RNA-based gene expression profiles generated by our group have revealed numerous RNA markers that are differentially expressed in SSLs and cADNs . Here, using ISH, we verified the accuracy of 9 of the 12 markers putatively capable of distinguishing between these two major premalignant tumor types. Those that appeared to be SSL-specific, however, were unable to differentiate these lesions from HPs. Evidently, the gene-expression trajectories underlying the early stages of serrated tumorigenesis in these two serrated precursor lesions are common (see Introduction). HP-specific markers were also not found among additional 9 RNAs investigated in this study. The advanced TSAs (>10mm diameter) we investigated showed mixed staining patterns reflecting the coexistence in each lesion of serrated and cADN-like histologic features (Table 1).
One of the three genes that displayed particularly high expression in SSLs and HPs was VSIG1 (V-set and immunoglobulin domain containing 1) (Figure 2, Supplementary Figure 4). Absent in cADNs and the normal colon mucosa, VSIG1 expression was very high along almost the entire length of the serrated crypts in SSLs and HPs, except the very bottom of the crypt and its surface. TSAs, in contrast, display very little VSIG1 expression or none at all (e.g., TSA 2 in Table 1). The expression that is observed is reflected by patchy staining confined to glands with an SSL-like phenotype. The strikingly different VSIG1 expression patterns in serrated crypts (highly expressed) and those of the normal colorectal mucosa (unexpressed) might one day be exploited to improve detection of flat serrated lesions using fluorescein-labeled anti-VSIG1 antibodies during colonoscopy .
The VSIG1 protein, a member of the junctional adhesion molecule family, is normally expressed in the gastric mucosa and testis . Its ectopic expression in serrated colorectal lesions, which has been documented at both the transcript and protein levels , is thought to reflect aberrant differentiation toward a gastric-cell phenotype during the development of these tumors. SSLs and HPs also acquire expression of other molecules typically found in the gastric mucosa, e.g., ANXA10 (Annexin 10), a known marker of the normal mucosa of the stomach . ANXA10 belongs to the calcium-dependent phospholipid-binding annexin protein family, and its function is currently unknown. Like VSIG1, ANXA10 is unexpressed in the normal colon mucosa and in most cADNs. (In rare cases, moderate expression can be observed in a few cells or crypts on the surface of cADNs.) ANXA10 can be considered a bona fide marker of serrated glands in SSLs and HPs (Figure 2 and Supplementary Figure 5), and it is also encountered fairly often in cells on the surface of TSAs and in their SSL-like glands.
The third gene that was highly expressed in SSLs and HPs, ACHE (acetylcholinesterase), is instead typically expressed in conducting tissues, including those of the enteric nervous system, and at neuromuscular junctions . It terminates signal transmission by hydrolyzing the neurotransmitter acetylcholine at cholinergic synapses in the brain and at neuromuscular junctions, and pharmacologic inhibition of this enzymatic activity is used to treat colonic pseudo-obstruction . We found high levels of ACHE mRNA in epithelial cells at the surface of the normal colorectal mucosa, but even higher levels were found in serrated glands, extending about half-way down toward the base of the crypts (Figure 2, Supplementary Figure 6). By contrast, ACHE expression is markedly lower in cADNs, although it may be found in some cells on the surface of adenomatous villi. In TSAs, some serrated glands are strongly positive for ACHE expression, and this feature might be used to better visualize the serrated component of these polyps. ACHE is also expressed in some stromal cells—probably lymphocytes—and in some cells of lymphocytic folliculi. As expected, it is also strongly expressed in the submucosal plexi (Supplementary Figure 6).
Like VISG1, ANXA10 and ACHE, SEMG1 (Semenogelin 1) also emerged as a good marker of the serrated pathway of tumorigenesis . It is not expressed in cADNs or in the normal colon mucosa, but moderate levels are found in SSLs and HPs, along the length of serrated crypts and to a somewhat lesser extent at the crypt bases and mouths (Figure 3 and Supplementary Figure 7). SEMG1 expression is also appreciable in regions of TSAs where the serrated glandular differentiation is more obvious. This gene, too, is involved in a curious form of dysregulated cell-fate differentiation that occurs during serrated tumorigenesis. SEMG1 (like SEMG2, which is also strongly expressed is some SSLs ), is typically expressed in seminal vesicles, and the SEMG1 protein is a major component of the semen coagulum . Prostate specific antigen-mediated cleavage of SEMG1 yields functional polypeptides that favor semen liquefaction and enhanced sperm motility, and increased sperm levels of SEMG1 are often associated with asthenospermia . It is tempting to hypothesize that an abundant ectopic secretion of semenogelins into the lumens of serrated colon crypts might favor the formation of a tenacious mucus matrix, which would explain the presence of the adhesive mucus cap that often covers SSLs .
AQP5 (Aquaporin 5) was also confirmed as a very good marker of serrated tumors: completely absent in the normal mucosa and in cADNs, AQP5 transcript is highly or very highly expressed at the bases of all serrated crypts in SSLs and HPs (Figure 3 and Supplementary Figure 8), and patchy, high-level expression was also observed in two of the three TSAs we investigated (TSAs 1 and 3 in Table 1). There was no evidence of AQP5 expression in TSA 2, which was characterized by a more pervasive dysplastic, cADN-like histology than that seen in TSAs 1 and 3. Ectopic expression of this gene in serrated colorectal glands is another example of tumor-associated, phenotypic dys-differentiation. AQP5 encodes a water-channel membrane protein normally expressed in the bronchi, salivary glands, stomach, and testis . AQP5 mutations and polymorphisms are associated with palmoplantar keratoderma  and with outcomes in patients with acute respiratory distress syndrome . Differentiation of alveolar epithelial cells from type II to type I in the lungs is transcriptionally regulated by the p300/beta-catenin complex (but not by the CREB-binding protein/beta-catenin complex), with a concomitant increase in the expression of AQP5 . This finding suggests a functional relationship between APQ5 and Wnt signaling, but evidently not with the canonical Wnt signaling pathway, which is constitutionally active at the base of normal colorectal crypts. The fact that AQP5 is unexpressed in the normal colorectal mucosa and highly expressed at the bottom of serrated crypts (Figure 3 and Supplementary Figure 8) suggests that a variant form of Wnt signaling, likely resembling that reported for alveolar epithelial cells, is active during serrated tumorigenesis. Kleeman et al.  reported that the Wnt signaling activation observed in CRCs arising through the serrated pathway is ligand-dependent, i.e., resulting from mutations in genes encoding RNF43 or RSPOs proteins, which amplify Wnt signal transmembrane transduction. Increased expression of AQP5 mRNA has also been demonstrated in MMR-deficient CRCs arising via the serrated pathway , suggesting that such variant Wnt signaling might be upregulating the expression of this gene in the bases of serrated crypts.
In contrast, Wnt signaling activation is ligand-independent in CRCs arising along the conventional tumorigenic process acting in cADNs, i.e., tumors with mutations in APC or CTNNB1 genes encoding the intracellular signal transduction proteins Adenomatous polyposis coli or β-catenin, respectively. The constitutive activation in cADNs of this canonical Wnt signaling at the base of normal colorectal crypts upregulates the expression of well-known Wnt target genes, such as CMYC, CD44, NKD1 and AXIN2 . NKD1 (naked cuticle homolog 1) encodes a protein that negatively regulates canonical Wnt signaling via mechanisms that are still incompletely understood . For these reasons, we tested NKD1 mRNA expression in this study for its potential to distinguish cADNs (where high levels were expected) from SSLs and HPs (where expression was very low and confined to a few cells at the bases of serrated crypts) (Table 1, Figure 5 and Supplementary Figure 14).
Immunohistochemical staining patterns are sometimes difficult to interpret owing to the low specificity of the available antibodies, and this limitation would have been highly relevant for many of the targets we investigated in this study. The ISH protocol we used involves hybridization of multiple probes that are complementary to the RNA targets, thereby providing highly specific results, and its sensitivity is also high thanks to the use of a series of complementary amplification molecules (see Methods). Unlike immunohistochemistry, ISH also allows visualization within the tissue of noncoding RNAs, such as LINC00520, which we found to be upregulated in serrated lesions using RNA sequencing . Our present findings verify the validity of LINC00520 as a new marker of the serrated pathway: it is moderately to highly expressed in the upper half of the serrated crypts in SSLs and HPs, virtually absent in cADNs, except in a few cells at the mouth of the glands, and expressed at low to moderate levels at the surface of normal crypts (Table 1, Figure 3 and Supplementary Figure 9). Therefore, like ACHE, the LINC00520 gene is normally expressed in the superficial epithelium of normal colorectal crypts, and this expression is markedly upregulated in serrated lesions, where it extends deep into the abnormal crypts. This long noncoding RNA regulates endothelial nitric oxide synthase expression  and may play role in breast tumorigenesis , but its epigenetic regulatory function in the colorectal epithelium is completely unknown.
Two of the serrated-specific targets investigated in this study encode transcription factors, ZIC5 (Zinc finger protein of the cerebellum) and FOXD1 (Forkhead box D1). They are essential for embryonic development of specific tissues  but absent in most adult tissues, including the normal intestinal mucosa. Developmental transcription factors like these are often found to be ectopically re-expressed in specific tumor cells, and this is the case for ZIC5 and FOXD1 in serrated colorectal tumor cells (Table 1, Figure 4 and Supplementary Figures 10 and 11). Their mRNAs were consistently present in the serrated lesions we investigated: FOXD1 labeling was observed along the entire longitudinal axis of serrated crypts, whereas ZIC5 was generally confined to the lower half. ZIC5 and FOXD1 expression levels in serrated lesion were both low, probably because their mRNAs (like those of most transcription factors) are relatively unstable .
Interestingly, ZIC2 and its neighbor, ZIC5, displayed the same staining patterns in serrated lesions (Supplementary Figure 10). Expression of these two genes inhibits the transcriptional activity of beta-catenin/TCF (i.e., the canonical Wnt signaling that occurs in the adult stem cell compartment of the intestinal epithelium), thereby disrupting intestinal epithelial homeostasis . Their re-expression during serrated tumorigenesis once again points to a switch from the canonical Wnt signaling active during conventional adenomatous tumorigenesis to a fundamentally different variant form of this signaling cascade, as previously discussed for AQP5 and NKD1. Indeed, ZIC5/2 expression has been reported in APC-wildtype and MMR-deficient colon cancer cell lines, but levels were almost undetectable in APC-mutant and MMR-proficient lines .
RNA-sequencing data  revealed SSL-specific upregulation of APOBEC1 (Apolipoprotein B mRNA editing catalytic subunit 1) and, as reported by others , MUC5AC (Mucin 5AC) expression (in comparison with normal mucosa). Topographical analysis of the ISH tissue-staining patterns confirmed the RNA-sequencing data, but it also highlighted a risk of error associated with conclusions based exclusively on this type of data. As shown in Table 1, Figures 4 and 5, and Supplementary Figures 12 and 13, neither APOBEC1 nor MUC5AC can be considered a bona fide marker of the serrated pathway: both are more strongly expressed in SSLs than they are in cADNs, but cADNs do consistently express both genes, albeit in more restricted areas of the glands or in a patchy pattern. (These staining patterns explain why random selection of endoscopic biopsies for RNA extraction and sequencing can lead to underestimated expression levels of certain genes in some tumor types.)
The RNA-editing enzyme APOBEC1, which deaminates apolipoprotein B mRNA Cytosine666>Uracil in the small intestine , is moderately expressed in the surface epithelium of the colorectal mucosa (Supplementary Figure 12). Its high-level expression in SSLs extends down into the serrated crypts but stops short of the crypt base. It remains to be seen whether apolipoprotein B editing and/or APOBEC1-mediated DNA mutagenesis (i.e., C>T transitions stemming from unrepaired cytosine deaminations) are increased in these neoplastic crypts . It is interesting to note that C>T transitions at CpG dinucleotides are over-represented in the DNA mutation signature of CRCs displaying MMR-deficiency , which, as discussed above, is caused by CIMP-mediated silencing of MLH1 expression.
As for MUC5AC, its tumor-associated expression represents another example of dysregulated neoplastic cell-fate differentiation. MUC5AC encodes a typical gel-forming glycoprotein found in normal gastric and respiratory tract epithelial cells , and it proved to be an excellent marker of goblet cells in all the tumors we investigated (especially SSLs, which are typically goblet-cell-rich) but not of the goblet cells found in the normal colorectal mucosa (Supplementary Figure 13). The goblet-cell differentiation that occurs in serrated lesions (and in some areas of cADNs and TSAs) thus appears to be epigenetically distinct from that seen in normal colorectal crypts.
Whole-section analysis with ISH or immunohistochemistry facilitates characterization of expression pattern heterogeneity within a tumor, a feature that can be missed with RNA sequencing analysis of random biopsies, as exemplified by our experience with MUC5AC and APOBEC1. When used with reliable antibodies, immunohistochemistry can also identify tumor-specific changes that can escape detection by ISH. A recent example involves AGRN protein expression in the muscularis mucosae of SSLs, which has been shown to distinguish SSLs from HPs despite the fact that the two lesions display similar levels of AGRN mRNA expression levels in the lower half of their serrated crypts  (see also Supplementary Figure 2 and 3: AGRN mRNA expression patterns from  and , respectively).
Nine targets were chosen as putative HP-specific tissue staining markers  (RNA-sequencing data shown in Supplementary Figure 3), but none of the nine were verified as such by our ISH findings. Table 1, Figure 6 and Supplementary Figures 15 and 16 show two examples, the HOXD13 and HOXB13 genes. They belong to two different homeobox gene families of transcription factors that play crucial roles in vertebrate embryonic development . Their expression in adult tissues is restricted to the distal colon and prostate (both genes), the vagina (HOXD13), and the urinary bladder (HOXB13) . ISH confirmed that HOXD13 is expressed in the normal mucosa of the distal colon, the rectum in particular. It is also expressed at crypt bases in HPs, especially those located in the rectum (Table 1), but low expression is detectable also in the other tumor types investigated. Like HOXD13, HOXB13 expression is restricted to the distal colon and rectum (generally at higher levels than HOXD13), and it is expressed in HPs as well as all other tumor types, with levels in distal-colon tumors that are far higher than those in their proximal-colon counterparts. These two genes are more appropriately considered bona fide markers of the normal epithelium of the distal colon and rectum rather than of HPs. As discussed above, the fact that HPs are much more likely to arise in these segments than SSLs probably explains why these genes would appear to be HP-specific on the basis of RNA-sequencing data.
FAM3B, which encodes a signaling protein normally expressed in the endocrine pancreas and gastrointestinal tract , was chosen as a putative negative marker of HPs, i.e., one whose nonexpression is specific to these lesions (Supplementary Figure 3). This assumption was also explained by the staining pattern of the normal mucosa: unlike the previously discussed mRNAs, FAM3B is expressed only in the proximal segments of the normal colon, where HPs are rare. It was also variably expressed in the other lesion types from all colorectal segments, but the highest levels were found in tumor glands of proximal-colon lesions (Figure 6, Supplementary Figure 17).
The search for markers that can clearly distinguish SSLs from HPs is obviously going to be difficult. Kanth et al.  have identified additional candidate markers that were not included in our investigation. Using RT-PCR, they recently assessed the performance of a 7-marker panel that included five of those candidates, as well as SEMG1 and ZIC5/2. The panel differentiated SSLs from HPs with 89% sensitivity and 88% specificity . However, the gene expression differences between these two types of serrated tumors were less significant when distal- rather than proximal-colon SSLs were considered, suggesting that distinguishing between serrated lesions arising in the same colon segment is still likely to be problematic.
One of the obvious limitations of our study is that it was conducted on an undeniably small number of colorectal lesions. Our aim here was to verify biomarker candidates on the basis of our previously reported RNA-sequencing data. Becasue the ISH protocol entails synthesis of branched DNA molecules, it is expensive, at least when done manually, as it was in this study, where the priority was to test a relatively large number of promising markers instead of evaluating a few markers in numerous lesions. However, this technique could easily be used with the automatic robotic instruments routinely used for immunohistochemistry in all pathology laboratories. This would reduce costs considerably and greatly facilitate next-step efforts to provide the reliable validation of the most promising markers in larger series of colorectal tumors representing all histologic types, sizes, and colorectal segments of origin.
Descriptive findings like ours can clearly have impact in the clinics, but they can (and should) also serve as springboards for research into the functional significance during serrated colorectal tumorigenesis of the dysregulated gene expression discussed above. The cascade of molecular events that characterizes this process appears to involve a dramatic epigenetic reprogramming, whose early stages are reflected by the recently described proto-CIMP phenotype  and intriguing forms of aberrant differentiation at the cellular and tissue levels.