications of not correcting TMB based on the panel being used, we first obtained mutations from 6867 cancer patients across nine tissue origins with MSK-IMPACT panel-covered regions sequenced.^{13, 14} All the patients have the treatment outcomes with either ICI or non-ICI (Supplementary Table 4). After calculating TMB from the panel sequencing data, we predicted WES TMB based on our cancer type and panel type-specific regression model for each sample. Then, we assessed the hazard ratio as a measure of the effect size of ICI treatment, and we tested how the observed panel TMB vs. predicted WES TMB affected the optimal TMB threshold and the hazard ratio estimates.

As lung cancers with high TMB generally benefit from ICI treatment,²⁰ we first checked the hazard ratio of ICI treatment for lung cancer patients with different TMB thresholds. Without TMB adjustment, patients derive benefit from ICI treatment when the observed panel TMB reaches 10 mut/Mb. However, after adjustment using our model a TMB threshold of > 7 mut/Mb was sufficient for ICI treatment benefit based on predicted WES TMB (Fig. 4A). Importantly, when stratifying lung cancer patients treated with ICIs based on a cutoff of 10 mut/Mb, while the overall survival between patients with TMB ≥10 mut/Mb (TMB-H) and TMB < 10 mut/Mb (TMB-L) is significantly different (P < 0.01, Fig. 4B and C), the hazard ratio is substantially closer to 0 for predicted TMB-H compared with the original TMB-H group, indicating a stronger improvement in survival (hazard ratio, 0.36; 95% CI: 0.19, 0.68 versus 0.56; 95% CI: 0.4, 0.78, Supplementary Fig. 4A and B). We further assessed the thresholds of TMB cutoff on the benefit of ICI treatment for other cancer types. The amount of benefit and TMB cutoffs for effective ICI treatment was different for each cancer type. Generally, variation between the observed panel-based TMB and predicted WES-based TMB was ubiquitously present (Supplementary Fig. 5). These results suggest that TMB as a predictor to stratify patients for ICI treatment is heavily dependant on the approach from which the TMB is calculated. As cancer panels generally inflate the TMB estimation compared to WES, the TMB cutoff value for favourable ICI treatment would be smaller than 10 mut/Mb when the TMB is estimated from WES data.

Fig. 4.

Clinical implication of TMB threshold derived from panels and predicted whole exome. (A) HR of ICI treatment on patient survival for different TMB cutoff values for LUAD. Overall survival time comparison between groups that are separated by panel TMB (B) and predicted TMB (C) using the cutoff value as 10 mut/Mb for lung cancer. The HR of TMB for overall survival from multivariate Cox regression model is indicated. CI, confident interval; ICI, immune checkpoint inhibitor; HR, hazard ratio; LUAD, lung cancer; N.S., P > 0.05; ObsTMB, observed panel TMB; PreTMB, corrected TMB by the two-layer Poisson model; Pval, P value; TMB, tumour mutation burden; TMB-H, TMB high; TMB-L, TMB low.

4. Discussion

TMB as a predictor has been widely used to stratify patients to receive ICI treatment in clinical practice. As the calculation of TMB is very straightforward, that is, simply counting the number of mutations across the cancer genome, it can be measured by either target regions or whole-exome regions. We show that cancer gene-based panels overestimate TMB calculation, potentially leading to misclassification of a subset of patients for ICI therapy. The overestimation is largely due to the increased frequency of positively selected mutations in cancer genes, which cannot simply be addressed by the removal of known hotspot mutations.

Somatic mutations for TMB calculation in the cancer genome accumulate throughout the whole life of a cancer patient and undergo positive selection.²¹ Cancer cells with growth advantage are positively selected, and as these cells expand, mutations that confer growth advantage are also amplified. This would cause these positively selected mutations to be distributed unevenly across cancer genomes and potentially be enriched in cancer-related genes. On the other hand, well-established local genomic determinants also can contribute to variations in the distribution of mutations across the genome.¹⁹ Taking all these together, positive selection in combination with local determinants underlies the discrepancy between panel-derived TMB and WES-derived TMB.

We tried different approaches to address the discrepancy in TMB between panel sequencing and WES. First, we consider the effect of different local determinants on TMB estimation using the Dirichlet approach, which combines the mutation prediction probability from a binomial regression model to optimise TMB evaluation. In spite of considering these local determinants, we found that a two-layer Piosson model incorporating only genomic sequence characteristics yields similar performance and can more easily be generalised to any cancer type. The performance of the Dirichlet method heavily depends on prior knowledge, particularly the size of the panel, the similarity of properties between the panel and whole exome, as well the similarity of the new sample to the observed cohort. For these reasons, the Dirichlet approach may be difficult to implement in practice, as there is generally substantial heterogeneity amongst cancer patients.²² In contrast, the Poisson calibration model considers the sequence contents of the target regions of the panels and the whole exome, by estimating calibration parameters that corrects the panel TMB to the whole exome TMB. For tumours with low TMB, the Poisson model performs better than the Dirichlet approach, while they present comparable performance for tumours with relative high TMB, e.g. MSI-H CRC samples. As there is high heterogeneity amongst tumours of different types, the Poisson model is more applicable to different cancer types. Thus, our model can help to harmonise TMB derived from panels and WES and overcome discrepancies in a cancer-type-agnostic and panel-specific manner.

We also asessed potential clinical implications of the discrepancy between the panel and WES TMB estimates. The cutoff value of TMB at 10 mut/Mb has been approved by the FDA to stratify patients to receive ICI treatment. The determination of this cutoff value is largely based on clinical trials with cancer panel sequencing data. Recently, Friedman et al. assessed the therapeutic effect of atezolizumab for high TMB patients with advanced solid tumours and found that patients with TMB ≥ 16 mut/Mb have a durable response to atezolizumab. In comparison, there is limited activity for patients with TMB ≥ 10 and < 16 mut/Mb.²³ In spite of this, the variation of TMB derived from different genomic regions might lead to patient misclassification. We re-evaluate the benefits of ICI treatment for patients with different TMB values derived from targeted panels, before and after calibration using our developed models. Consistent with a previous report,²⁴ patients could benefit from ICI treatment when the panel TMB is at least 10 mut/Mb. But for the predicted WES TMB, patients usually benefit when the TMB is at least 7 mut/Mb. This suggests that many patients with TMB derived from WES data will be excluded from ICI treatment, even though they may in fact benefit from ICI treatment. This discrepancy has strong implications for clinical practice. Due to limitations of data investigated in this study, prospective trials are required to further confirm the clinical implications.

Additionally, mounting evidence demonstrates that high TMB alone is not sufficient as a biomarker to predict ICI treatment response.^25,26 Based on an evaluation of 137 patients with advanced CRC who were treated with ICIs, Benoit Rousseau et al. found that for all TMB high (10 mut/Mb) patients, only these with mismatch repair deficiency status or with pathogenic mutations in polymerase ε (POLE) or polymerase δ1 (POLD1) have favourable survival.²⁷ Extension of this analysis to 1661 patients with various cancer types supports this observation and further revealed that, in addition to these two genetic subtypes, only patients with hypermutated tumours who benefited from ICI had cancers strongly associated with environmental exposures, such as UV radiation or cigarette smoking. Therefore, the aetiology of the hypermutation phenotype may influence the benefit of ICI treatment.

In summary, we demonstrated that substantial discrepancies may exist in TMBs calculated from exome vs. targeted panel sequencing. Cancer panels generally overestimate TMB estimation compared to WES. The overestimation is largely due to the positive selection for mutations in cancer genes. Importantly, we developed a statistical model to harmonise TMB estimation, which could benefit clinical practice.

Declaration of competing interest

The authors declare that they have no conflict of interests.