Introduction

Colorectal cancer (CRC) is the second deadliest cancer in the United States, with an estimated 147,950 individuals diagnosed and 53,200 deaths in 20201. Although there have been great reductions in CRC incidence and mortality widely attributed to increased screening2, the incidence of CRC has been increasing in individuals less than 50 years of age at a rate of 2% per year since 19943. Molecular analysis has shown that < 20% of early-onset CRC cases are explained by genetically determined hereditary syndromes4 and a variety of environmental factors have been postulated to underlie its increase5, suggesting that a unitary cause of early-onset CRC will be elusive. With early-onset CRC manifesting as a heterogenous disease caused by a multitude of factors, there is a pressing need to identify the distinct molecular subtypes of CRC that are overrepresented in early-onset cases.

Somatic mutation of the adenomatous polyposis coli (APC) gene is the initiating event in approximately 80% of all CRCs, but APC mutations are significantly less frequent in early-onset CRCs6,7,8. APC is a structural and regulatory component of a destruction complex, which responds to WNT stimulation by inhibition of degradation of the stem cell-promoting transcription factor β-catenin, encoded by the CTNNB1 gene9. Failure to regulate β-catenin by degradation due to mutational inactivation of APC throws colorectal epithelial cells into a continuous “WNT-activated” state; they no longer require activation by WNTs to maintain the stem cell compartment10. The fact that early-onset CRCs more frequently lack an APC mutation suggests that many of these tumors depend on alternative molecular mechanisms. In mismatch repair-deficient CRCs, which exhibit microsatellite instability (MSI) and constitute 12–15% of all CRCs, APC mutations are also significantly less frequent and BRAF mutations constitute a dominant driver mechanism11. What initiates and drives the carcinogenetic process in microsatellite stable (MSS) CRCs that lack APC mutations? Here, we comprehensively compare molecular profiles of MSS APC mutation-positive CRCs (APCmut+) and MSS APC mutation-negative (APCmut–) CRCs to identify novel APC-independent mechanisms driving CRC subtypes.

Methods

Analyses of genomic alterations in case series

To formulate a discovery series, we obtained colon adenocarcinoma (COAD) and rectal adenocarcinoma (READ) data from The Cancer Genome Atlas (TCGA) in the Genomic Data Commons portal (Supplementary File 1). Curated somatic nucleotide variant and copy number data were extracted using TCGAbiolinks and FireBrowse, respectively12. Deep deletions, amplifications, and gene fusions13 were identified. We excluded hypermutable cases by removing MSI-high cases based on clinical data and by removing cases with > 700 mutations. CRC samples were classified as APCmut– if they lacked a non-silent mutation or deep (homozygous) deletion in APC or lacked a mutation in CTNNB114. With these filtration steps, we had 63 APCmut– samples and 362 APCmut+ samples. We compared genomic alterations between APCmut– and APCmut+ CRCs by Fisher’s exact test and tested mutual exclusivity by CoMEt15. For more details on the bioinformatics analysis, see Supplementary Methods.

For validation series, we used the CPTAC-216 and GSE3589617 datasets, because they were the only CRC datasets with APC mutation status and gene expression data available from the cBioPortal, the International Cancer Genome Consortium or studies utilized by Guinney et al. to determine consensus molecular subtypes of CRC18. CPTAC-2 was downloaded from cBioPortal and GSE3896 from synapse.org18. In the CPTAC-2 dataset, we identified 11 APCmut– CRCs and 70 APCmut+ CRCs. Because GSE35896 did not have whole exome sequencing or copy number data, we could not filter for hypermutation, APC deep deletion, or CTNNB1 mutations. Based on the data available, we classified 16 out of 56 MSS CRC samples as APCmut–.

Transcriptomic analyses

For TCGA and CPTAC-2 series, we obtained HTSeq count files and used the edgeR and limma pipeline to normalize the counts matrix. For GSE35896, we used the RMA normalized data. Genes that had less than one count per million in more than half the samples were discarded. MBatch analysis showed no evidence of batch effects.

We used limma to identify differentially expressed genes between APCmut– and APCmut+ CRCs (Padj < 0.05). PathView was used to map differentially expressed genes onto the KEGG WNT canonical signaling pathway, with the node sum parameter set to “max.abs”19. Gene set enrichment analysis (GSEA) was performed using fgsea20. For GSEA input, we used the Gene Ontology (GO) biological process gene sets from MSigDB and ranked the genes by the t-statistic from our differential expression analysis. The Cytoscape application EnrichmentMap was used to visualize all statistically significant GO terms (Padj < 0.05)21. CIBERSORTx was used to impute the fraction of immune cells based on gene expression data from the TCGA, GSE35896, and CPTAC-2 datasets22.

To characterize the theoretical WNT ligand sensitivity of APCmut– tumors, we defined a score by the normalized expression of RSPO3 minus the sum of the normalized expression levels of RNF43 and ZNRF3:

$$ {\text{WNT}}_{{{\text{LS}}}} = {\text{Z}}\left[ {RSPO3_{mRNA - z} - (RNF43_{mRNA - z} + ZNRF3_{mRNA - z} )} \right], $$

where the subscript mRNA-z indicates the z-score of the expression value relative to all samples including tumors and normals. The WNTLS score for each tumor was then compared to the maximum WNT ligand expression over all 12 WNTs in the same tumor.

DNA methylation analyses

Methylation was assayed by TCGA using Illumina Human Methylation 450 arrays and data was accessed using TCGAbiolinks. Preprocessing and normalization were carried out with the R package minfi23. MBatch analysis showed no evidence of batch effects. Differentially methylated regions (DMRs) were identified using DMRcate and annotated with annotatr24,25. For DMRs that spanned multiple gene regions, we selected the gene with the most significant beta-values. To quantify methylation of a DMR, we took the average of all the statistically significant beta-values associated with the DMR.

DepMap data analyses

DepMap data was obtained from https://depmap.org/portal/download/26. CRC cell lines were selected excluding those with MSI and with > 800 mutations. To distinguish APCmut– from APCmut+ CRC cell lines, we used the same filtering steps we used for the TCGA dataset. To assess the effect of CRISPR knockouts, we applied a Welch’s two-sample t-test statistic to the dependency scores of APCmut– and APCmut+ cell lines. Dependency scores were extracted from the file “Achilles_gene_dependency.csv” on the DepMap portal.

Ethics statement

Ethics approval is not required for this study because it does not involve human participants or animal subjects.

Results

Age effect in APC mut– CRCs

To identify characteristics that distinguished APCmut– from APCmut+ CRCs, we compared molecular profiles between the two groups in a discovery series from the TCGA, then validated the results in two additional publicly available series. CRC samples that exhibited MSI or were hypermutated were excluded from our study, because tumors with these characteristics constitute a well-defined subtype11. In addition to separating MSS and non-hypermutated CRCs by APC mutation status, samples that contained a CTNNB1 mutation14 or deep deletion of APC were also classified as APCmut+. After applying these filtration steps, we classified 63 of 425 (15%) of the MSS CRCs in the TCGA dataset as APCmut–. In the GSE35896 validation dataset, 16 out of 56 (29%) CRCs were classified as APCmut– and in the CPTAC-2 dataset, 11 out of 81 (14%) CRCs were classified as APCmut–.

We tested clinical features that might be statistically associated with TCGA APCmut– CRCs (Table 1). In agreement with previous studies6,7,8, APCmut– CRCs were diagnosed at a younger age (61.4 in APCmut– vs. 66.4 in APCmut+), and 63% of tumors diagnosed < 50 were APCmut–. APCmut– CRCs were also younger in the CPTAC-2 dataset (61.5 in APCmut– vs. 65.5 in APCmut+), but this observation did not reach statistical significance (p = 0.24). (Age of diagnosis was not available for the GSE35896 dataset.) In addition to age, TCGA APCmut– CRCs were more prevalent in Asians (p = 0.005), were more likely to be classified as CpG island methylator phenotype (CIMP) high (p = 0.02), and were more likely to be diagnosed later than stage one (p = 0.035).

Table 1 Comparison of clinical features in APC mutation-positive (APCmut+) and APC mutation-negative (APCmut–) colorectal cancers.
Full size table

WNT signaling mutations in APC mut– CRCs

To identify distinctive somatic mutations, we compared non-silent nucleotide variants, gene amplifications, deep gene deletions, and gene fusions in APCmut– and APCmut+ CRCs (Fig. 1A). The top three most statistically different genomic alterations specific to APCmut– CRCs were PTPRK-RSPO3 gene fusions (p = 1.3 × 10–5), RNF43 mutations (p = 4.7 × 10–5) and BRAF mutations (p = 1.9 × 10–4). These genetic alterations have been identified in CRC previously with evidence for mutual exclusivity with APC mutations27,28,29,30. (The RNF43 mutation G659Vfs*41, which is associated with MSI CRCs, was not present in the tumors analyzed here as MSI tumors were excluded from this analysis31). Eight of nine BRAF mutated APCmut– CRCs had the oncogenic V600E BRAF mutation. Six of eight RNF43 mutated APCmut– CRCs had mutations that caused premature protein truncations, whilst one sample had a previously identified missense mutation, R554G. These findings suggested that BRAF and RNF43 mutations are associated with tumor progression in MSS APCmut– CRCs.

Figure 1
figure 1

WNT signaling mutations in APCmut– CRCs. (A) Fraction of APCmut– CRCs from the TCGA dataset with gene mutations, amplifications, deep deletions, and fusions that were significantly more common in APCmut– in comparison to APCmut+ CRCs. The top 10 are shown by p-value ranking, most significant (left) to least (right). (B) OncoPrint diagram showing the top 10 statistically significant mutations associated with APCmut– CRCs and the gene fusion PTPRK-RSPO3 for the 63 APCmut– CRCs.

Full size image

Based on the mutated genes in Fig. 1A and the PTPRK-RSPO3 gene fusion, we found that only 37 out of 63 samples (59%) contained a genomic alteration that distinguished APCmut– from APCmut+ CRCs (Fig. 1B). No pairwise combination of genes were statistically mutually exclusive. However, PTPRK-RSPO3 gene fusions and RNF43 mutations never co-occurred and were found in 23% of the APCmut– CRCs. After disregarding overlapping genomic alterations, BRAF mutations were the next most abundant (10%), followed by mutations in ADGRL1 (6%), ERBB3 (5%), and ZAP70 (5%). Supporting these findings, we found that BRAF mutations in the GSE35896 dataset and mutations in RNF43, ERBB3, and ZAP70 in the CPTAC-2 dataset were more frequent in APCmut– CRCs than in APCmut+ CRCs (Supplementary Fig. 1). (No additional mutation information was provided with the GSE35896 dataset.)

Enhanced sensitivity to extracellular WNT in APC mut– CRCs

Because a distinctive somatic mutational mechanism was not evident in over 40% of APCmut– CRCs, we examined transcriptomics data for further distinguishing molecular characteristics. Strikingly, in differential gene expression analysis of the TCGA dataset, RNF43 was the most differentially expressed gene between the two tumor groups (Padj = 4.6 × 10–15; Fig. 2A), with a -0.98 log2 fold decrease in mean expression level in APCmut– CRCs. Consistent with these results, RNF43 was also down-regulated in APCmut– CRCs in the GSE35896 and CPTAC-2 validation datasets (Fig. 2B). RNF43 and its family member ZNRF3 are membrane-bound E3 ubiquitin ligases that actuate the degradation of low-density-lipoprotein-related protein (LRP)-FZD WNT receptors. Binding of R-spondins to leucine-rich repeat-containing G-protein coupled receptors (LGR) leads to sequestration and membrane clearance of RNF43 and ZNRF3 from the cell surface32,33,34. The transcriptional down-regulation of RNF43 we found in APCmut– CRCs suggested that these tumors may express higher levels of LRP-FZD receptors at the cell surface, and consequently be more responsive to extracellular WNTs.

Figure 2
figure 2

Enhanced sensitivity to extracellular WNT in APCmut– CRCs. (A) Volcano plot representing the results from differential expression analysis between APCmut– and APCmut+ CRCs. Labeled points are the genes with an Padj < 0.0005. Blue points were downregulated in APCmut– CRCs and red points upregulated. (B) Comparison of RNF43 gene expression in APCmut– CRCs, APCmut+ CRCs, and normal colon samples in the TCGA, GSE35896 and CPTAC-2 datasets. Two-sample t-tests with a two tailed p-value were used to test statistical significance. (C) Differentially expressed genes (Padj < 0.05) between APCmut– and APCmut+ CRCs from TCGA were mapped onto the KEGG canonical WNT signaling pathway. Blue labeling represents genes downregulated in APCmut–; red labeling represents upregulated genes. (D) Unsupervised clustering analysis of APCmut– CRCs from the TCGA dataset using differentially expressed genes (Padj < 0.05). (E) Scatter plot showing estimation of activation potential of extracellular WNT signaling. Each point is the mean for individual groups. The y-axis represents a group’s apparent sensitivity to extracellular WNT signaling using the WNT ligand sensitivity score. The x-axis represents a group’s WNT stimulation potential by quantifying each sample’s maximum WNT ligand expression.

Full size image

Because WNT signaling was implicated by these results, we sought to determine the extent to which other factors in the canonical WNT signaling pathway were differentially expressed between APCmut– and APCmut+ CRCs (Fig. 2C). Consistent with the results above, we observed that other genes involved in extracellular WNT signaling were dysregulated, namely RSPO3 and ZNRF3. Differences in RSPO3 and ZNRF3 mRNA expression showed a similar trend in the validation datasets and were statistically significant in select cases (Supplementary Fig. 2). We did not observe any differential expression of the extracellular WNT regulator genes LGR4, LGR5, LGR6, or LRP-FZD receptors. The fact that LGR4-6 were not differentially expressed between APCmut– and APCmut+ CRCs was consistent with the finding that RSPO3 does not require interaction with LGRs to potentiate WNT signaling35 and LRP-FZD receptor levels are regulated post-transcriptionally36. When we compared gene expression of APCmut– and APCmut+ CRCs to normal samples and mapped genes onto the canonical WNT signaling pathway, changes in gene expression in WNT signaling were similar between these two tumor types (Supplementary Fig. 3). These results suggested that both types of CRCs exploit changes in WNT signaling. However, based on the mutation and expression data, APCmut– CRCs appear to favor dysregulation of genes involved in response to extracellular WNT signaling, whereas APCmut+ CRCs are stuck in the “on” state and are WNT signal-transduction incompetent.

To determine the fraction of APCmut– CRCs that operate via enhanced sensitivity of extra-cellular WNT, we performed unsupervised hierarchical clustering using all differentially expressed genes (Padj < 0.05) between APCmut– and APCmut+ CRCs in the TCGA dataset (Fig. 2D). APCmut– CRCs clustered into two prominent groups, referred to here as Cluster 1 (CL1) and Cluster 2 (CL2). Most APCmut– CRCs with a PTPRK-RSPO3 fusion, BRAF mutation, or RNF43 mutation were in CL2. To characterize the expression profiles of APCmut– CRCs in the context of extracellular WNT signaling, we computed a summarized score defined as RSPO3 expression minus the sum of RNF43 and ZNRF3 expression. This score represents a theoretical WNT ligand sensitivity (WNTLS) based on the known function of RSPO3 in increasing ligand sensitivity, and RNF43 and ZNRF3 in decreasing ligand sensitivity33,34. We examined how the WNTLS score tracked with maximum WNT ligand expression (Fig. 2E; see Methods for more details). Consistent with our expectation, APCmut– CRCs with RNF43 mutations had higher WNTLS scores than APCmut+ CRCs and higher maximum WNT expression than normals. Interestingly, APCmut– CRCs with PTPRK-RSPO3 fusions had the highest WNTLS score, but had the lowest expression of WNT ligands compared to other CRCs. Inconsistencies in how CRCs with PTPRK-RSPO3 fusions and CRCs with RNF43 mutations enhance their sensitivity to extracellular WNT signaling may be due to different selective pressures during cancer evolution.

APCmut– CRCs with BRAF mutations also exhibited higher WNTLS and higher WNT ligand expression, similar to APCmut– CRCs with RNF43 mutations. Importantly, APCmut– CRCs from CL2 that did not have BRAF mutations, RNF43 mutations, or PTPRK-RSPO3 fusions exhibited higher WNTLS scores compared to APCmut+ CRCs. In contrast, APCmut– CRCs from CL1 exhibited WNTLS scores similar to APCmut+ CRCs. APCmut– CRCs from the GSE35896 and CPTAC-2 datasets also clustered into two groups with high and low WNTLS scores (Supplementary Fig. 4). Given the importance of WNT signaling in CRC, these results suggest that other WNT-related mechanisms drive CL1 APCmut– CRCs. By transcriptomic analysis, CL1 APCmut– CRCs were practically indistinguishable from APCmut+ CRCs; however, GSEA showed enrichment of oxidative phosphorylation genes (Supplementary Figs. 5 and 6), implicating mitochondrial activation in CL1 APCmut– tumorigenesis. These results were supported by data from the DepMap CRISPR screen that demonstrated dependence of APCmut– CRC cell lines on oxidative phosphorylation complexes in the mitochondria (Supplementary Fig. 5E).

APC mut– CRCs associated with immune infiltration

GSEA analysis showed that GO terms related to the adaptive immune response were upregulated in APCmut– compared to APCmut+ CRCs (Fig. 3A). To further investigate immune system involvement in APCmut– CRCs, we employed the bulk tissue deconvolution method CIBERSORTx22. In agreement with the GSEA results, the CIBERSORTx absolute score was increased in APCmut– compared to APCmut+ CRCs in all three CRC datasets (Fig. 3B). The CIBERSORTx absolute score was highest in APCmut– CRCs with BRAF or RNF43 mutations and CL2 APCmut– CRCs without mutations (Fig. 3C). Because these APCmut– CRCs had more infiltrating immune cells than those with PTPRK-RSPO3 fusions, we tested whether any of the 22 immune cell types were associated with expression of WNT agonist ligand RSPO3 (Fig. 3D). We found that M2 macrophages had the strongest positive Pearson correlation with RSPO3 expression. M2 macrophages and RSPO3 expression were also significantly correlated in the GSE35896 and the CPTAC-2 datasets. Macrophage expression of RSPO3 was shown in a study of patients with pulmonary fibrosis37.

Figure 3
figure 3

APCmut– CRCs associated with immune infiltration. (A) GSEA results of differential gene expression analysis of APCmut– versus APCmut+ CRCs from the TCGA dataset. Red clusters represent GO terms enriched among upregulated genes in APCmut– CRCs and blue clusters correspond to down-regulated processes. (B) CIBERSORTx absolute score in CRCs from the TCGA, GSE35896 and CPTAC-2 datasets. Two-sample t-tests with a two-tailed p-value were used to test statistical significance. (C) Violin plot of CIBERSORTx absolute score across subtypes of APCmut– CRCs. (D) Expression of RSPO3 in APCmut– and APCmut+ CRCs plotted against their individual M2 macrophage scores identified from the CIBERSORTx algorithm. Pearson correlation was performed to determine statistical significance.

Full size image

APC mut– CRCs have higher AXIN2 methylation

Because we found an association between APCmut– CRCs and CIMP-high previously7, we identified differentially methylated regions (DMRs) between APCmut– and APCmut+ CRCs. APCmut– CRCs were globally more hypermethylated than APCmut+ CRCs, with a particular excess in promoter regions (Fig. 4A). Comparing the top ten hypermethylated and hypomethylated DMRs, we did not observe the same statistically significant genes as we did in the mutation and expression analyses (Fig. 4B). However, when we tested correlation of RNF43 expression with DNA methylation levels of DMRs and with RNA expression, we found that methylation and gene expression of AXIN2 had the highest correlations (Fig. 4C). RNF43 gene expression was also significantly correlated with AXIN2 expression in the GSE35896 and CPTAC-2 datasets (Fig. 4D; methylation data was not available in these datasets). Increased AXIN2 DNA methylation was associated with decreased RNF43 expression in a subset of APCmut– CRCs that did not have one of the common somatic mutations (Fig. 4E). Similar to our findings with RSPO3 expression, we found that M2 macrophages correlated most with AXIN2 DNA methylation (Fig. 4F).

Figure 4
figure 4

APCmut– CRCs have higher AXIN2 methylation. (A) Bar plot comparing total number of hyper-methylated and hypo-methylated differentially methylated regions (DMRs) between APCmut– and APCmut+ CRCs from the TCGA dataset. (B) Top 10 APCmut– hypermethylated and hypomethylated DMRs between APCmut– and APCmut+ CRCs from TCGA. Red bars represent APCmut– CRC differentially hypermethylated genes and blue bars represent APCmut– CRC differentially hypomethylated genes. (C) Bar plot representing DMRs with strongest correlations with RNF43. Blue bars represent the top 10 DMRs with the highest Pearson gene expression correlation with RNF43 gene expression. Red bars represent the Pearson correlation between the average differentially methylated beta values and RNF43 expression for these differentially methylated regions. (D) Scatter plots of RNF43 expression and AXIN2 expression of both APCmut– and APCmut+ CRCs in the TCGA, GSE35896, and CPTAC-2 datasets. Pearson correlation was performed to determine statistical significance. (E) Matched comparison between Z-normalized AXIN2 average beta values and Z-normalized RNF43 expression of APCmut– CRCs. (F) Scatter plot of AXIN2 average beta values and the CIBERSORTx M2 macrophage score of APCmut– and APCmut+ CRCs from the TCGA dataset. Pearson correlation was performed to measure statistical significance.

Full size image

AP2M1 gene expression associated with earlier onset in APC mut– CRCs

Age of onset was not different in CL1 and CL2 APCmut– CRCs (Fig. 5A). To identify gene expression changes linked to earlier age of onset in APCmut– CRCs, we separated APCmut– CRCs into two groups based on the median expression of each gene and performed a logrank test between these two groups, using the age at diagnosis as the event variable. Expression of AP2M1 best distinguished earlier onset APCmut– CRCs from later onset APCmut– CRCs (Fig. 5B), and higher AP2M1 expression was associated with earlier onset in APCmut– relative to APCmut+ CRCs (Fig. 5C).

Figure 5
figure 5

AP2M1 gene expression associated with earlier-onset in APCmut– CRCs. (A) A comparison of age between APCmut– clusters identified from Fig. 2D. A two-sample t-test with a two-tailed p-value was used to determine statistical significance. (B) Top 10 statistically significant genes based on a logrank test whose median gene expression best separates age of CRC diagnosis of APCmut– CRCs from the TCGA dataset. (C) Kaplan–Meier plot representing association between the age at CRC diagnosis and median separation of AP2M1 expression in APCmut– and APCmut+ CRCs from the TCGA dataset. (D) Flowchart of two molecular mechanisms that may be involved in the development of APCmut– CRC.

Full size image

Discussion

Most CRCs are initiated by somatic mutation of the gene APC, leading to ligand-independent, constitutive activity of the WNT pathway. In the present study, we found two alternate ways in which APCmut– CRCs may activate the WNT pathway. APCmut– tumors clustered into two groups according to their transcriptomic profiles (Fig. 5D). One cluster (CL2) exhibited a variety of molecular alterations that were consistent with the hypothesis that these tumors have enhanced sensitivity to extracellular WNT ligands. In particular, the most significant change was downregulation of RNF43, which is expected to result in increased levels of WNT receptors and greater sensitivity to extracellular WNTs. AXIN2 methylation was highly correlated with RNF43 downregulation. AXIN2 and RNF43 are negative regulators of WNT signaling that are transcriptionally activated by nuclear β-catenin, consistent with the notion that epigenetic silencing of negative regulators plays a critical role in tumor formation in ligand-dependent, APCmut– CRCs. Similarly, PTPRK-RSPO3 gene fusions drive R-spondin signaling, which is also expected to reduce RNF43 levels at the cell surface, upregulate WNT receptors, and enhance sensitivity to extracellular WNTs. We defined a WNT ligand sensitivity score to quantify this signature of extracellular WNT signaling in a sample-specific fashion and found a high WNTLS score was associated with CL2 APCmut– CRCs in multiple independent datasets.

Germline mutations in RNF43 have been previously associated with serrated polyposis families, and somatic mutations in RNF43 and BRAF have been associated with sporadic serrated adenomas38. In a preliminary analysis, we found that CL2 APCmut– CRCs expression profiles appear to be more similar to serrated adenomas than CL1 APCmut– CRCs and APCmut+ CRCs, according to two published gene signatures39,40 (data not shown), but these results need further investigation.

We also found that CL2 APCmut– CRCs have a higher level of immune infiltration compared to APCmut+ and CL1 APCmut– CRCs, especially in APCmut– CRCs that had RNF43 or BRAF mutations. M2 macrophages had the strongest association with potentiating WNT signaling through its significant correlations with RSPO3 expression and AXIN2 DNA methylation. Previous studies have shown that macrophages have the capability to express RSPO3 and stimulate WNT signaling in response to tissue damage41,42. The association of CL2 APCmut– CRCs with M2 macrophages suggests the etiology of this cancer subtype is tied to chronic tissue stress and inflammation that eventually favors a clone with hypersensitivity to WNT. We suggest that CL2 APCmut– CRCs may be sensitive to porcupine inhibitors or anti-WNT/anti-DKK1 biologics. We note that AXIN2 methylation has been previously identified in APCmut– CRCs as a potential biomarker for ligand-dependent tumors that would respond to anti-WNT-based therapies such as porcupine inhibitors43,44,45.

In contrast, the other cluster (CL1) of APCmut– CRCs was associated with low WNTLS score and may be dependent on enhanced mitochondrial activation. APCmut– CRC cell lines from the DepMap database had a strong dependency on mitochondrial activation relative to APCmut+ CRC cell lines. We are cautious in interpreting these data, because the observed effectiveness of mitochondrial disruption of the APCmut– CRC cell lines may be due to the absence of immune cells in vitro. One potential reason why some APCmut– CRCs become dependent on enhanced mitochondrial activation is because mitochondria can stimulate the WNT pathway independently of WNT ligands46. Moreover, intestinal epithelial cell-specific knockout of TFAM, a transcription factor required for replication of mitochondria DNA, drastically reduced tumor formation in APCmin/+ mouse models47. Therefore, we suggest that enhanced activation of mitochondria is a second, independent mechanism by which APCmut– CRCs exploit WNT signaling in tumor progression. These findings also suggest that mitochondria inhibitors may be a promising therapeutic option for CL1 APCmut– CRCs.

Although APCmut- tumors overall exhibit a lower age of onset than APCmut+ tumors, we found no difference in age of onset between CL1 and CL2, suggesting that both extracellular WNT sensitivity and mitochondrial activation contribute to the incidence of early-onset CRC. We performed a APCmut–wide analysis to determine what gene expression feature was most associated with age of onset and found that earlier-onset APCmut– CRCs had increased expression of AP2M1. AP2M1 plays an important role in clathrin-mediated endocytosis48. A recent study showed that when insulin binds to an insulin receptor, IRS1 and IRS2 recruit AP2M1 to initiate insulin receptor endocytosis49. Thus, an increase of AP2M1 may suggest increased insulin signaling. Importantly, insulin can activate both the PI3K pathway and the MAPK pathway, which may in turn play a role in enhancing both mitochondrial activation and immune infiltration, thus contributing to driving both CL1 and CL2 subtypes of APCmut– CRC50,51,52. Other studies have found that individuals with type two diabetes are at a greater risk for early-onset CRC53,54.

Early-onset CRC is a rapidly advancing public health emergency, and it is associated with a lack of mutation in APC. Our comprehensive genomic analysis has uncovered two classes of APCmut– CRCs, one which potentiates WNT signaling through sensitivity to extracellular signaling, and the other which exhibits mitochondrial activation. Future research should test the effect of anti-WNT biologics and mitochondrial inhibitors in organoid models and in vivo and compare the efficacy of AXIN2 methylation and WNT ligand sensitivity score in identifying anti-WNT sensitive tumors.