Developmental priming of cancer susceptibility

SUMMARY

DNA mutations are necessary drivers of cancer, yet only a small subset of mutated cells go on to cause the disease. To date, the mechanisms that determine which rare subset of cells transform and initiate tumorigenesis remain unclear. Here, we take advantage of a unique model of intrinsic developmental heterogeneity (Trim28^+/D9) and demonstrate that stochastic early life epigenetic variation can trigger distinct cancer-susceptibility ‘states’ in adulthood. We show that these developmentally primed states are characterized by differential methylation patterns at typically silenced heterochromatin, and that these epigenetic signatures are detectable as early as 10 days of age. The differentially methylated loci are enriched for genes with known oncogenic potential. These same genes are frequently mutated in human cancers, and their dysregulation correlates with poor prognosis. These results provide proof-of-concept that intrinsic developmental heterogeneity can prime individual, life-long cancer risk.

INTRODUCTION

Cancer is triggered by oncogenic DNA mutations^1–3. However, these mutations are also found at relatively high rates in otherwise ‘normal’ tissues, and not every mutation is oncogenic across all tissues^4–13. In other words, the oncogenic potential of DNA mutations are cell-, tissue-, and temporal- specific^14,15. The molecular basis of this context specificity comprises one of the biggest unanswered questions in cancer biology.

Pioneering studies over the last decades have implicated epigenetic regulation as a key mediator of context specificity. Notable examples include demonstrations that cell-type and differentiation-stage specific differences in epigenetic control determine when and where transformation occurs^14,16–23. What is typically overlooked in human genetics and epidemiology is that, in addition to the epigenetic programs that emerge and drive cell differentiation, another layer of intrinsic epigenetic variation arises during development that is at least partially stochastic in nature²⁴. Indeed, these epigenetic changes occur at rates several orders of magnitude higher than mutations²⁵, and an unequal distribution of epigenetic marks can drive phenotypic discordance for instance between MZ twins or isogenic mice^25–27. This ‘intrinsic developmental heterogeneity’ is distinct from the epigenetic changes triggered by external environmental exposures and from the large literature of in utero and early-life environmental insults that can increase cancer risk (e.g., estrogens, pesticides, alcohol, and over- or under-nutrition^28–33). While an impressive theoretic framework has been developed for how intrinsic developmental heterogeneity impacts cancer³⁴, to our knowledge, the notion has never been demonstrated experimentally. This has been due in part to a lack of proper models.

Because tumor initiation involves some degree of randomness, testing the relationship between developmental heterogeneity and cancer susceptibility requires measurement of the distribution of observed outcomes comparing distinct intrinsic epigenetic states; essentially, it requires an isogenic model, raised in tightly controlled environments, but bearing more than one reproducible intrinsic epigenetic states^35,36. TRIM28 is an epigenetic regulator that plays an important role in heterochromatic gene silencing^37–41.TRIM28 loss-of-function models have implicated the protein in cancer in complex and tissue-specific manner⁴². TRIM28, however, is also a master regulator of organism-level developmental heterogeneity⁴³. Our prior work showed that genetically and environmentally identical Trim28^+/D9 haplo-insufficient mice emerge into adulthood as two distinct populations (or developmental morphs) characterized by differences in body mass composition⁴⁴. The Trim28^+/D9 mouse thus meets the requirements for testing intrinsic developmental heterogeneity effects; it provides a model sensitized to detect the long-term phenotypic consequences of two distinct developmental programming states. Here, we leverage this unique model and ask if isogenic populations with reproducibly distinct intrinsic epigenetic states might exhibit differential cancer susceptibility. We show that the two Trim28^+/D9 developmental morphs develop distinct types, timing and severity of cancer. We identify a signature of DNA hypo-methylated genes, installed well before weaning, that stratify mice for cancer risk and outcome. These same genes are frequently mutated in human cancers and their dysregulation correlates with poor prognosis, suggesting that if conserved, this novel mode of action has the potential to impact a broad portion of the population.

RESULTS

Trim28^+/D9 mice exhibit a novel multi-cancer syndrome.

To test if developmental heterogeneity influences cancer, we crossed B6J.Tp53^+/R270H mice with FVB.Trim28^+/D9 animals (Fig. 1A). The Tp53^+/R270H mouse is a multi-cancer syndrome (MCS) model⁴⁵, while the Trim28^+/D9 mouse is sensitized to exhibit reproducible and exaggerated developmental heterogeneity^43,44,46,47. Both lines were highly backcrossed and cohorts thus yielded isogenic offspring with one of four genotypes: wild-type (WT), Trim28^+/D9 single heterozygotes (Trim28), Tp53^R270H/+ single heterozygotes (Tp53), and Tp53^R270H/+;Trim28^+/D9 compound heterozygotes (Tp53/Trim28). Parental ID, litter-size, and housing density were all carefully recorded to maximize our ability to associate cancer outcomes with developmental heterogeneity while reducing confounders. The experiment tracked animals from birth to endpoint (70 weeks of age), monitoring every individual for signs of cancer 2–3 times per week, with periodic measures of morphological, growth, and metabolic characteristics (Fig. 1A). Early-life ear biopsies were obtained at 10 days of age for epigenomic profiling (Fig. 1A). At sacrifice, all animals underwent a systematic 21-organ dissection protocol in which tissues were isolated, processed for histology, and scored by a pathologist. Cancer events were divided into aggressive (i.e., animals requiring euthanasia prior to 70 weeks) and endpoint events (i.e., animals reaching the 70-weeks endpoint without evidence of sickness). The final dataset comprised 137 animals with 79 malignant and 34 benign primary tumors.

Figure 1. Trim28^+/D9 mice exhibit a novel multi-cancer syndrome.

A) Schematic of the experimental plan. We mated the Tp53^+/R270H multi-cancer syndrome model (MCS) with the Trim28^+/D9 developmental heterogeneity model. All the resulting genotypes were screened for health issues and tumor development. Tissues and masses were harvested either before the endpoint (aggressive) or at the endpoint of the study (70 weeks). Histopathological evaluation was performed on the harvested tissues and masses to determine the presence of tumors. Body, fat, and lean mass composition data were collected at multiple timepoints. Early life biopsies were collected at day 10 of age (pre-weaning). Created with BioRender.com. B) Kaplan-Meier survival probability for each genotype. Log-rank test, p=0.046. N=60 (6 WT, 17 Tp53^R270H/+, 15 Trim28^+/D9, 22 Tp53^R270H/+;Trim28^+/D9). C) Percentage of animals found with aggressive tumors for each genotype (relative to the total number of animals screened for each genotype). N=60 (6 WT, 15 Trim28^+/D9, 17 Tp53^R270H/+, 22 Tp53^R270H/+;Trim28^+/D9). D) Top panel: prevalence of malignant (black bars) or benign (white bars) aggressive tumors of for each genotype, expressed as percentage relative to the total number of tumors found in each genotype. N=76 (total tumors: 1 in WT, 16 in Trim28^+/D9, 29 in Tp53^R270H/+, 30 in Tp53^R270H/+;Trim28^+/D9). Bottom panel: prevalence of distinct malignant aggressive tumor types for each genotype, expressed as percentage relative to the total number of malignant aggressive tumors found in each genotype. N=52 (total malignant aggressive tumors: 1 in WT, 8 in Trim28^+/D9, 18 in Tp53^R270H/+, 25 in Tp53^R270H/+;Trim28^+/D9). E) Fraction of animals harboring 1 or multiple aggressive malignant tumors prior to endpoint in the different genotypes. N=36 animals sacrificed prior to endpoint (1 WT, 7 Trim28^+/D9, 12 Tp53^R270H/+, 16 Tp53^R270H/+;Trim28^+/D9). F) Tissues targeted by malignant aggressive tumors in the different genotypes. Top panel: mouse anatomy plots, with non-targeted in light-grey and targeted tissues colored according to the affected genotype: WT in black, Trim28^+/D9 in orange, Tp53^R270H/+ in green, Tp53^R270H/+;Trim28^+/D9 in purple. Bottom panel: percentage of animals with specific organs targeted by malignant aggressive tumors in the different genotypes. N=36 animals sacrificed prior to endpoint (1 WT, 7 Trim28^+/D9, 12 Tp5^R270H/+, 16 Tp53^R270H/+;Trim28^+/D9). Created with BioRender.com

As expected^45,48,49, Tp53^R270H/+ siblings exhibited a high penetrance multi-cancer syndrome (MCS), with 76% of Tp53^R270H/+ mice succumbing to aggressive tumors before the 70-week endpoint (Fig. 1B-C). We found 24 total primary malignant tumors in Tp53^R270H/+ animals, primarily consisting of carcinomas and sarcomas. Eighteen of these developed before the 70-week study endpoint (Fig. 1C) and were widely distributed throughout the body (Fig. 1D-F and S1E). We were surprised to find that Trim28^+/D9 heterozygotes also showed a substantially reduced survival probability, similar to that of Tp53^R270H/+ animals (Fig. 1B; mean survival probability of 59.9 and 56.5 weeks, respectively). Health-monitoring and the 21-organ histopathology analysis revealed that reduced Trim28^+/D9 survival was due to its own MCS. Trim28^+/D9 animals showed similar time-to-detection and tumor burden to Tp53^R270H/+ animals (Fig. 1B, 1D, top, and -1E). That said, Trim28^+/D9-triggered MCS showed several differences relative to Tp53^R270H/+-triggered MCS. First, Trim28^+/D9 animals specifically developed rare germ-cell tumors (Fig.S1B-C). Second, Trim28^+/D9 animals only showed one case of sarcoma, whereas sarcomas were common in Tp53^R270H/+ animals (Fig.1D, bottom, and S1C). Overall, carcinomas dominated the landscape of malignant primary tumors across genotypes, with representative histological examples shown in Appendix 1. Tp53^R270H/+;Trim28^+/D9 compound heterozygotes showed the lowest survival probability of all genotypes (Fig. 1B), with target tissue and tumor type distributions consistent with the presence of both alleles (Fig. 1F and 1D). Stratifying the data into aggressive and endpoint samples indicated that the early and late pathologies within each genotype were largely constant (Fig. 1A-E and S1A-E). As expected, age-associated carcinomas were over-represented in WT animals (Fig.S1C and S1E, right). The few sarcomas that were observed in Trim28^+/D9 animals were found at endpoint (Fig.S1C). Collectively, these data demonstrate that Trim28^+/D9 triggers a novel MCS, similar in timing and penetrance to Tp53^R270H/+.

TRIM28-dependent developmental heterogeneity primes cancer outcomes.

Consistent with the literature^43,44,46,47, Trim28^+/D9 animals showed marked variation in body mass at 16 weeks of age (Fig. 2A) and separated statistically (MClust) into two distinct developmental ‘morphs’ (reproducible phenotypic forms), referred to here as Trim28^+/D9-heavy and Trim28^+/D9-light (Fig. 2B-C, S2A-). Tp53^R270H/+;Trim28^+/D9 compound heterozygotes also showed high variation in body mass, indicating that the developmental heterogeneity effect of Trim28^+/D9 is maintained in the presence Tp53^R270H/+. In compound heterozygotes, however, bimodality could not be statistically resolved (Fig. 2A-C, and Fig.S2A-B). Neither WT nor Tp53^R270H/+ heterozygotes showed significant variation phenotypes (Fig. 2A-B), nor of a bimodal distribution (Fig. 2A-C, and Fig.S2A-B). We validated the unbiased developmental morph calling using Rmixmod and found 100% congruence with MClust (cluster detection and morph assignment) (Fig.S2C). The phenotypic distinctions between morphs are transient over the long-term (Fig.S2A). They were also consistent with our previous findings⁴⁴. This developmental bifurcation in Trim28^+/D9 animals is a critical feature of the model, because it enables comparison of cancer outcomes between groups of isogenic animals with distinct yet reproducible developmental trajectories (Fig. 2B and Fig.S2A).

Figure 2. TRIM28-dependent developmental heterogeneity primes cancer outcomes.

A) Distribution of body mass at 16 weeks of age for each genotype. Each dot represents one animal. Levene’s test, Benjamini–Hochberg adjusted p-value for multiple comparisons, threshold of significance for padj<0.05: * p<0.05; ** p<0.01. WT vs Trim28^+/D9, padj=0.0016; WT vs Tp53^R270H/+, padj=0.2174; WT vs Tp53^R270H/+;Trim28^+/D9, padj=0.0152; Trim28^+/D9 vs Tp53^R270H/+, padj=0.0016; Trim28^+/D9 vs Tp53^R270H/+;Trim28^+/D9, padj=0.4220; Tp53^R270H/+;Trim28^+/D9 vs Tp53^R270H/+, padj=0.0178. N=129 (16 WT, 33 Trim28^+/D9, 42 Tp53^R270H/+, 38 Tp53^R270H/+;Trim28^+/D9). B) Fat and lean mass data at 16 weeks for the different genotypes. Each dot represents one animal. Overlaid density estimation of the data by MClust for each genotype. The dot size represents the classification certainty (probability) as calculated by MClust for each animal. N=137 (18 WT, 34 Trim28^+/D9, 44 Tp53^R270H/+, 41 Tp53^R270H/+;Trim28^+/D9). C) Bimodality index determined for each genotype as the ratio between the MClust-determined Bayesian Information Criterion (BIC) for 2 clusters over 1 cluster from fat and lean mass data at 16 weeks (same data as in Fig. 2A). D) Proportion of Trim28^+/D9-heavy and -light animals affected by malignant aggressive tumors (expressed as percentage of animals developing tumors relative to the total animals in each population). Two-sample test for equality of proportions without continuity correction, significance for p<0.05: -heavy vs -light, p=0.0308, χ²=4.6667: * p<0.05. N=14 (7 -heavy, 7 -light). E) Kaplan-Meier survival probability of Trim28^+/D9-heavy and -light animals. Log-rank test, p=0.049: * p<0.05. N=14 (7 -heavy, 7 -light). F) Prevalence of distinct aggressive tumor types in Trim28^+/D9-heavy and -light animals (expressed as percentage relative to the total number of aggressive tumors found in each population). N=21 (total aggressive tumors, including malignant and benign, 2 in -heavy and 19 in -light). G) Distribution of malignant aggressive tumor types in Trim28^+/D9-heavy and -light animals, expressed as percentage of animals with a particular tumor type relative to the total number of animals with malignant aggressive tumors in each population. The colored bars on the right identify the main tumor type, as in Fig. 2F: age-related carcinoma in red; carcinoma in gold; leukemia in green. N=6 (1 -heavy, 5 -light). H) Fraction of Trim28^+/D9-heavy and -light animals that died before the endpoint of the study harboring 1 or multiple malignant aggressive tumors. N=6 (1 -heavy, 5 -light). I) Tissues targeted by malignant aggressive tumors in the different genotypes. Top panel: mouse anatomy plots, with non-targeted in pale grey and targeted tissues colored according to the affected morph: dark orange for -heavy, pale orange for -light. Bottom panel: percentage of animals with specific organs targeted by malignant aggressive tumors in the different morphs. N=6 (1 -heavy, 5 -light). J) Key histological examples of aggressive malignant tumors and targeted tissues for -heavy and -light animals. I- AML targeting the bone marrow in a Trim28^+/D9-heavy animal (same as Appendix 1B-IX). II- Prostatic ACA in a Trim28^+/D9-light animal. III- Seminal vesicles ACA Trim28^+/D9-light animal. IV- Bronchoalveolar CA in a Trim28^+/D9-light animal. V- Gastroesophageal junction SCC in a Trim28^+/D9-light animal. VI- Colon ACA in a Trim28^+/D9-light animal.

Because Trim28^+/D9 animals exhibited both a reproducible developmental bifurcation and a novel MCS, we focused on the Trim28^+/D9 genotype to test for associations between intrinsic developmental heterogeneity and cancer outcomes. Using the unbiased MClust assignment of Trim28^+/D9 animals to their -light or -heavy morph, we compared survival, tumor incidence and associated outcomes between the two groups (Fig. 2D-J). Whereas 86% of Trim28^+/D9-heavy animals reached the study endpoint illness-free, the majority of Trim28^+/D9-light animals showed aggressive MCS (Fig. 2D-J). Mean survival times were significantly different between Trim28^+/D9-light and -heavy morphs (55.4 and 67.3 weeks, respectively) (Fig. 2E). The single Trim28^+/D9-heavy animal that required pre-endpoint euthanasia had a bone-marrow-derived acute myeloid leukemia (AML; Fig. 2E-I, and 2J panel I). In contrast, Trim28^+/D9-light animals exhibited up to four different primary tumors per animal (Fig. 2H), mainly consisting of carcinomas, age-related carcinomas, and benign tumors (Fig 2F). Tumors in Trim28^+/D9-light animals were found in seminal vesicles, prostate, pancreas, lungs, esophagus, and colon (Fig. 2G, 2I and 2J panels II-VI). Histopathology of the Trim28^+/D9-heavy AML and representative Trim28^+/D9-light carcinomas are shown in Fig. 2J and Appendix 1B. Consistent with the differential timing of cancer onset between -light and -heavy animals, the endpoint analysis was dominated by -heavy tumors (Fig.S2D-H). Thus, Trim28^+/D9-light developmental morphs exhibit an accelerated MCS. These data provide genetic evidence that TRIM28-dependent epigenetic variation in development controls later-life cancer outcomes.

Trim28^+/D9-dependent cancer susceptibility states are distinguished by distinct early-life epigenomes.

We reasoned that if developmentally programmed epigenetic differences impact cancer susceptibility and outcomes late in life, then these differences should be detectable early on. We therefore performed DNA-methylation profiling on biopsies taken from all genotypes and animals at 10 days of age (i.e., before weaning). We used ear clips as they are minimally invasive for young mice and are similar to tissues used to identify early-life epigenetic signatures in humans^50–55. We used Illumina Infinium Mouse Methylation BeadChips to quantify DNA methylation state at ~285,000 CpG sites that included CpGs in all annotated genes, functional RNAs, and cis-regulatory regions of the mouse genome. Global DNA methylation profiles were highly correlated across genotypes, indicating a robust technical approach and high sample quality (Fig. S3A, row 1–4). Trim28^+/D9 and Tp53^R270H/+;Trim28^+/D9 compound heterozygotes had ~3 times as many differential methylated CpG loci (DML) than Tp53^R270H/+ alone (Fig. S3B). DMLs in Trim28^+/D9 and Tp53^R270H/+;Trim28^+/D9 animals overlapped strongly (Fig. S3A-3D), indicating that Trim28^+/D9 substantially and reproducibly changes the early life methylome even in the presence of the Tp53^R270H/+ mutation.

Relative to WT, Trim28^+/D9 biopsies were skewed towards reduced methylation (Fig. S3E). This is consistent with TRIM28’s known role in gene silencing ^{37,38,40,41,56}. Interestingly, Tp53^R270H/+ animals also showed early-life epigenetic changes, and these showed some similarity to Trim28^+/D9-induced changes (Fig. S3F). These data are consistent with reports that TP53 regulates DNA methylation^57–59. Trim28^+/D9 hypomethylated DMLs were enriched almost exclusively in regions known to be targeted by TRIM28, with probe set enrichment analysis (PSEA) revealing annotations for heterochromatin, monoallelic methylation, Polycomb-silencing, CTCF, TRIM28-binding, and H3K9me3 (Fig. S3G). These data show that full TRIM28 dosage is required to maintain early life DNA methylation fidelity at heterochromatic regions.

Importantly, substantial DNA methylation differences were also detected between isogenic Trim28^+/D9 animals that would go on to become -light versus -heavy, even though there are no detectable phenotypic differences at 10 days of age (Fig. 3A, S3A row 5–6, and S3H-I). We found a total of 1133 DMLs between the two Trim28^+/D9 morphs, and a clear skew towards hypomethylation in the -light, cancer-prone morph (Fig. 3A, S3I). Similar to the genotype as a whole, PSEA revealed differential methylation predominantly at regions of monoallelic methylation and imprinting, including the Kcnq1-Kcnq1ot1 cluster, H19, and Peg3 (Fig. 3B top panel, green, and bottom panel, pale orange), and regions annotated as H3K9me3- and H3K27me3-decorated (Fig. 3B top panel, light blue). A search for overlap with transcription factor binding (Fig. 3B top panel, purple) revealed strong and specific enrichment for chromatin binders involved in DNA-methylation (MBD1, MECP2, C17orf96, DPPA2 and TRIM28 itself) and Polycomb silencing machinery (SUZ12, EZH2, C17orf96, RNF2, AEBP2, PCGF2, CBX7, BMI1, and JARID2). Enrichments were also observed for probes within the epigenetic aging clock; enhancers; and at select subsets of transcriptional start sites (TssBiv and TssFlnk; Fig 3B top panel, green and red). Collectively, these data suggest that Trim28^+/D9-light animals have more permissive chromatin at regions that would otherwise be silenced. To our knowledge, these data represent the first epigenetic signatures of developmentally programmed cancer susceptibility states.

Figure 3. Trim28^+/D9-dependent cancer susceptibility states are distinguished by distinct early-life epigenomes that are enriched for epigenetic regulators and bona fide oncogenes.

A) The heatmap reports z-score(log1p) transformed beta values of differentially methylated probes in Trim28^+/D9-heavy vs -light animals. Effect size cut-off = 0.1; p-value cut-off = 0.001. N=24 (15 Trim28^+/D9-heavy vs 9 Trim28^+/D9-light). B) Enriched features (upper panel) and genes (lower panel) for differentially methylated probes between Trim28^+/D9-heavy vs -light animals. False Discovery Rate (FDR) cut-off = 0.01 (one-tailed Fisher’s exact test). C) Gene-ontology (GO) enrichment plot of biological processes enriched in DML between Trim28^+/D9-heavy vs -light animals. The size of the dots represents the gene-ratio, while the color represents the adjusted p-value for each term; Benjamini–Hochberg adjusted p-value cut-off = 0.05. D) Odds ratio from Fisher’s exact tests, testing the statistical significance of overlaps between genes enriched by hypo-methylated probes in Trim28^+/D9-heavy (dark orange) and -light animals (pale orange), and the COSMIC Cancer Gene Census (https://cancer.sanger.ac.uk/census). P-value = 0.0002, *** p<0.001. E) Left panel: Kaplan-Meier survival probability as percentage of the total population and the time of survival in months. Log-rank test, p=1.516e⁻⁶. Total cases analyzed = 10967. Right panel: Kaplan-Meier disease-free survival probability. Log-rank test, p=1.27e⁻⁹. In both panels all TCGA PanCancer Atlas patients mutated in genes from the Trim28^+/D9-light hypo-methylated signature (pale orange, n=3766) are compared to individuals not mutated in the same genes (black, n=7180). F) Heatmap of the effects on overall survival probability of mutations in the indicated genes and tumor tissues. The analysis includes all samples from TCGA and non-TCGA studies with no overlapping samples, from cBioPortal (N=69223 samples). Tumor tissues are separated in two main branches according to sample number informing the analysis (left: >3000 samples; right: <3000 samples). G) Volcano plot showing the type of interaction for pairwise mutations in genes from the Trim28^+/D9-light hypo-methylated signature, in all TCGA PanCancer Atlas patients. Red indicates co-occurrence and blue mutually exclusivity of pairwise mutations. One-sided Fisher’s Exact Test; Benjamini–Hochberg adjusted p-value cut-off = 0.05. Total cases analyzed = 10967. H) Our model suggests that TRIM28 buffers intrinsic developmental heterogeneity via heterochromatin silencing. By modulating a differentially methylated cancer-related gene set primes two distinct developmental trajectories for cancer susceptibility and outcomes, with one of the trajectories being more resistant and the other prone to cancer.

The Trim28^+/D9-sensitive epigenome is enriched for epigenetic regulators and bona fide oncogenes.

We next examined the specific genes that were differentially methylated between the -light and -heavy morphs at 10 days of age (DMGs; Fig. 3B bottom panel). Interestingly, DMGs hypomethylated in Trim28^+/D9-light animals were also enriched in epigenetic regulators of gene expression, heterochromatin formation, heterochromatin organization, genomic imprinting, DNA methylation and DNA alkylation, among others (Fig. 3B, bottom panel, pale orange, and 3C). Therefore, light-morphs exhibit hypomethylation at coding regions of epigenetic silencers and their targeted genomic regions (Fig. 3B, top left panel).

We also queried three independent resources to gain insight into whether dysregulation of the -light and -heavy DMGs alone would be expected to impact cancer outcomes. We search for their enrichment in the Jensen DISEASE database of disease-gene associations⁶⁰, and found that -light and -heavy DMGs were enriched for pathways underpinning human development and cancer (Fig. S3J), findings consistent with phenotypes in the mouse. Second, we tested for presence of -light / -heavy DMG orthologs in the COSMIC Cancer Gene Census (COSMIC), a database of high confidence oncogenes and tumor-suppressor genes⁶¹. In agreement with the aggressive MCS phenotype of -light morphs, DMGs ‘activated’ (hypomethylated) in Trim28^+/D9-light animals were significantly enriched for known oncogenes (GNAS, JAK3, MYCN, HMGA2) (Fig. 3D). Finally, we used the TCGA PanCancer Atlas⁶² and compared cancer outcomes of individuals with or without mutations in -light or -heavy specific DMGs. Interestingly, patients with mutations in Trim28^+/D9-light hypomethylated DMGs showed reduced overall survival probability (Fig. 3E, left panel) and a striking difference in time-to-relapse (Fig. 3E, right panel) compared to patients bearing other mutations. Stratifying the data by tumor-type showed wide-spread tumor promoting effects in orthologs of essentially all Trim28^+/D9-light DMGs, with significantly reduced survival rates across many tumor types (Fig. 3F and S3K). Noteworthy in the stratification analysis, prostatic adenocarcinoma showed cancer-accelerating associations with orthologs of almost all -light specific DMGs (Fig. 3F and S3K, leftmost column); prostatic adenocarcinoma was the most prevalent cancer type specifically in the Trim28^+/D9-light animals (Fig. 2G). No differences were found for the same analyses performed for orthologs of hypermethylated DMGs (i.e., those hypomethylated in -heavy animals) (Fig. S3L). Collectively, these data identify the hypomethylated Trim28^+/D9-light DMGs as putative mediators of the altered cancer susceptibility states found in Trim28^+/D9 animals.

As a final assessment of on-coregulatory potential, we tested for mutational co-occurrence of the same DMG ortholog sets in the TCGA PanCancer Atlas. Co-occurrence of mutations within groups of genes may suggest that those alleles can provide additive or synergistic tumor survival advantages. Consistent with the phenotype of the Trim28^+/D9-light morphs above, co-occurrence of mutations in -light DMG orthologs was markedly over-represented in human primary tumors, suggesting that they are part of a co-regulated signature (Fig. 3G). Thus, the early-life Trim28^+/D9-dependent epigenome is enriched for bona fide oncogenes. Collectively, these data suggest a model whereby intrinsic differences in early-life epigenetic programming may determine cancer outcomes (Fig. 3H).

DISCUSSION

Early-life epigenetic heterogeneity as a regulator of differential cancer susceptibility.

Here we identify (TRIM28-buffered) intrinsic developmental heterogeneity as a novel determinant of cancer susceptibility. The data show that TRIM28 haploinsufficiency generates two reproducible developmental morphs at the organismal level and that these differ in their cancer susceptibility; one “resistant”, and one more “susceptible”. Conceptually, the result is similar to epigenetic heterogeneity described within tumors⁶³ and between tumors^64,65, except at the inter-organismal level. How meta-stable states between identical cells or organisms are imposed remains unclear, though pioneering work on meta-stability of variegating reporters implicate epigenetic silencing machinery^66–70. Along those lines, it has been suggested that a key condition for the emergence of alternate cellular states is the epigenetic reorganization of the genome^34,71,72. Feinberg and Levchenko³⁴ recently provided an innovative theoretical framework for how genetic and epigenetic networks generate meta-stable partitions and alternate cellular functional states, a potential energy landscape model that include energy wells or ‘attractors’. DNA mutations and/or changes in epigenetic topology (via DNA mutation, DNA methylation, or histone modifications, for instance) alter that landscape and create new/alternate attractor states. Our data suggest that these same concepts hold true at the organismal scale, and that these differences can have real consequences for cancer outcomes. They suggest that TRIM28-dependent silencing helps define the shape of the potential energy landscape (e.g., by controlling the depth of or barrier between attractor states). In the same way that oncogenic mutations have different effects depending on cellular developmental stage^73–77, our data suggest that oncogenesis can also be influenced by stochastic, organismal epigenetic programs that are established in development.

Role of epigenetics in the developmental origin of cancer.

This study is one of the first attempts to bridge two key questions in the field of cancer epigenetics. One is a “cell of origin” question: how does the epigenetic state of a cell permit, support, or resist oncogenic transformation^14,16–23? The second is: how do early-life epigenetic cues change or modulate cancer risk (between individuals)^35,36,78,79? Addressing these questions requires isogenic models and identical environments, in part because individual development involves stochastic processes.

Our data show that inter-individual differences in early-life epigenome organization can dictate differential cancer development, prevalence, and survival. This finding complements prior work indicating that H3K9me3 (a target of TRIM28 action) most strongly correlates with tumor mutation density²⁰. Together with the latter work, the inter-individual epigenetic differences identified in our current study suggest that one potential mechanism for the observed differential cancer outcomes is an altered sensitivity to mutations between the two morphs. We also detected important differences in Polycomb-targeted genes, poised, and bivalent regions between the cancer susceptibility morphs. During tumorigenesis, these regions are particularly sensitive to regulation via DNA methylation and may be correlated with cell “stemness”^80–83. Regardless, the provocative implication arising from our data is that individual cancer susceptibility may have just as much to do with the epigenetic ‘background’ we are born with as it does DNA mutations, external environmental insults, or the cell of origin. Just as prior work demonstrated that epigenetic dysregulation at specific genes drives tumorigenesis in specific tissues or developmental stages, we would expect that the DMLs and DMGs identified here also have tissue-specific and developmental-stage specific effects. Analogous situations arise when different tumors exhibit mutations in different genes from a common or shared biochemical pathway^84–92, or when cancer-associated mutations have different effects depending on the cell differentiation stage^93–97. It would be interesting to understand when (during development) the different epigenetic backgrounds in the Trim28^+/D9 cancer susceptibility morphs become “activated”, and why some tissues seem to be more sensitive to tumorigenesis than others.

These novel findings differ from TRIM28’s published roles^42,98 that include context-specific oncogene^99–109 and tumor suppressor^110–115 functions, roles that were primarily determined from complete (homozygous) knockouts. In contrast to homozygous knockout models, the Trim28^+/D9 haplo-insufficient mouse exhibits near normal levels of TRIM28 throughout the body⁴⁴. Given TRIM28’s presence in multiple complexes (Trim28/KAP1 co-repressor complex, NuRD, CoREST, PML-NB, BORG/TRIM28, ZMYM2-TRIM28, MAGE-Trim28, or HUSH), this is an important distinction between the models. Indeed, TRIM28 homozygous deletion is embryonic lethal^116–118. The DNA-methylation differences between -light and -heavy morphs suggest that Trim28^+/D9 specifically impacts TRIM28’s silencing function. This fundamental difference between models may explain for instance why TRIM28 knockout models develop liver tumors^111,112, and Trim28^+/D9 mice do not.

Limitations of the study and future directions.

This work provides proof-of-concept that early-life, epigenetically distinct developmental programs can result in differential cancer susceptibility. The ability to show this effect in multiple tissue types is both a strength and a limitation of this study. Indeed, by focusing on an MCS model, we demonstrate that differential susceptibility is a property of the entire organism and can identify responsive tissues. At the same time, thousands of animals would be needed to draw the same conclusions for all the observed cancer sub-types including rare cancers. Likewise, and because the Trim28^+/D9 mutation in this model occurs in the whole body, it will be difficult to use the Trim28^+/D9 mouse (by itself) to dissect the molecular mechanisms down-stream of TRIM28 and heterochromatin disruption that underlies each of the observed cancers. A natural extension of this study is therefore to refine the model to understand the mechanistic basis of developmentally ‘primed’ cancer susceptibility for each cancer type.

Our data show that the epigenetic distinction between the two cancer susceptibility morphs is already evident by day 10, before weaning. Other open questions therefore relate to the precise timing of the observed (epigenetic) bifurcation, and the nature of any cell-intrinsic factors that might skew development towards one or the other epigenetic background. Our finding that the Trim28^+/D9 DMLs and DMGs are enriched for known human oncogenes and that they suggest a more permissive chromatin state at otherwise silenced regions, hints to a possible generalization of the model. If we can identify similarly sensitive regions of the human cancer genome, then we will be better equipped to optimally stratify and treat patients.

METHODS

Origin and maintenance of mice

FVB/NJ.Trim28^+/MommeD9 animals (Trim28^+/D9) were generated in the Whitelaw lab¹¹⁹. B6.129S4-Trp53<tm3.1Tyj>/J animals (Tp53^+/R270H) were originally generated in the Jacks lab⁴⁵ and purchased from Jackson Laboratories (stock #008182). Both lines were fully backcrossed for more than 10 generations (FVB/NJ and B6 respectively). After bringing lines in-house, they were both maintained by internal breeding with wild-type siblings. Approximately 270 F1 hybrids were generated by crossing 8-week-old FVB.Trim28^+/D9 males with 8-week-old B6.Tp53^R270H/+ females. One male was crossed with two females in the same cage, and females separated after plug checking the next morning. From these crosses, we generated 137 males (18 WT, 44 Tp53^+/R270H, 34 Trim28^+/D9, and 41 Tp53^R270H/+;Trim28^+/D9) and 133 females (30 WT, 36 Tp53^+/R270H, 32 Trim28^+/D9, and 35 Tp53^R270H/+;Trim28^+/D9). 114 animals were screened for tumors: 60 males (6 WT, 17 Tp53^+/R270H, 15 Trim28^+/D9, and 22 Tp53^R270H/+;Trim28^+/D9) and 54 females (8 WT, 21 Tp53^+/R270H, 8 Trim28^+/D9, and 17 Tp53^R270H/+;Trim28^+/D9). Only male offspring were analyzed; females exhibited unusually low levels of both phenotypic heterogeneity and cancer incidence precluding analysis.

All animals were fed breeder chow (Lab diet, 5021 cat. #0006540) ad libitum and housed in individually ventilated cages (Tecniplast, Sealsafe Plus GM500 in DGM Racks). All animals were kept on a 12-hour light/dark cycle at an average ambient temperature of 23 °C and 35% humidity. The maximum capacity per cage is 5 animals, and each cage was enriched with Enviro-dri (The Andersons, Crink-l’Nest) and cardboard dome homes (Shepherd, Shepherd Shack Dome). Whenever possible, same-sex siblings and same-sex animals from different litters were combined (~20 days of age) to co-house isogenic animals. At 4, 8, 16, 32, 40, 50, 60, and 70 weeks of age (or at euthanasia), mice were weighed and scanned via EchoMRI for fat and lean mass composition (EchoMRI^™, EchoMRI^™-100H), in the morning. All protocols were approved by Institutional Animal Care and Use Committee under protocols 19–0026, 22–09-036, 18–10-028, and 21–08-023, at Van Andel Institute (VAI, USA).

Genotyping

Ear punch biopsies were collected at 10 days and placed in a 20 µl reaction mix composed of genomic DNA lysis buffer (100 mM Tris-HCl pH 8.5, 5 mM EDTA, 0.2% SDS, 100 mM NaCl) supplemented with 20 mg proteinase K (Thermo Scientific, EO0491). Biopsies were digested with a thermal cycling protocol consisting of 55 °C for 16 hours, 95 °C for 10 minutes, and a 4 °C hold (lid at 105 °C). Thereafter, 160 µl of nuclease-free water (Invitrogen, AM9938) was added to each lysate to achieve 180 µl final volume. The PCR reaction (Thermo Scientific, EP0703) for Trim28 and Tp53 alleles comprised 1 µl diluted biopsy lysate and 19 µl reaction master mix (1X DreamTaq Buffer, 0.2 mM dNTPs, 0.1 µM primer forward and reverse mix, 2 U DreamTaq DNA Polymerase, in nuclease-free water), with PCR primer and thermal cycling conditions listed in the Supplementary Tables 1 and 2, respectively. To verify the presence of each point-mutation, 20 µl of each PCR product was digested with either 0.5 µl XceI/NspI (for Trim28^+/D9; Themo Scientific, FD1474) or 0.5 µl MslI (for Tp53^R270H/+; New England BioLabs, R0571L) in a final reaction volume of 30 µl. Digestion products (~700 bp WT Trim28, ~250 bp + ~450 bp Trim28^+/D9, ~500 bp WT Tp53, ~200 + ~300 bp Tp53^R270H/+) were detected on a 3% agarose gel (Fisher Scientific, BP160-500) in 1X TAE, with GelRed as the intercalating dye (Biotium, 41003).

Statistical analyses of developmental heterogeneity

We used Levene’s test¹²⁰ on body, fat, or lean masses independently to test for homoscedasticity (or equality of variances) across genotypes. P-values were subsequently adjusted for multiple comparisons using the Benjamini-Hochberg method¹²¹. We used MClust (version 5.4.9)¹²² for iterative Expectation-Maximization (EM) maximum-likelihood estimation in parameterized Gaussian mixture models. We chose to regularize with a prior to achieve smoother Bayesian Information Criterion (BIC). The uncertainty in the classification was used as graphical parameter for the lean/fat mass data plots, and for weighing the Log-Rank p-values in mouse Kaplan-Meier plots. Since most mice were classified with high confidence, the effect of this correction is negligible. We validated the results from MClust with Rmixmod (version 2.1.8)¹²³ using unsupervised classification and density estimation with 3 criteria: BIC, ICL, and NEC. Both methods were used to cluster 16-week fat and lean mass data for each genotype.

We used generalized additive models (GAMs)¹²⁴ to model the fat or lean mass changes over time per group. We used random-effect splines to model a random slope and random intercept for each animal (by week). We then used the “emmeans” R package^125,126 to compare the overall slope of the fat or lean mass by group. We ran a separate GAM to model the differences in fat or lean mass within each timepoint. For this model, we included a random-effect spline for each animal but excluded the spine for a random slope by week. Again, we used the “emmeans” package to compare the differences in the fat or lean mass within each timepoint. We used a two-sample test of proportions¹²⁷ under the “stats” R package^128,129 to examine the differences in proportion of animals that died with cancer before and after the endpoint in the Trim28 -heavy and -light trajectories. The p-values from those tests were adjusted for multiple testing using the Benjamini-Hochberg method.

Health monitoring

Professional vivarium staff checked mice for overall health, well-being, and the presence of any abnormal mass/tumor 2–3 times per week. Mice were euthanized if they showed any of the following symptoms: >20% weight loss, tumors that were ~15% of total body weight, tumor ulcerations, discharge or hemorrhage from the tumor, limited ambulation, reduced appetite and drinking, limited defecation or urination, abnormal gait or posture, labored breathing, lack of movement, or hypothermia. Mice with reported health issues or that reached the study endpoint (70 weeks of age) were euthanized via CO₂ asphyxiation and cervical dislocation.

Tissue harvesting

The following tissues were dissected and immediately fixed in 10% NBF solution (3.7–4% formaldehyde 37–40%, 0.03 M NaH₂PO₄, 0.05 M Na₂HPO₄, in distilled water with final pH of 7.2± 0.5): epidydimal white adipose tissue (eWAT); uterus or preputial glands, seminal vesicles, and testis (depending on sex); bladder; pancreas; spleen; intestine; stomach; mesenteric fat; liver; kidneys; heart; lungs; thymus; brain; the ninth breast; hindlimb muscles; and hindlimb bones. We also recovered spine, ribs, skull, skin, and any other mass if they appeared to be involved with a tumor or disease. The volume of fixative was at least 15–20 times greater than the volume of tissue. Specimens > 2.5 mm thick were cut to ensure adequate fixation. All the tissues but eWAT, mesenteric fat, uterus, and bones (including spine) were fixed for 48 hours. The fat-rich tissues (eWAT, mesenteric fat, uterus) were fixed for 72 hours. The bony tissues (bones and spine) were fixed for 1 week followed by 1 week decalcification in 14% EDTA (14% free-acid EDTA at pH 7.2, adjusted with NH₄OH). After each incubation, all the tissues were moved to 70% ethanol.

Tissue preparation for histology

All tissues were embedded in paraffin by the Van Andel Institute Pathology and Biorepository Core following internal standard operating procedures. Upon receipt, tissues were dehydrated and cleared using a Tissue-Tek VIP 5 (Sakura) and an automated protocol consisting of 60’ in 70% ethanol; 60’ in 80% ethanol; 2 x 60’ in 95% ethanol; 3 x 60’ in 100% ethanol; 2 x 30’ in xylene; and 1 x 30’ and 1 x 45’in paraffin. Tissues were embedded in paraffin with a Leica EG1150 embedding center. Three 5-µm thin sections spaced 150 µm apart were cut from each tissue for hematoxylin and eosin (H&E) staining using a Leica rotary microtome. The remaining tissue was conserved as a paraffin embedded tissue block. H&E staining was performed with a Tissue-Tek Prisma Plus Automated Slide Stainer (Sakura) and Tissue-Tek Prisma H&E Staining Kit #1.

Pathology evaluation

Standard 5-µm thick tissue sections stained with H&E were assessed for tumors and dysplastic lesions by a board-certified pathologist dedicated to this study. Most samples were provided to the pathologist in a blinded manner. Tumors were broadly classified as carcinomas, germ cell tumors, leukemias, lymphomas, and sarcomas. A detailed classification was provided based on the tissue of origin.

Mouse DNA methylation array

Ear punch biopsies were collected as described above, and DNA purified from lysate using a DNeasy Blood & Tissue Kit (QIAGEN, 69504) with slight modifications. After tissue digestion, the lysate was brought to 220 µl with 1X PBS. We then we followed steps 2–7 of the Quick-Start protocol. DNA was eluted with two washes of 100 µl Buffer AE, and purified DNA quantified by Qubit fluorometry (Life Technologies). Then, 6–500 ng of DNA from each sample was bisulfite-converted using the Zymo EZ DNA Methylation Kit (Zymo Research, Irvine, CA USA) following the manufacturer’s protocol and the specified modifications for the Illumina Infinium methylation assays. After conversion, all bisulfite reactions were cleaned using the Zymo-Spin binding columns and eluted in 12 µL of Tris buffer. Following elution, bisulfite-converted and restored DNA was processed through the Illumina mouse methylation array protocol. The mouse methylation array contains >285K probes for CpG islands, translation start sites, enhancers, imprinted loci, and other regions, along with strain-specific SNP probes¹³⁰. To perform the assay, 7 µl of converted DNA was denatured with 4 µl 0.1N sodium hydroxide. DNA was then amplified, hybridized to the Infinium bead chip, and an extension reaction performed using fluorophore-labeled nucleotides per the manufacturer’s protocol. Arrays were scanned on the Illumina iScan platform, and probe-specific calls were made using Illumina Genome Studio v2011.1 software to generate IDAT files.

DNA methylation analysis

Analysis of IDAT files was performed using the default SeSAMe pipeline¹³¹ and its wrapper pipeline SeSAMeStr¹³². Fifty-eight independent biological replicates from WT, Trim28, Tp53 and Trim28/Tp53 male animals were analyzed. Data pre-processing and quality controls were performed using SeSAMe default parameters and the pre-processing code ‘TQCDPB’. All samples showed a detection rate >90% and no dye bias. In all differential DNA methylation analyses, the effect size cutoff was set to 0.1 (i.e., 10% differential DNA methylation) and the p-value cutoff was <0.05, unless otherwise specified in the figure legends. For all differential DNA methylation analyses between isogenic Trim28 - heavy and -light animals, the effect size cutoff was set to 0.05 (i.e., 5% differential DNA methylation) and the p-value cutoff was <0.05. For all differential analyses, we included batch effect as a covariate in the model. Other technical and biological effects/bias (i.e., detection rate, initial DNA concentration, litter) were evaluated but not included in the model because they were co-linear with the batch effect and/or did not separate groups in dimensional reduction analyses. Global DNA methylation correlation analysis was performed using the ‘chart.Correlation’ function from the ‘PerformanceAnalytics’ R package. Similarity between samples was calculated as the sum of squared residuals from linear regressions. Principal component analysis (PCA) of beta values was performed on SeSAMeStr pipeline output using the R function ‘prcomp’. Heatmap visualization of differentially methylated loci (DML) was performed using the R package ‘ComplexHeatmap’¹³³. For heatmap visualization, beta values were modelled and weighted using the Mclust certainty score in limma¹³⁴. Probe enrichment analysis was performed using SeSAMe KnowYourCG module. Gene ontology analysis of probes-enriched genes was performed using the R package ‘clusterProfiler’¹³⁵. Gene enrichment in the Jensen DISEASES database¹³⁶ was performed using the R package ‘enrichR’¹³⁷. Further data visualization of SeSAMe/SeSAMeStr output was perform in R using Rstudio.

TCGA PanCancer Atlas data analyses

The TCGA PanCancer Atlas⁶² encompasses 32 studies and 10967 samples. All preliminary analyses were performed using cBioPortal^138,139, and outputs were used to replot and visualize the data in Rstudio. All Kaplan-Meier curves were generated by selecting all TCGA PanCancer Atlas cases that harbor mutations in either the Trim28^+/D9 -heavy or -light gene signatures, and comparing them with cases without mutations in the same genes. Statistical significance for differences between groups in all Kaplan-Meier curves was tested by log-rank tests with a p-value cut-off = 0.05. Mutual co-occurrence or exclusivity of pairwise mutations in genes within either the Trim28^+/D9 -heavy or -light gene signatures were tested by one-sided Fisher’s Exact Test, with a Benjamini–Hochberg adjusted p-value cut-off = 0.05. Statistical significance of overlaps between genes within either the Trim28^+/D9 -heavy or -light gene signatures and the COSMIC Cancer Gene Census⁶¹ was tested by a Fisher’s exact test with a p-value cut-off = 0.05. The effect of mutations at the Trim28^+/D9 -light signature genes was assessed in each group of tumors from all available tissue types in cBioPortal. This analysis comprised all samples from TCGA and non-overlapping samples from cBioPortal (N=69223 samples). Definition of ‘harmful’ and ‘protective’ mutations was based on the ratio of the median months survival between the samples with no mutations and the samples carrying mutations (i.e., median months survival unaltered/altered samples >1 = ‘harmful’ mutation; median months survival unaltered/altered samples <1 = ‘protective’ mutation). As the number of significant hits in this analysis is biased by the total number of samples for each targeted tissue, we split the tumor tissues into low (<3000) and high (>3000) samples number and showed them separated in the heatmap visualization. The same analysis was performed on the overall survival and the disease-free survival.

CONSORTIA

PERMUTE

J. Andrew Pospisilik, Ilaria Panzeri, Luca Fagnocchi, Stefanos Apostle, Emily Wolfrum, Zachary Madaj, Jillian Richards, Holly Dykstra, Tim Gruber, Mitch McDonald, Andrea Parham, Brooke Armistead, Timothy J. Triche Jr., Zachary DeBruine, Mao Ding, Ember Tokarski, Eve Gardner, Joseph Nadeau, Christine Lary, Carmen Khoo, Ildiko Polyak, Qingchu Jin.