A quantitative and multiplexed approach to uncover the fitness landscape of tumor suppression in vivo

Abstract

Cancer growth is a multi-stage, stochastic evolutionary process. While cancer genome sequencing has been instrumental in identifying the genomic alterations that occur in human tumors, the consequences of these alterations on tumor growth remains largely unexplored. Conventional genetically engineered mouse models enable the study of tumor growth in vivo, but they are neither readily scalable nor sufficiently quantitative to unravel the magnitude and mode of action of many tumor suppressor genes. Here, we present a method that integrates tumor barcoding with ultra-deep barcode sequencing (Tuba-seq) to interrogate tumor suppressor function in mouse models of human cancer. Tuba-seq uncovers genotype-dependent distributions of tumor sizes with great precision. By combining Tuba-seq with multiplexed CRISPR/Cas9-mediated genome editing, we quantified the effects of eleven tumor-suppressor pathways that are frequently altered in human lung adenocarcinoma. With unprecedented resolution, parallelization, and precision Tuba-seq enables broad quantification of tumor suppressor gene function.

INTRODUCTION

Genome sequencing has catalogued the somatic alterations in human cancers and identified many putative tumor suppressor genes^1–3. However, the identification of recurrent genomic alterations does not necessarily indicate their functional importance to cancer growth, and the impact of gene inactivation remains difficult to glean from cancer genome sequencing data alone^4,5.

The impacts of tumor suppressor gene losses on neoplastic growth have been investigated using knockdown, knockout, and overexpression studies in cell lines as well as in genetically engineered mouse models. The near-optimal growth of cancer cell lines in culture, widespread pre-existing genetic and epigenetic changes, and the lack of the autochthonous microenvironment limit the ability of these systems to provide insight into how tumor suppressor genes constrain the expansion of tumors in vivo. In contrast, genetically engineered mouse models of human cancer enable the introduction of defined genetic alterations into normal adult cells which results in the initiation and growth of tumors within their natural in vivo setting⁶. While Cre/loxP-based genetically engineered mouse models have become a mainstay for the analysis of tumor suppressor gene function, these systems are neither readily scalable nor sufficiently quantitative.

Recently, CRISPR/Cas9-mediated genome editing in somatic cells has increased the throughput of in vivo analyses of gene function in autochthonous cancer models^7–10. While these systems increase the scale of in vivo functional analyses, they continue to rely on relatively crude measurements of tumor growth, limiting their application to the analysis of tumor suppressors with the most dramatic effects. The lack of rigorously quantitative systems to analyze tumor suppressor function in vivo has precluded a broad understanding of tumor suppressor pathways that constrain tumor growth.

In other settings, molecular barcoding has enabled precise, multiplexed quantification of evolutionary fitness, selection, and clonal growth^11–17. Here, we describe Tuba-seq which combines tumor barcoding and high-throughput sequencing with genetically engineered mouse models to quantify tumor growth with unprecedented resolution. Precise quantification of individual tumor sizes uncovered the impact of inactivating different tumor suppressor genes. Integration of these methods with multiplexed CRISPR/Cas9-mediated genome editing enabled the parallel inactivation and functional quantification of a panel of putative tumor suppressor genes. This method is a rapid, multiplexed, and highly quantitative platform to study the impact of genetic alterations on cancer growth in vivo and uncover novel modes of tumor growth.

RESULTS

Tumor barcoding with ultra-deep barcode sequencing (Tuba-seq) enables precise and parallel quantification of tumor sizes

Oncogenic KRAS is a key driver of human lung adenocarcinoma, and early stage lung tumors can be modeled using LoxP-Stop-LoxP Kras^G12D knock-in mice (Kras^LSL-G12D/+) in which expression of Cre in lung epithelial cells leads to the expression of oncogenic Kras^{G12D 18,19}. LKB1 and P53 are frequently mutated tumor suppressors in human lung adenocarcinomas (Supplementary Fig. 1a)²⁰. Additionally, Lkb1- and p53-deficiency increases tumor burden in mouse models of oncogenic Kras^G12D-driven lung tumors^21–22. In Viral-Cre-induced mouse models of lung cancer large number of tumors can be initiated simultaneously and individual tumors can be stably tagged by lentiviral-mediated DNA barcoding^23,24. Therefore, we determined whether high-throughput sequencing of the lentiviral barcode region from bulk tumor-bearing lungs could quantify the number of neoplastic cells within each uniquely barcoded tumor (Supplementary Fig. 1b).

To interrogate the growth of oncogenic Kras^G12D-driven lung tumors as well as the impact of Lkb1- and p53-deficiency on tumor growth, we initiated lung tumors in Kras^LSL-G12D/+;Rosa26^LSL-Tomato (KT), KT;Lkb1^flox/flox (KLT), and KT;p53^flox/flox (KPT) mice with a library of Lentiviral-Cre vectors containing >10⁶ unique barcodes (Lenti-mBC/Cre; Fig. 1a and Supplementary Fig. 1b). KT mice developed widespread hyperplasias and small tumor masses (Fig. 1b and Supplementary Fig. 1c). Interestingly, while KLT mice had large tumors of relatively uniform size, KPT mice had a very diverse range of tumor sizes (Fig. 1b).

Figure 1. Tuba-seq combines tumor barcoding with high-throughput sequencing to allow parallel quantification of tumor sizes.

a, Schematic of Tuba-seq pipeline to assess lung tumor size distributions. Tumors were initiated in Kras^LSL-G12D/+;Rosa26^LSL-Tomato (KT), KT;Lkb1^flox/flox (KLT), and KT;p53^flox/flox (KPT) mice with Lenti-mBC/Cre (a pool of lentiviral vectors containing ~10⁶ random 15-nucleotide DNA barcodes (BC)). Tumor sizes were calculated via bulk barcode sequencing.

b, Fluorescence dissecting scope images of lung lobes from KT, KLT, and KPT mice with Lenti-mBC/Cre initiated tumors. Lung lobes are outlined with white dashed lines. The titer of Lenti-mBC/Cre is indicated. Different titers were used in different genetic background to generate approximately equal total tumor burden despite differences in overall tumor growth. Scale bars in upper panels = 5 mm. Scale bars in lower panels = 1 mm.

c, Tumor size distributions in KT, KLT, and KPT mice (number of mice per group is indicated). Each dot represents a tumor. The area of each dot is proportional to the number of cancer cells in each tumor. A dot corresponding to the approximate number of cancer cells in a 1mm diameter spherical tumor is shown to the right of the data for reference.

To quantify the neoplastic cell number in every lesion using high-throughput sequencing, we PCR-amplified the integrated lentiviral-barcode region from bulk lung DNA from each mouse and sequenced this to an average depth of >10⁷ reads/mouse (Fig. 1a and Supplementary Note). Tumor sizes varied by over one-thousand-fold (Fig. 1c). Barcode reads from small lesions could represent unique tumors or be generated from recurrent sequencing errors of similar barcodes from larger tumors. To minimize the occurrence of these spurious tumors, we aggregated reads expected to be derived from the same tumor barcode using an algorithm that generates a statistical model of sequencing errors (DADA2: Fig. 2 and Supplementary Fig. 2)²⁵. To enable the conversion of read count to cancer cell number, we added cells with known barcodes to each lung sample at a defined number, prior to tissue homogenization and DNA extraction, and normalized tumor read counts to “benchmark” read counts from these cells (Fig. 1a and Supplementary Fig. 3).

Figure 2. Tuba-seq is a robust and reproducible method to quantify tumor sizes.

a, DADA2, a denoising algorithm designed for deep sequencing of amplicon data, eliminates recurrent read errors that can appear as spurious tumors. Cell lines with known barcodes were added to each lung sample (5×10⁵ cells each). Recurrent read errors derived from these known barcodes generate spurious tumors which are greatly reduced by DADA2

b,c, Individual lesion sizes (b) and size profiles of tumors at the indicated percentiles (c) of technical replicate sequencing libraries prepared from an individual bulk lung sample.

d, Analysis of the effect of variation in read depth, GC content of the DNA barcodes, and diversity of the barcode library on tumor size calling. Tumors were partitioned into thirds corresponding to high, moderate, and low levels of each technical parameter. Whiskers capped at 1.5 IQR.

e, Size distributions across five KLT mice. Sizes of the tumors at the indicated percentiles in individual mice are connected by a line.

f, Tumors in each KLT mouse were partitioned into two groups (see Methods) and the profiles of these groups were compared. Sizes of the tumors at the indicated percentiles in an individual mouse are connected by a line.

Tuba-seq is highly reproducible between technical replicates and is insensitive to typical variation in many technical variables (Figure 2b–d, Supplementary Fig. 4 and Supplementary Note). Tumor size distributions were also highly reproducible between mice of the same genotype (R² >0.98; Figure 2e, Supplementary Fig. 4g and Supplementary Note). Indeed, unsupervised hierarchical clustering of size distributions separated mice according to their genotype, even when tumors were induced with different Lenti-mBC/Cre titers (Supplementary Fig. 4d). Differences in the spectrum of tumor sizes between mice of the same genotypes were greater than the differences between two fractions of tumors within the same mouse indicating that Tuba-seq is more precise than the intrinsic variability between mice (Fig. 2e,f). Thus, Tuba-seq rapidly and precisely quantified the number of neoplastic cells within thousands of lung lesions in KT, KLT, and KPT mice (Fig. 1c, Supplementary Fig. 4c and Supplementary Note).

Analysis of tumor sizes uncovers two modes of tumor suppression

To assess the effect of p53- or Lkb1-deficiency on tumor growth, we calculated the number of neoplastic cells in the tumors at different percentiles within the distribution. While tumors in KLT mice were consistently larger than KT tumors, deletion of p53 allowed only a small fraction of tumors to grow to exceptional sizes (Fig. 1c and Fig. 3).

Figure 3. Massively parallel quantification of tumor sizes enables probability distribution fitting across multiple genotypes.

a, b, Tumor size at the indicated percentile in KLT (n=5) mice (a) and KPT (n=3) mice (b) versus tumor size at the indicated percentile in KT mice (n=7). Each percentile was calculated using all tumors from all mice of each genotype 11 weeks after tumor initiation with Lenti-mBC/Cre.

c, Tumor sizes at the indicated percentiles for each genotype relative to KT tumors at the same percentiles. Error bars are 95% confidence intervals obtained via bootstrapping. Percentiles that are significantly differently from the corresponding KT percentiles are in color.

d, Tumor size distributions were most closely fit by a lognormal distribution. Tumors in KLT mice are best described by a lognormal distribution throughout their entire size spectrum (middle). The tumor size distributions in KT mice (left) and KPT mice (right) were better explained by combining a lognormal distribution at smaller scales with a power-law distribution at larger scales. Power-law relationships decline linearly on log-log axes, consistent with rare, yet very large tumors within the top ~1% of tumors in KT mice and ~10% of tumors in KPT mice. Note, only tumors in KPT mice exceed 10⁶ cancer cells after 11 weeks, consistent with p53-deficiency enabling the generation of the largest tumors in this study.

To better understand the difference in p53- and Lkb1-deficient tumor growth, we defined the mathematical distributions that best fit the tumor size distributions in KT, KLT, and KPT mice. Lkb1-deficient tumors were lognormally distributed across the full range of the distribution consistent with exponential tumor growth with normally distributed rates (Fig. 3d). To estimate average tumor size without allowing very large tumors to greatly shift this metric, we calculated the maximum likelihood estimator of the mean number of cancer cells given a lognormal distribution of tumor sizes (LN mean). By this measure KLT tumors had, on average, 7-fold more cancer cells than KT tumors (Fig. 3a,c)²⁶. Despite greater tumor burden and visibly larger tumors in KPT mice, p53-deficiency did not increase LN mean. Instead, p53-deficient tumors were power-law distributed at large sizes and the elevated tumor burden was driven by rare, exceptionally large tumors (Fig. 3d, and Supplementary Note)²⁷. A Power-law distribution is consistent with p53-deficiency allowing tumors to acquire additional rare, yet profoundly tumorigenic events that drive subsequent rapid growth^28–30.

Generation of a library of barcoded lentiviral vectors for multiplexed CRISPR/Cas9-mediated inactivation of tumor suppressor genes

To simultaneously quantify the tumor-suppressive function of many known and candidate tumor suppressor genes in parallel, we combined Tuba-seq and conventional Cre-based mouse models with multiplexed CRISPR/Cas9-mediated in vivo genome editing (Fig. 4a–c). Assessing different tumor genotypes within individual mice minimized the effect of mouse-to-mouse variability and maximizes the resolution of Tuba-seq (Supplemental Note).

Figure 4. Rapid quantification of tumor suppressor phenotypes using Tuba-seq and multiplexed CRISPR/Cas9 mediated gene inactivation.

a, Schematic of the Lenti-sgTS-Pool/Cre vector that contain a two-component barcode with an 8-nucleotide “sgID” sequence linked to each sgRNA as well as a random 15-nucleotide random barcode (BC).

b, Lenti-sgTS-Pool/Cre contains four vectors with inert sgRNAs and eleven vectors targeting known and candidate tumor suppressor genes. Each sgRNA vector contains a unique sgID and a random barcode. NT = Non-Targeting.

c, Schematic of multiplexed CRISPR/Cas9-mediated tumor suppressor inactivation coupled with Tuba-seq to assess the function of each targeted gene on lung tumor growth in vivo. Tumors were initiated with Lenti-sgTS-Pool/Cre virus in KT and KT;H11^LSL-Cas9 (KT;Cas9) mice.

d, Bright field (top) and fluorescence dissecting scope images (bottom) of lung lobes from KT and KT;Cas9 mice 12 weeks after tumor initiation with Lenti-sgTS-Pool/Cre. Lung lobes are outlined with white dashed lines in the fluorescence images. Viral titer is indicated. Scale bars = 5 mm.

Initiation of tumors with Lentiviral-sgRNA/Cre vectors targeting either the tdTomato reporter or Lkb1 tumors in mice with an H11^LSL-Cas9 allele confirmed efficient Cas9-mediated gene inactivation in lung tumors (Supplementary Fig. 5)⁸. Next, we selected eleven known and putative lung adenocarcinoma tumor suppressor genes representing diverse pathways and identified efficient sgRNAs targeting each gene (Fig. 4b and Supplementary Fig. 1a)^20,31. To quantify the number of cancer cells in each tumor using Tuba-seq, we diversified each Lenti-sgRNA/Cre vector with a two-component barcode consisting of a unique 8-nucleotide “sgID” specific to each sgRNA and a random 15-nucleotide barcode (BC) to uniquely tag each tumor (sgID-BC; Fig. 4a,b and Supplementary Fig. 6–7).

Parallel quantification of tumor suppressor function in vivo

To quantify the effect of inactivating each gene on lung tumor growth in parallel, we initiated tumors in KT and KT;H11^LSL-Cas9 (KT;Cas9) mice with a pool of the eleven barcoded Lenti-sgRNA/Cre vectors and four barcoded Lenti-sgInert/Cre vectors (Lenti-sgTS-Pool/Cre; Fig. 4b,c). Despite receiving a lower dose of virus than KT mice, KT;Cas9 mice had an increase in the number and size of macroscopic tumors 12 weeks after tumor initiation (Fig. 4d and Supplementary Fig. 8a). To determine the number of cancer cells in each tumor with each sgRNA, we amplified the sgID-BC region from bulk tumor-bearing lung DNA, deep sequenced the product, and applied our Tuba-seq analysis pipeline. For each sgRNA, the number of cancer cells in tumors at different percentiles was normalized to tumors from the sgInert distribution (Fig. 5a). We also determined the relative LN mean size of tumors containing each of the eleven tumor-suppressor-targeting sgRNAs (Fig. 5b). These analyses confirmed the known tumor-suppressive function of Lkb1, Rb1, Cdkn2a, and Apc in Kras^G12D-driven lung tumor growth (Fig. 5a,b and Supplementary Fig. 6b, Supplementary Fig. 8)^7,22,32,33.

Figure 5. Tuba-seq uncovers known and novel tumor suppressors with unprecedented resolution.

a, Analysis of the relative tumor sizes in KT;Cas9 mice 12 weeks after tumor initiation with Lenti-sgTS-Pool/Cre identified six tumor growth suppressing genes. Relative size of tumors at the indicated percentiles represents merged data from 8 mice, normalized to the average size of sgInert tumors. 95% confidence intervals are shown. Percentiles that are significantly greater than sgInert are in color.

b, Estimates of mean tumor size, assuming a lognormal tumor size distribution, identified sgRNAs that significantly increase growth in KT;Cas9 mice. Bonferroni-corrected, bootstrapped p-values are shown. p-values < 0.05 and their corresponding means are bold.

c, Relative size of the 95th percentile tumors (left), lognormal (LN) mean (middle), and lognormal (LN) p-value (right) for tumors with each sgRNA in KT and KT;Cas9 mice 12 weeks after tumor initiation, and KT;Cas9 mice 15 weeks after tumor initiation.

d,e, The relative size of the 95^th percentile tumor and the lognormal statistical significance determined by Tuba-seq identified plotted against the average fold change in ΔsgID representation and their associated p-values (e and f). Error bars in (e) are 95% confidence intervals. Dotted lines in (f) indicate the 0.05 significance threshold. Dot color corresponds to the sgRNA color in Figure 4b.

f, Tumor size at the indicated percentile from KT;Cas9 mice with Lenti-sgSetd2#1/Cre initiated tumors versus Lenti-sgNeo2/Cre initiated tumors (N=4 mice/group). Percentiles were calculated using all tumors from all mice in each group.

We analyzed an additional cohort of KT;Cas9 mice 15 weeks after tumor initiation with Lenti-sgTS-Pool/Cre. We confirmed the tumor-suppressive effect of all tumor suppressors identified at 12 weeks post-tumor initiation (Fig. 5c and Supplementary Fig. 8d–e). Importantly, both the LN mean and the relative number of cancer cells in the 95^th percentile tumor were reproducible (Fig. 5c and Supplementary Fig. 8).

Identification of p53-mediated tumor suppression and recapitulation of tumor size distributions within the tumor suppressor pool

Consistent with the distribution of tumor sizes in KPT mice, neither LN mean nor the analysis of tumors up to the 95^th percentile uncovered an effect of targeting p53 in KT;Cas9 mice with Lenti-sgTSPool/Cre-initiated tumors (Fig. 5). As anticipated, Lenti-sgp53/Cre-initiated tumors exhibited a power-law distribution at larger sizes and sgp53 was enriched within the largest tumors in these mice (Supplementary Fig. 9a,b). The effect of targeting p53 was greater at the later 15-week time point consistent with the known role of p53 in limiting tumor progression (Supplemental Fig. 9)^21,29.

In KT;Cas9 mice with Lenti-sgTSPool/Cre initiated tumors, Lenti-sgLkb1/Cre-initiated tumors exhibited a lognormal distribution of tumor sizes consistent with our data from KLT mice (Fig. 1c, 2d and Supplementary Fig. 10a). Thus, both p53- and Lkb1-deficient tumors generated through somatic genome-editing have similar size distributions to tumors initiated using floxed alleles. Even in this pooled setting, quantification of individual tumor sizes can uncover characteristic distributions of tumor sizes upon tumor suppressor inactivation.

Identification of Setd2 and Rbm10 as suppressors of lung tumor growth in vivo

In addition to appropriately uncovering tumor suppressors with known effects on lung tumor growth in vivo, Tuba-seq also identified the methyltransferase Setd2 and the splicing factor Rbm10 as suppressors of lung tumor growth. Setd2 is the sole histone H3K36me3 methyltransferase and may also affect genome stability by methylating microtubules^34–36. SETD2 is frequently mutated in several major cancer types, including lung adenocarcinoma ^{2,20,31,33,37}. Setd2 inactivation dramatically increased tumor size and these tumors exhibited a lognormal distribution of sizes (Supplementary Fig. 10) Splicing factors have emerged as potential tumor suppressors in many cancer types and components of the spliceosome are mutated in 10–15% of human lung adenocarcinomas ^2,20,31,38. Rbm10 inactivation significantly increased the number of cancer cells in the top 50% of lung tumors and increased the LN mean size (Fig. 5a,b). These data suggest that the absence of Setd2-mediated lysine methylation and aberrant pre-mRNA splicing each have profound pro-tumorigenic effects in lung adenocarcinoma.

Tuba-seq is a precise and sensitive method to quantify tumor suppression in vivo

By initiating many lesions per mouse, barcoding every lesion, pooling multiple sgRNAs into each mouse, and including inert sgRNAs with the pool we could identify and correct for multiple sources of biological and technical variation (Supplementary Note). Measuring the size of each tumor was considerably more precise and sensitive to the growth effect of tumor suppressor inactivation than bulk measurements (Figure 5). Interestingly, two thirds of our identified tumor suppressors (Apc, Rb1, Rbm10, and Cdkn2a) were only identified when we considered the number of neoplastic cells in each barcoded tumor, but not when we only considered the fold change in sgID representation (Fig. 5). In fact, the precision of effect size estimates, statistical significance, and ability to detect tumor suppressors with small effect were all improved using the Tuba-seq pipeline (Fig. 5d,e).

Confirmation of on-target CRISPR/Cas9-mediated genome editing

As an orthogonal approach to investigate the selection for tumor suppressor inactivation and to confirm on-target sgRNA-mediated genome editing, we PCR-amplified and deep-sequenced each sgRNA-targeted region from bulk lung DNA from Lenti-sgTS-Pool/Cre infected KT;Cas9 mice. A relatively high fraction of Setd2, Lkb1, and Rb1 alleles had inactivating indels at the targeted sites consistent with on-target sgRNA activity and the expansion of tumors with inactivation of these genes (Supplementary Fig. 9c–f and 11a,c).

This analysis also confirmed that all targeted genes contained indels (Fig. 6a). Although all of the genes included in our pool are recurrently mutated in human lung adenocarcinoma (Supplementary Fig. 1a)^20,31, Arid1a, Smad4, Keap1, and Atm were not identified as tumor suppressors (Fig. 5, Supplementary Fig. 8d–e,h and 11a). The lack of tumor-suppressive function of Atm is consistent with results using an Atm^floxed allele³⁹, and we confirmed the lack of tumor-suppressive function of Smad4 in vivo (Supplementary Fig. 11d,e). For these genes, changes in gene expression or environmental state, additional time, or coincident genomic alterations may be required for inactivation of these pathways to confer a growth advantage in lung cancer cells.

To further validate the tumor-suppressive effect of Setd2, we induced tumors in KT and KT;Cas9 mice with lentiviral vectors containing an inert sgRNA (sgNeo2) or either of two sgRNAs targeting Setd2. KT;Cas9 mice with tumors initiated with either Lenti-sgSetd2/Cre vector developed large adenomas and adenocarcinoma, and exhibited greater overall tumor burden than KT mice with tumors initiated with the same virus (Supplementary Fig. 12). Analysis of tumor sizes by Tuba-seq confirmed a nearly four-fold increase in the number of cancer cells in Setd2-deficient tumors relative to control tumors (Fig. 5f and Supplementary Fig. 12). Importantly, the validation of Setd2-mediated tumor suppression by conventional methods required more mice than our initial screen of eleven putative tumor suppressors emphasizing the benefit of multiplexing sgRNAs to increase throughput and decrease costs.

DISCUSSION

While many putative tumor suppressors have been identified from cancer genome sequencing, limited strategies exist to test their function in vivo in a rapid, systematic, and quantitative manner (Supplementary Table 1)^{1,2,4,7,20,31}. By combining DNA barcoding, high-throughput sequencing, and CRISPR/Cas9-mediated genome editing, Tuba-seq not only increases the throughput of these analyses, but also enables exceptionally precise and detailed quantification of tumor growth in vivo.

Interestingly, tumors initiated at the same time, within the same mouse, with the same genomic alterations grew to vastly different sizes after only 12 weeks of growth. Thus, additional spontaneous alterations, differences in the state of the initial transformed cell, or the local microenvironment may impact how rapidly a tumor grows and its capacity for continued expansion in these model systems. This growth variability identified by Tuba-seq, also revealed properties of gene function. p53-deficiency generates a tumor size distribution that is power-law distributed for the largest tumors, consistent with a Markov process where very large tumors are generated by additional, rarely acquired driver mutations (Supplementary Note)²⁷. Conversely, Lkb1 inactivation increases the size of a majority of lesions, consistent with the role of Lkb1 in restraining proliferation⁴⁰. Interestingly, Setd2 has recently been suggested to methylate tubulin, and Setd2-deficiency can lead to genomic instability which would be expected to generate power-law distributed tumor growth³⁴. However, the size distribution of Setd2-deficient lung tumors was strictly lognormal, suggesting that the main impact of Setd2 loss is the induction of gene expression programs that generally dysregulate growth (Fig. 5f and Supplementary Fig. 10b,c).

The scale of our analyses, which evaluated thousands of individual tumors, dramatically improved our ability to identify functional tumor suppressor genes (Fig. 5d–e). Unlike conventional floxed alleles, CRISPR/Cas9-mediated genome editing in the lung generates homozygous null alleles in approximately half of all tumors (Supplementary Fig. 5d). Thus, while the lack of uniform homozygous deletion of targeted genes would reduce the tumor suppressive signal from bulk measurements, by barcoding and analyzing each tumor, Tuba-seq effectively overcomes this technological limitation.

By analyzing a large number of tumor suppressors, our data suggest that early neoplastic cells reside in an evolutionarily nascent state where many tumor suppressor alterations were adaptive and conferred a large growth advantage. In contrast, tumor suppressor alterations in cancer cell lines often provide little advantage and can even be detrimental⁴¹. This is consistent with cancer cell lines residing in a much more mature evolutionary state, approaching optimal growth fitness due to their origin from advanced-stage disease as well as selection for optimal proliferative ability in culture. Furthermore, the intimate link between tumor suppression and many aspects of the in vivo environment underscores the importance of analyzing the effects of tumor suppressor loss in tumors in vivo^42–44.

Interestingly, the frequency of tumor suppressor alterations in human cancer did not directly correspond to the magnitude of their tumor suppressor function. While variation in mutation rates, inclusive fitness, and genetic context likely contribute to the frequency of mutations in human cancer, our findings highlights the need for rapid and quantitative methods to determine the functional importance of lower-frequency putative tumor suppressors that may be profoundly important for individual patients.

Tuba-seq will contribute to our understanding of cancer pathogenesis in many other ways. It should permit investigation of more complex combinations of tumor suppressor gene loss, and facilitate analysis of other aspects of progression. Tuba-seq is adaptable to studying other cancer types and genes that normally promote, rather than inhibit tumor growth^8,10,45,46. Finally, these methods may enable the investigation of genotype-specific therapeutic responses, ultimately leading to more precise and personalized patient treatment.

ONLINE METHODS

Step-by-step protocol

Protocols for plasmid barcoding and library preparation for Tuba-seq are available as Supplementary Protocols and on Protocol Exchange.

Mice and tumor initiation

Kras^LSL-G12D (K), Lkb1^flox (L), p53^flox (P), R26^LSL-Tomato (T),and H11^LSL-Cas9 (Cas9) mice have been described^8,19,47–49. Lung tumors were initiated by intratracheal infection of mice as previously described ¹⁸ using lentiviral-Cre vectors at the titers indicated. Tumor burden was assessed by fluorescence microscopy, lung weight, and histology as indicated. All experiments were performed in accordance with Stanford University Institutional Animal Care and Use Committee guidelines.

Generation of barcoded Lenti-mBC/Cre and Lenti-sgPool/Cre vector pools

To enable quantification of the number of cancer cells in individual tumors in parallel using high-throughput sequencing, we diversified lentiviral-Cre vectors with a short barcode sequence that would be unique to each tumor by virtue of stable integration of the lentiviral vector into the initial infected lung epithelial cell. We generated tumors in a variety of mouse backgrounds with two different pools of barcoded lentiviral vectors. The first was a pool of ~10⁶ uniquely barcoded variants of Lenti-PGK-Cre (Lenti-millionBC/Cre; Lenti-mBC/Cre, generated by pooling six barcoded Lenti-U6-sgRNA/PGK-Cre vectors) which we used to analyze the number of cancer cells in tumors induced in Kras^LSL-G12D/+;R26^LSL-Tomato (KT), Kras^LSL-G12D/+;p53^flox/flox;R26^LSL-Tomato (KPT), and Kras^LSL-G12D/+;Lkb1^flox/flox;R26^LSL-Tomato (KLT) mice (Figure 1). The second was a pool of 15 barcoded Lenti-U6-sgRNA/PGK-Cre vectors which we used to assess the tumor suppressive effect of candidate tumor suppressor genes in three different genetic backgrounds by infecting KT;H11^LSL-Cas9 (KT;Cas9) and KT mice. Our Lenti-sgInert/Cre vectors included three sgRNAs that target the NeoR gene within the Rosa26^LSL-Tomato allele, which are actively cutting, but functionally inert, negative control sgRNAs.

Design, generation, and screening of sgRNAs

We generated lentiviral vectors carrying Cre as well as an sgRNA targeting each of 11 known and putative lung adenocarcinoma tumor suppressors: sgLkb1, sgP53, sgApc, sgAtm, sgArid1a, sgCdkn2a, sgKeap1, sgRb1, sgRbm10, sgSetd2, and sgSmad4. Vectors were also generated carrying inert guides: sgNeo1, sgNeo2, sgNeo3, sgNT1, and sgNT3. All possible 20-bp sgRNAs (using an NGG PAM) targeting each tumor suppressor gene of interest were identified and scored for predicted on-target cutting efficiency using an available sgRNA design/scoring algorithm¹⁰. For each tumor suppressor gene, we selected three unique sgRNAs predicted to be the most likely to produce null alleles; preference was given to sgRNAs with the highest predicted cutting efficiencies, as well as those targeting exons conserved in all known splice isoforms (ENSEMBL), closest to splice acceptor/splice donor sites, positioned earliest in the gene coding region, occurring upstream of annotated functional domains (InterPro; UniProt), and occurring upstream of known human lung adenocarcinoma mutation sites ^{20,31,50–53}. Lenti-U6-sgRNA/Cre vectors containing each sgRNA were generated as previously described ⁸. Briefly, Q5 site-directed mutagenesis (NEB E0554S) was used to insert sgRNAs into the parental lentiviral vector containing the U6 promoter as well as PGK-Cre. The cutting efficiency of each sgRNA was determined by infecting LSL-YFP;Cas9 ⁸ cells with each Lenti-sgRNA/Cre virus. Forty-eight hours after infection, flow cytometric quantification of YFP-positive cells was used to determine percent infection. DNA was then extracted from all cells and the targeted tumor suppressor gene locus was amplified by PCR.

PCR amplicons were Sanger sequenced and analyzed using TIDE analysis to quantify percent indel formation⁵⁴. Finally, the indel percent determined by TIDE was divided by the percent infection of LSL-YFP;Cas9 cells, as determined by flow cytometry, to determine sgRNA cutting efficiency. The most efficient sgRNA targeting each tumor suppressor gene of interest was used for subsequent experiments. sgRNAs targeting Tomato and Lkb1 have been described previously^7,8, and we previously validated an sgRNA targeting p53 (unpublished data). Primers sequences used to amplify target indel regions for the top guides used in this study are below:

	F primer (5′ → 3′)	R primer (5′ → 3′)
sgApc_1	TGACTTTGCAGGGCAAGTTT	CCCACTCCCCTGTTACCTTT
sgArid1a_3	CAGCAGTCCCCAACTCCATA	GGAGCCATTTCTTGGGGTTA
sgAtm_3	GCCCCAAGTGAGAATCAGTG	AGCTCTGGCTCCTTGTGGAT
sgCdkn2a_2	GGCTTCTTTCTTGGGTCCTG	GGCTCATTTGGGTTGCTTCT
sgKeap1_2	CTGAGCCAGCAACTCTGTGA	GGCCTATCCCACTTCTGAGC
sgRb1_3	AACTGTGCTGGTGTGTGCAA	ACACCACCACCACCATCATC
sgRbm10_3	CAAAGCTGGAAGCGAGACTG	CTGGCTGGAGCTGTGAGAGT
sgSetd2_1	TCTGCAAGTTCAAGCGATGA	TGGATTCAGGTGACCTAGATGG
sgSetd2_2	CCTCCAGCCGCTCCTCAT	GAACGCCGAACCTAAGCAG
sgSmad4_3	GCCTTTCTGTGGAAATGGAA	TTCCAGGCTGAGTGGTAAGG
sgNeo_1	TTGTCAAGACCGACCTGTCC	CCACCATGATATTCGGCAAG
sgNeo_2	TCTGGACGAAGAGCATCAGG	GCTCCAATCCTTCCATTCAA
sgNeo_3	CGCTGTTCTCCTCTTCCTCA	TGGATACTTTCTCGGCAGGA

Barcode diversification of Lenti-sgRNA/Cre

After identifying the best sgRNA targeting each tumor suppressor of interest, we diversified the corresponding Lenti-sgRNA/Cre vector with a known 8-nucleotide ID specific to each individual sgRNA (sgID; green) and the 15-nucleotide random barcode (BC; purple) (see Supplementary Fig. 4a).

	Primer (5′ → 3′)
Universal Reverse Primer	AGCTAGGGATCCGCCGCATAACCAGTG
Barcoded Forward Primer	AGCTAGTCCGG AA TT AA ATGCCCAAGAAGAAGAGGAAGGTGTC

These primers were used to PCR amplify a region of the Lenti-PGK-Cre vector that included the 3′ end of the PGK promoter and the 5′ part of Cre. PCR was performed using PrimeSTAR^® HS DNA Polymerase (premix) (Clontech, R040A) and PCR products were purified using the Qiagen^® PCR Purification Kit (28106). The PCR insert was digested with BspEI and BamHI and ligated with the Lenti-sgRNA-Cre vectors cut with XmaI (which produces a BspEI compatible end) and BamHI.

To generate a large number of uniquely barcoded vectors, we ligated 300 ng of each XmaI, BamHI-digested Lenti-sgRNA-Cre vector with 180ng of each BspEI, BamHI-digested PCR product using T4 Ligase (NEB, M0202L) and standard protocols (80 μl total reaction volume). Ligations were PCR purified using the Qiagen^® PCR Purification Kit to remove residual salt. To obtain a pool of the greatest possible number of uniquely barcoded Lenti-sgRNA/Cre vectors, 1 μl of purified ligation was transformed into 20 μl of ElectroMAX DH10B cells (Thermo Fisher, 18290015). Cells were electroporated in 0.1 cm GenePulser/MicroPulser Cuvettes (Bio-Rad, 165-2089) in a BD MicroPulser™ Electroporator (Bio-Rad,165-2100) at 1.9kV. Cells were then rescued by adding 500 μl media and shaking at 200 rpm for 30 minutes at 37°C. For each ligation, bacteria were plated on seven LB-Amp plates (1 plate with 1 μl, 1 plate with 10 μl, and 5 plates with 100 μl). The following day, colonies were counted on the 1 μl or 10 μl plate to estimate the number of colonies on the 100 μl plates, and this was used as an initial estimation of number of unique barcodes associated with each ID.

10 ml of liquid LB-Amp was added to each plate of bacteria to pool the colonies. Colonies were scraped off of the plates into the liquid, and all plates from each transformation were combined into a flask. Flasks were shaken at 200 rpm for 30 minutes at 37°C to mix. DNA was Midi-prepped using the Qiagen^® HiSpeed MidiPrep Kit (12643). DNA concentrations were determined using a Qubit dsDNA HS Kit (Invitrogen, Q32851).

As a quality control measure, the sgID-BC region from each Lenti-sgRNA-sgID-BC/Cre plasmid pool was PCR amplified with GoTaq Green polymerase (Promega M7123) following manufacturer’s instructions. These PCR products were Sanger sequenced (Stanford PAN facility) to confirm the expected sgID and the presence of a random BC. Since BspEI and XmaI have compatible overhangs but different recognition sites, the Lenti-sgRNA-sgID-BC/Cre vectors generated from successful ligation of the sgID/BC lack an XmaI site. Thus for pools that had a detectable amount of unbarcoded parental Lenti-sgRNA/Cre plasmid as determined by Sanger sequencing (>5%), we destroyed the parental unbarcoded vector by digesting the pool with XmaI (NEB, 100μl reaction) using standard methods. These re-digested plasmid pools were re-purified using the Qiagen^® PCR Purification Kit and concentration was redetermined by NanoDrop.

Generation of Lenti-mBC/Cre and Lenti-TS-Pool/Cre

To obtain a library with approximately 10⁶ associated barcodes to use in our initial experiments in mice that lacked the H11^LSL-Cas9 allele, we pooled six sgID-BC barcoded vectors to create Lenti-million Barcode/Cre (Lenti-mBC/Cre). We then pooled the barcoded Lenti-sgRNA-sgID-BC/Cre vectors (sgLkb1, sgp53, sgApc, sgAtm, sgArid1a, sgCdkn2a, sgKeap1, sgNeo1, sgNeo2, sgNeo3, sgNT1, sgRb1, sgRbm10, sgSetd2, and sgSmad4) to generate Lenti-sgTS-Pool/Cre. All plasmids were pooled at equal ratios as determined by Qubit concentration prior to lentivirus production.

Production, purification, and titering of lentivirus

Lentiviral vectors were produced using polyethylenimine (PEI)-based transfection of 293T cells with the lentiviral vectors and delta8.2 and VSV-G packaging plasmids. Lenti-mBC/Cre, Lenti-sgTS-Pool/Cre, Lenti-sgTomato/Cre, Lenti-sgLkb1, Lenti-sgSetd2#1/Cre, Lenti-sgSetd2#3/Cre, Lenti-sgNeo2/Cre, and Lenti-sgSmad4/Cre were generated for tumor initiation. Sodium butyrate (Sigma Aldrich, B5887) was added at a final concentration of 0.2 mM eight hours after transfection to increase production of viral particles. Virus-containing media was collected 36, 48, and 60 hours after transfection, concentrated by ultracentrifugation (25,000 rpm for 1.5–2 hours), resuspended overnight in PBS, and frozen at −80°C. Concentrated lentiviral particles were titered by infecting LSL-YFP cells (a gift from Dr. Alejandro Sweet-Cordero), determining the percent YFP-positive cells by flow cytometry, and comparing the infectious titer to a lentiviral preparation of known titer.

Generation of “benchmark” cell lines

Three uniquely barcoded Lenti-Cre vectors with the sgID “TTCTGCCT” were used to generate benchmark cell lines that could be spiked into each bulk lung sample at a known cell number to enable the calculation of cancer cell number within each tumor. Plasmid DNA from individual bacterial colonies was isolated using the Qiagen^® QIAprep Spin Miniprep Kit (27106). Clones were Sanger sequenced, lentivirus was produced as described above, and LSL-YFP cells were infected at a very low multiplicity of infection such that approximately 3% of cells were YFP-positive after 48 hours. Infected cells were expanded and sorted using a BD Aria II™ (BD Biosciences). YFP-positive sorted cells were replated and expanded to obtain a large number of cells. After expansion, cells were re-analyzed for percent YFP-positive cells on a BD LSR II™ analyzer (BD Biosciences). Using this percentage, the number of total cells needed to contain 5 × 10⁵ integrated barcoded lentiviral vectors was calculated for each of the three cell lines and cells were aliquoted and frozen based on this calculation.

Summary of all mouse infections

Genotype	Virus Type	Viral Titer
KT	Lenti-mBC/Cre	6.8×105
KTlow	Lenti-mBC/Cre	1.7×105
KPT	Lenti-mBC/Cre	1.7×105
KLT	Lenti-mBC/Cre	1.7×104
KT	Lenti-TS-Pool/Cre	9.0×104
KT;Cas9	Lenti-TS-Pool/Cre	2.2×104
KT;Cas9	Lenti-sgNeo2/Cre	9×103
KT;Cas9	Lenti-sgSetd2#1/Cre	9×103
KT;Cas9	Lenti-sgSetd2#2/Cre	9×103
KT	Lenti-sgSmad4/Cre	105
KT;Cas9	Lenti-sgSmad4/Cre	105

Isolation of genomic DNA from mouse lungs

For experiments in which barcode sequencing was used to quantify the number of cancer cells in each tumor the whole lungs from each mouse were homogenized using a Fisher TissueMeiser. 5 × 10⁵ cells from each of the three individually barcoded benchmark cell lines were added at the time of homogenization. Tissue was homogenized in 20 ml lysis buffer (100mM NaCl, 20mM Tris, 10mM EDTA, 0.5% SDS) with 200 μl of 20 mg/ml Proteinase K (Life Technologies, AM2544). Homogenized tissue was incubated at 55°C overnight. To maintain accurate representation of all tumors, DNA was phenol-chloroform extracted and ethanol precipitated from ~ 1/10^th of the total lung lysate using standard protocols. For lungs weighing less than 0.3 grams, DNA was extracted from ~1/5^th of the total lung lysate, and for those weighing less than 0.2 grams, DNA was extracted from ~3/10^th of the total lung lysate to increase DNA yield.

Preparation of sgID-BC libraries for sequencing

Libraries were prepared by amplifying the sgID-BC region from 32 μg of genomic DNA per mouse. The sgID-BC region of the integrated Lenti-sgRNA-BC/Cre vectors was PCR amplified using one of 24 primer pairs that contain TruSeq Illumina^® adapters and a 5′ multiplexing tag (TruSeq i7 index region indicated in purple):

	Primer (5′ → 3′)
Universal Forward Primer	AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCTGCGCACGTCTGCCGCGCTG
Reverse Primer	CAAGCAGAAGACGGCATACGAGAT GTGACTGGACTTCAGACGTGTGCTCTTCCGATCCAGGTTCTTGCGAACCTCAT

We used a single-step PCR amplification of sgID-BC regions, which we found to be a highly reproducible and quantitative method to determine the number of cancer cells in each tumor. We performed eight 100 μl PCR reactions per mouse (4 μg DNA per reaction) using OneTaq 2× Master Mix with Standard buffer (NEB, M0482L) with the following PCR program:

1. 94C 10 min

2. 94C 30sec

3. 55C 30sec

4. 68C 30sec

5. GO TO 2 (34×)

6. 68C 7min

7. 4C infinity

Pooled PCR products were isolated by gel electrophoresis and gel extracted using the Qiagen^® MinElute Gel Extraction kit. The concentration of purified PCR products from individual mice was determined by Bioanalyzer (Agilent Technologies) and pooled at equal ratios. Samples were sequenced on an Illumina^® HiSeq to generate 100bp single-end reads (ELIM Biopharmaceuticals, Inc).

Identifying distinct sgRNAs and tumors via ultra-deep sequencing

The unique sgID-BC identifies tumors. These sgID-BCs were detected via next generation sequencing on the Illumina^® HiSeq. The size of each tumor, with respect to cell number, was expected to roughly correspond to the abundance of each unique sgID-BC pair. Because tumor sizes varied by factors larger than the rate of read sequencing errors, distinguishing true tumors from recurrent read errors required careful analysis of the deep-sequencing data.

To this end, tumors and their respective sgRNAs were identified in three steps: (i) abnormal and low quality reads were discarded from the ultra-deep sequencing runs, (ii) unique barcode pileups were bundled into groups that we predicted to arise from the same tumor, and (iii) cell number was estimated from these bundles in the manner that proved most reproducible.

Read pre-processing

Reads contained a two-component DNA barcode (an 8-nucleotide sgID and a 21-nucleotide barcode sequence that contains 15 random nucleotides) that began 49 nucleotides downstream of our forward primer. We discarded unusual reads, specifically: those that lacked the flanking lentiviral sequences, those that contained unexpected barcodes, and those with high error rates. This was accomplished in three steps (Supplementary Fig. 2a):

1. We examined the 12 lentiviral nucleotides immediately upstream and downstream of the sgID-BC. These 12 nucleotides were identified using pairs of adjacent 6-mer search strings, such that each 6-mer could tolerate one mismatch. Although we expected these 12 nucleotides to begin at position 37 within the read, we did not require this positioning or leverage this information. A nested 6-mer approach (with two opportunities to identify the lentiviral sequences flanking the sgID-BC) was used to minimize read discarding. For ~7–8% of reads, this 2^nd 6-mer match salvaged the read, i.e. the 6-mers immediately flanking the sgID-BC were not as expected (despite our tolerance of one mismatch) yet the 6-mers immediately outside of these inner 6-mer sequences were recognizable and allowed us to salvage the read and identify the barcodes. Salvaging reads is not particularly critical for estimating tumor sizes, however it is critical for accurate estimation of read error rates because the non-barcoded regions of our reads were used to estimate sequencing error rates and, therefore, should not be biased against read errors.

2. We then discarded reads in which the sgID-BC deviated in length by greater than two nucleotides in either direction. Because our first barcode was expected to contain one of the 15 sgIDs, we discarded reads that did not match one of these 15 sequences. One mismatch and one indel were permitted in the matching.

3. We then end-trimmed each read such that 18 bp flanked either end of the sgID-BC. We then filtered the trimmed reads according to quality score, retaining those that were predicted to contain no more than two sequencing errors⁵⁵. We also discarded reads with uncalled bases in the second (random) barcode and rectified uncalled bases elsewhere.

In these three stages, 14% of reads were discarded at stage one, ~7% at stage two, and <2% at stage three.

We then examined those reads that failed at each stage. By performing BLAST searches, we determined that those reads discarded at stage one often contained uninformative sequences corresponding to artifacts from either our preparation (Phi X bacteriophage genome and mouse genome) or other samples paired with us on the lane (common plasmid DNAs). In stage two, we found that reads with aberrant barcode lengths often contained large indels or had one or both of their sgID-BC completely missing. Lastly, very few reads were discarded in stage three due to the fact that internal regions of the reads exhibited higher quality scores than the termini of reads. As a consequence of this trend, it is common practice to end-trim reads prior to discarding those reads predicted to contain greater than two sequencing errors²⁵, as we did.

Clustering of unique read pileups via DADA2

sgID-BC reads were aggregated into sets of identical sequences and counted. The counts of unique DNA barcode pairs do not directly correspond to unique tumors because large tumors are expected to generate recurrent sequencing errors (Supplementary Fig. 2b). We therefore spent considerable effort developing a method to distinguish small tumors from recurrent sequencing errors arising from large tumors. Consider, for example, that a tumor of 10 million cells will produce sequencing-error pileups that mimic a 10–100 thousand-cell tumor, if the error rate is 0.1–1% (a typical rate, given the limitations of PCR amplification and Illumina^® sequencing machines). DADA2 has been used previously to address this issue in barcoding experiments involving ultra-deep sequencing¹². However, because it was designed for ultra-deep sequencing of full-length Illumina amplicons²⁵, we had to tailor and calibrate it for our purposes.

In DADA2, the likelihood of barcode pileups resulting from a recurrent sequencing error of a larger pileup depends upon:

1. The abundance of the larger pileup,

2. The specific nucleotide differences between the smaller and larger pileups, and

3. The average quality scores of the smaller pileup at the variant positions.

Factors one and two are, at first, considered heuristically (to maximize computational speed) and then more precisely (when needed) via a Needleman-Wunsch algorithm. DADA2 splits a cluster into two when the probability that a smaller pileup was generated by sequencing errors is less than Ω. Therefore, this value represents a threshold for splitting larger clusters. When this threshold is large, read pileups are split permissively (many called tumors, perhaps dividing large tumors), and when Ω is small, read pileups are split restrictively (few called tumors, perhaps aggregating distinct small tumors).

The likelihood of sequencing errors was inferred from our ultra-deep sequencing data. Phred quality scores provide a theoretical estimate of sequencing error rates, however these estimates tend to vary from Illumina^® machine to Illumina^® machine and do not account for the specifics of our protocol (including, for example, occasional errors introduced via PCR amplification, despite our use of high-fidelity polymerase). Ordinarily, DADA2 will estimate sequencing error rates simultaneously with the unique DNA clusters; however, our lentiviral constructs had non-degenerate regions outside of our sgID-BC region that were used to estimate sequencing error rates directly. Moreover, estimating error rates and barcode clusters jointly is more computationally intensive, requiring greater than 20,000 CPU-hours for clustering our entire dataset and exploring the relevant clustering parameters.

A sequencing error model was trained to each Illumina^® machine by:

1. Generating training pseudo-reads by concatenating the 18 nucleotides immediately upstream of our sgID-BC with the 18 nucleotides immediately downstream of the barcodes, then

2. Clustered these pseudo-reads using a single run of DADA2.

3. Using the error rates estimated from this training run to cluster the sgID-BC using a single run of DADA2.

We used a very low value of Ω = 10⁻¹⁰⁰ to estimate sequencing errors in the training run, as we expected only one cluster of lentiviral sgID-BC-flanking sequences. Altering this value does not affect training results appreciably, but we nonetheless observed occasional very small derivative clusters from our lentiviral sequence even at this value. These derivative clusters are presumably rare DNA artifacts and never amounted to >2% of our processed reads. We used a very stringent DADA2 run to estimate sequencing errors because a more permissive threshold might over-fit sequencing errors and underestimate sequencing error rates, while the less permissive approach of estimating error rates directly from each read’s deviance from expectation (akin to a DADA2 run where Ω = 0) would not accommodate any DNA artifacts in our data and, therefore, overestimate sequencing error rates.

We trained sequencing error rates on each Illumina^® machine used in this study (seven in total). Training allowed the probability of every substitution type (A→C, A→T, etc) to be estimated. The error rates as a function of Phred quality score were determined using LOESS regression of the available data (Supplementary Fig. 2c)²⁵. In general, error rates were approximately two to three times higher than predicted by the Phred quality scores for transversions (and approximately consistent with expectations for transitions). This elevated error rate is typical²⁵ and may reflect miscalibration of the machines and/or mutations introduced during PCR.

We then clustered the dual barcodes that passed our pre-processing filters using DADA2. Barcodes were given seven nucleotides of non-degenerate lentiviral flanking regions so that any indels within the barcodes could be identified (without adequate flanking sequences, DNA alignment algorithms sometimes miscall indels as multiple point mutations). During clustering, we also required (i) that clusters deviate from each other by at least two bases (i.e. MIN_HAMMING_DISTANCE = 2), (ii) that new clusters only be formed when pileup size exceeded expectations under the error process by at least a factor of two (MIN_FOLD = 2), and (iii) that the Needleman-Wunsch algorithm consider only alignments with at most four net insertions or deletions (BAND_SIZE = 4, VECTORIZED_ALIGNMENT = FALSE). None of these choices affected the results appreciably, but they increased computational performance and offered additional verification that barcodes were aggregated into tumors of reasonable size.

Vetting and calibration of pipeline

We sequenced our first PCR-amplified, multiplexed DNA libraries (from KT, KLT, and KPT tumors) in triplicate to vet and design our tumor-calling approach.

Reproducibility was measured in three ways: (i) by measuring correlation between estimated cell abundances for all barcodes and all mice, (ii) by measuring the variation in the number of lesions called for each sgID in each mouse in our first experiment, and (iii) by measuring the variation in LN mean size for each sgID—a value that should be constant in mice not expressing Cas9. Because the read depth of our triplicate run naturally varied (40.1 × 10⁶, 22.2 × 10⁶, and 34.9 × 10⁶ reads after pre-processing), these three runs were performed on distinct Illumina^® machines with different sequencing error rates, and, because our initial lentiviral pool contained six different sgIDs with varying levels of barcode diversity, the technical variability in our vetting process well-approximated the technical variability of later experiments. In our tumor-size analysis pipeline, we found:

1. The mean abundance of our three “benchmark” DNA barcodes was more reproducible between replicate runs than the median abundance. Thus, this mean value of benchmark read abundance (corresponding to 500,000 cells) was used to convert read abundance into the absolute cell number of cancer cells in each tumor (Supplementary Fig. 3).

2. Ignoring reads with ≥ 2 errors from the consensus barcode of a cluster improved reproducibility. Typically, ~80–90% of reads in a barcode cluster were exact matches to the consensus barcode, while ~5% of reads were single errors from this read, and ~5–15% of reads deviated at ≥ 2 errors. These reads, with ≥ 2 errors, were poorly correlated between replicate runs and hampered our ability to reproducibly estimate absolute cell number/tumor size. Thus, these reads were excluded as we have neither enough confidence to consider these reads as unique lesions, nor enough confidence to count these reads towards the larger cluster.

3. The cluster-splitting proclivity of DADA2 was thresholded at Ω = 10⁻¹⁰ and we required that lesions contain ≥500 cells for Figures 1–3 and ≥1000 cells for Figures 4–6 to maximize reproducibility between replicate runs (Supplementary fig. 2d–f). Threshold parameters with high specificity (small Ω, high minimum cell number) called lesion sizes more reproducibly, whereas threshold parameters with high sensitivity (large Ω, low minimum cell number) called lesion quantities more reproducibly. Over-prioritizing only one facet of reproducibility would be imprudent. With two thresholds, considering different facets of measurement error, we better balanced these competing priorities.

With this pipeline, we interrogated the diversity of the barcode in our screen in several ways. First, we confirmed that nucleotides in this barcode were evenly distributed among A’s, T’s, C’s, and G’s (Supplementary Fig. 4b). Second, we found no evidence for an excess of repeated string (e.g. sequences AAAAA). Third, we calculated the number of random barcodes paired to each sgID in our lentiviral pool. Due to the large number of uniquely barcoded variants of each vector that we generated through our barcode ligation approach, (see Barcode diversification of Lenti-sgRNA/Cre) most barcodes that exist in our lentiviral pool were never detected in any lesions in any of the experiments (because diversity is much higher that total lesion number). Nonetheless, we still inferred the amount of barcode diversity from the observed barcodes.

To infer the barcode diversity of each sgID, we assumed that the probability of observing a barcode in i mice is Poisson distributed: P(k=i; λ) = λ^k e^−λ / k!, where λ_r = L_r/D_r, is a ratio of the number of called lesions L_r for each sgID r in our entire dataset (a known quantity) divided by the total number of unique barcodes D_r for each sgID (our quantity of interest). By noting that λ_r/(1 – e^−λr) = μ_non-zero, where μ_non-zero = Σ_i=1^∞P(k=i; λ_r) is simply the mean number of occurrences of each barcode that occurred once or more, we calculated D_r. Across our entire dataset, the average probability of the same barcode initiating two distinct tumors in the same mouse was 0.91%.

Good barcode diversity is also demonstrated by the highly-reproducible mean size of the six sgIDs in the Lenti-mBC/Cre experiment. If barcode diversity was low and barcodes overlapped often within a mouse, then the mean sizes of the less diverse sgIDs would increase—as two distinct tumors with the same barcode would be bundled together. However, the mean sizes of each sgID vary by <1% within replicate mice, thus refuting this possibility. We also assessed our ability to call sgIDs accurately, despite sequencing errors, by processing deep-sequencing runs in two ways: by identifying each read’s cognate sgID before clustering based on the raw read sequence or by identifying cognate sgIDs after clustering based on the consensus sequence of the cluster. Using either approach, 99.8% of reads paired to the same cognate sgID, thus providing assurance that sgIDs are accurately identified. We opted to employ the latter approach for our final analysis.

By thoroughly developing and vetting our tumor-calling pipeline, we salvaged an extra decade of size resolution. Our three DNA benchmarks (added to the lung samples at the very beginning of DNA preparation) (Supplementary Fig. 3) offer a glimpse of this resolution. Sequencing errors of the DNA benchmarks are easily identified by the DNA benchmark’s unique sgID and known secondary barcodes. While these sequencing errors are usually discarded, we can treat them as ordinary read pileups and observe the properties of potential sequencing errors. Without our calibrated analysis pipeline, the sequencing errors appear as lesions of ~10³ cells; with our pipeline, these sequencing errors emerge as lesions of ~10² cells—below our minimum cell threshold (Fig. 2a).

More importantly, our pipeline is robust to technical perturbations. We more intensively profiled reproducibility with two additional technical perturbations in two specific mice from the first experiment. First, a KLT 11-week mouse (JB1349) was sequenced at great depth and then randomly down-sampled ten-fold to typical read depth (this down-sampling was more dramatic than any variability in read depth actually detected throughout our study). Lesion sizes were very highly correlated in this first perturbation (Fig. 2b). Additionally, a KT 11-week mouse (IW1301) was amplified in two PCR reactions with different multiplexing tags (Fig. 2c). PCR and multiplexing appears to hamper reproducibility more than read depth, although reproducibility is good overall. These mice also display two encouraging reproducibility trends: (i) larger lesions/tumors were most consistent between replicates, and (ii) the overall shape (histogram) of tumor lesion sizes were better correlated between the replicates than individual tumors (e.g. r = 0.89 after log-transformation for each lesion in IW1301, whereas r = 0.993 for the abundance of tumors within the 60 histogram bins of Supplementary Fig. 2b). The excellent reproducibility of size histograms suggests that noise in our tumor size calls is generally unbiased.

Minimizing the influence of GC amplification bias on tumor-size calling

We define each tumor in our study by a size T_mrb corresponding to the mouse m that harbored it, the cognate sgRNA r identified by its first barcode, and a unique barcode sequence (consensus of the DADA2 cluster) b. Given the approximately lognormal structure of our data (Fig. 3d and data not shown), we log-transformed and normalized sizes such that τ_mrb = Ln(T_mrb/E_mr[T_mrb]). Here E_mr[T_mrb] = Σ_b T_mrb /N_mr is the expected lesion size for a given mouse m and sgRNA r and we will use this notation for expectation values. This notation—where aggregated indices are dropped from subscripts—is used throughout. GC biases were subtle: the coefficient of variation (CV) of E_mr[T_mrb] was 5.0%. This marginal distribution still exhibited a subtle dependence on the GC-content of the combined barcode sequence that was best described by a 4^th-order least-squares polynomial fit f₄ (b) of E_b[τ_mrb] (adjusted r² = 0.994). The sgIDs were all designed with well-balanced GC-content, however the second barcode comprises random sequences. While the multinomial process of generating barcodes made intermediate levels of GC-content most common, some deviation of GC-content was observed. Maximal values of f₄ (b) arise at intermediate GC-content, suggesting that PCR biases amplification towards template DNA of intermediate melting temperature. We subtracted the effects of this GC-bias from log-transformed values: t_mrb = Ln[T_mrb] – f₄ (b). This correction alters tumor sizes by 5% on average.

Calculation of in vitro cutting efficiency using the Lenti-TS-Pool/Cre virus

Cas9 expressing cell lines were infected with Lenti-TS-Pool/Cre virus and harvested after 48 hours. gDNA was extracted and targeted loci were amplified using the above primers.

Analysis of indels at target sites

To confirm CRISPR/Cas9-induced indel formation in vivo, the targeted region of each gene of interest was PCR-amplified from genomic DNA extracted from bulk lung samples using GoTaq Green polymerase (Promega M7123) and primer pairs that yield short amplicons amenable to paired-end sequencing:

	F primer (5′ → 3′)	R primer (5′ → 3′)
Apc	CATGGCATAAAGCAGTTACTACA	TCTCCTGAACGGCTGGATAC
Arid1a	CCAGTCCAATGGATCAGATG	GGGTACCCATGTCCTTGTTG
Atm	CACCCAGTTGACCCTATCTTC	CCGTTTTCGGAAGTTGACAG
Cdkn2a	CAACGTTCACGTAGCAGCTC	ACCAGCGTGTCCAGGAAG
Keap1	GGCTTATTGAGTTCGCCTACA	GCTGCTGCACGAGGAAGT
Rb1	GGTACCCGATCATGTCAGAGA	AAGGAACACAGCTCCCACAC
Rbm10	TACTCAGCCGCTTTCTTTGC	GAGGATTTGTTCCGCATCAG
Setd2	CTGTTGTGGTTGTGCCAAAG	TTTTCAGTTTGAGAACAGCCTTT
Smad4	TCGATTCAAACCATCCAACA	CTTGTGGAAGCCACAGGAAT
Lkb1	GGGCCTGTACCCATTTGAG	TGTCCCTTGCTGTCCTAACA
p53	CATCACCTCACTGCATGGAC	CAGGGGTCTCGGTGACAG
Neo1	GGCAGGATCTCCTGTCATCT	AGTACGTGCTCGCTCGATG
Neo2	CGGACCGCTATCAGGACATA	GAGCGGCGATACCGTAAAG
Neo3	GATCGGCCATTGAACAAGAT	CATCAGAGCAGCCGATTGT

PCR products were either gel-extracted or purified directly using the Qiagen^® MinElute kit. DNA concentration was determined using the Qubit HS assay, following manufacturer’s instructions. All 14 purified PCR products were combined in equal proportions for each mouse. TruSeq Illumina^® sequencing adapters were ligated on to the pooled PCR products with a single multiplexing tag per mouse using SPRIworks (Beckman Coulter, A88267) with standard protocols. Sequencing was performed on the Illumina HiSeq to generate single-end, 150-bp reads (Stanford Functional Genomics Facility).

Custom Python scripts were used to analyze the indel sequencing data. For each of the 14 targeted regions, an 8-mer was selected on either side of the targeted region to generate a 46 basepair region. Reads were required to contain both anchors and no sequencing errors were allowed. The length of each fragment between the two anchors was then determined and compared to the expected length. Indels were categorized according to the number of basepairs inserted or deleted.

The percent of indels for each individual locus in each individual mouse was calculated as follows:

$$\%\mathit{Indels}=\frac{\mathit{Total}\mathit{Reads}-\mathit{WildType}\mathit{Reads}}{\mathit{Total}\mathit{Reads}}$$

Then the average % of indels in the three Neo loci was calculated and the % indels at every other targeted locus was normalized to this value to generate the % Indels relative to Neo that are plotted in figure 6a.

Calculation of in vitro cutting efficiency using the Lenti-TS-Pool/Cre virus

Cas9 expressing cell lines were infected with Lenti-TS-Pool/Cre virus and harvested after 48 hours. gDNA was extracted and targeted loci were amplified using the above primers (see Analysis of indels at target sites). First, all primers were pooled and 15 rounds of PCR were performed using GoTaq Green polymerase (Promega M7123). These products were then used for subsequent amplification with individual primer pairs as described above. Sequencing libraries were prepared as described above.

Histology, immunohistochemistry, and tumor analysis

Samples were fixed in 4% formalin and paraffin-embedded. Immunohistochemistry was performed on 4 μm sections with the ABC Vectastain kits (Vector Laboratories) with antibodies against Tomato (Rockland Immunochemicals, 600-401-379), Smad4 (AbCam, AB40759) and Sox9 (EMD Milipore, AB5535). Sections were developed with DAB and counterstained with haematoxylin. Haematoxylin and eosin staining was performed using standard methods.

Sections from lungs infected with Lenti-sgTomato/Cre, were stained for Tomato and tumors were scored as positive (>95% Tomato positive cancer cells), Negative (no Tomato-positive cancer cells), or mixed (all other tumors). Tumors were classified and counted from a single section through all lung lobes from 4 independent mice.

Quantification of tumor area and barcode sequencing of tumors induced with Lenti-sgSetd2 and Lenti-sgNeo

Tumor-bearing lung lobes from mice infected with Lenti-sgSetd2#1/Cre, Lenti-sgSetd2#2/Cre or Lenti-sgNeo2/Cre virus were embedded in paraffin, sectioned, and stained with haematoxylin and eosin. Percent tumor area was determined using ImageJ.

The distribution of the number of cancer cells in individual tumors in KT;Cas9 mice infected with Lenti-sgSetd2#1/Cre and Lenti-sgNeo2/Cre was assessed by Illumina^® sequencing of their respective lentiviral barcodes and subsequent analysis as described above.

Western blotting for Lkb1 and Cas9

Microdissected Tomato-positive lung tumors from KT and KT;Cas9 mice with Lenti-sgLkb1/Cre initiated tumors were analyzed for Cas9 and Lkb1 protein expression. Samples were lysed in RIPA buffer and boiled with LDS loading dye. Denatured samples were run on a 4%–12% Bis-Tris gel (NuPage) and transferred onto a PVDF membrane. Membranes were immunoblotted using primary antibodies against Hsp90 (BD Transduction Laboratories, 610419), Lkb1 (Cell Signaling, 13031P), Cas9 (Novus Biologicals, NBP2-36440), and secondary HRP-conjugated anti-mouse (Santa Cruz Biotechnology, sc-2005) and anti-rabbit (Santa Cruz Biotechnology, sc-2004) antibodies.

Survival analysis of mice with Cas9 mediated inactivation of Smad4

To confirm lack of functional tumor suppression attributable to Smad4, KT and KT;Cas9 mice were infected intratracheally with 10⁵ Lenti-sgSmad4/Cre. Mice were sacrificed when they displayed visible signs of distress to assess survival.

Protocols and Vectors

Protocols for generation of barcoded vectors and library preparation for Tuba-seq analysis have been uploaded to protocol exchange and the following unbarcoded Lenti-pLL3.3-sgRNA/Cre vectors are available via AddGene:

Vector Name	AddGene ID
Lenti-sgNT1/Cre	66895
Lenti-sgNT3/Cre	89654
Lenti-sgNeo1/Cre	67594
Lenti-sgNeo2/Cre	89652
Lenti-sgNeo3/Cre	89653
Lenti-sgSmad4/Cre	89651
Lenti-sgSetd2#1/Cre	89649
Lenti-sgSetd2#2/Cre	89650
Lenti-sgRbm10/Cre	89648
Lenti-sgRb1/Cre	89647
Lenti-sgp53/Cre	89646
Lenti-sgKeap1/Cre	89645
Lenti-sgCdkn2a/Cre	89644
Lenti-sgAtm/Cre	89643
Lenti-sgArid1a/Cre	89642
Lenti-sgApc/Cre	89641
Lenti-sgLkb1/Cre	66894

Step-by-step protocol

Protocols for plasmid barcoding and library preparation for Tuba-seq are available as Supplementary Protocols and on Protocol Exchange.

Data Availability

Raw sequencing data is available upon request.

Code Availability

User-friendly code has been made available at https://github.com/petrov-lab/tuba-seq.