Introduction

The differentiation of a primary tumor and a metastatic lesion originating from a distant primary site is crucial in cancer diagnosis as it has tremendous effects on the individual treatment and prognosis. Over the last decades, there have been great advances in this field mainly relying on immunohistochemistry [1]. However, in some cases it is still very difficult or even impossible to distinguish primary and metastatic tumors. This is particularly true for the differentiation of pulmonary enteric adenocarcinoma and metastatic colorectal carcinoma to the lung.

The first description of pulmonary enteric adenocarcinoma dates back to 1991 [2]. However, pulmonary enteric adenocarcinomas were not included in the World Health Organization classification of lung tumors until 2015 due to the difficulty of the diagnosis [3, 4]. By definition, these tumors show an enteric morphology in >50% of the tumor mass. This pattern can consist of glandular, papillary, cribriform or solid growth, lined by tall-columnar tumor cells with luminal necrosis or prominent nuclear debris [5]. Furthermore, pulmonary enteric adenocarcinomas express at least one immunohistochemical marker typical for enteric differentiation (CDX2, CK20 or MUC2). The diagnosis of pulmonary enteric adenocarcinoma is further complicated by the fact that TTF-1 expression is frequently lost in non-small cell lung cancer with mucinous or enteric differentiation. Therefore, CK7 was considered to be a helpful marker to distinguish pulmonary enteric adenocarcinomas from metastatic colorectal cancer [6, 7]. However, there are several reports on CK7 negative pulmonary enteric adenocarcinomas and CK7 expression has been described in up to 17% of colorectal cancers [8,9,10,11]. A recent study determined an intermediate sensitivity (71%) and specificity (82%) for the combinatorial use of CK7 and CDX2 for pulmonary enteric adenocarcinoma identification [12]. Furthermore, apart from metastatic colorectal cancer, metastases from malignancies arising from the upper gastrointestinal tract, such as stomach or esophageal adenocarcinoma, often have to be considered as another differential diagnosis. This further complicates the diagnosis of pulmonary enteric adenocarcinoma, as CK7 expression is even more common in these tumors than in colorectal adenocarcinomas. In most instances, this disqualifies CK7 as a reliable marker to exclude a gastrointestinal primary. In addition, pulmonary enteric adenocarcinomas were reported to frequently lack Surfactant and Napsin A expression, two markers that are commonly used to demonstrate the pulmonary origin of an adenocarcinoma [13]. More recently, SATB2 and β-catenin immunohistochemistry have been described as a putative helpful diagnostic markers in this setting, correctly identifying six of seven investigated pulmonary enteric adenocarcinomas [14]. Apart from immunohistochemistry, pulmonary enteric adenocarcinomas are poorly characterized with only small case studies mainly focusing on single genetic alterations [2, 3, 10, 13, 15,16,17].

As it is currently impossible to distinguish pulmonary enteric adenocarcinomas from metastatic gastrointestinal adenocarcinomas based on conventional histomorphology and immunohistochemistry, gastrointestinal malignancies have to be ruled out by careful clinical and radiological examination. This is a complex, time-consuming, and stressful procedure for the patients and delays the initiation of optimal therapy. Therefore, there is an urgent need to identify reliable methods to differentiate pulmonary enteric adenocarcinomas from metastatic gastrointestinal cancer.

DNA methylation patterns have been shown to be very tissue specific [18, 19]. This led to the recent development of a methylation-based classification of brain tumors and novel algorithms to identify primary sites of cancers of unknown primary [20, 21]. Along these lines, we hypothesized that DNA methylation may also be able to distinguish pulmonary enteric adenocarcinomas from metastatic gastrointestinal cancer. Therefore, we used publicly available methylation data of primary pulmonary, colorectal as well as esophageal and gastric (upper gastrointestinal) adenocarcinomas to develop a machine learning-based algorithm that reliably distinguishes pulmonary enteric adenocarcinomas from metastatic gastrointestinal cancers. Furthermore, we investigated if there is molecular evidence of a separate pulmonary enteric adenocarcinoma methylation class among lung tumors and characterized our cohort for molecular alterations.

Material and methods

Patients and samples

IDAT files containing raw methylation data were obtained from The Cancer Genome Atlas and the Gene Expression Omnibus database [22]. After excluding duplicates and normal tissue samples, The Cancer Genome Atlas dataset consisted of 455 pulmonary, 391 colorectal, and 375 upper gastrointestinal adenocarcinoma specimens. The Gene Expression Omnibus dataset included 27 pulmonary (GSE83842 and GSE94785) and 32 colorectal adenocarcinoma samples (GSE77954, GSE77965, GSE98990 and GSE75546) [23,24,25,26,27].

Reference cohort: All samples were randomly numbered using the sample function from the R base package. The first 200 pulmonary, 200 colorectal, and 200 upper gastrointestinal adenocarcinoma specimens were assigned to a reference cohort (n = 600) that was used for generation of the machine learning model.

Validation cohort: The remaining samples were used as a validation cohort (n = 680) to validate the model generated using the reference cohort.

Test cases: Formalin-fixed paraffin embedded tissue samples were obtained from the archives of the Institute of Pathology of the Charité–University Hospital Berlin. We included six resection specimens and eight biopsy samples that met the morphological, immunohistochemical, and clinical criteria for pulmonary enteric adenocarcinoma. We purposely included biopsy samples as the vast majority of lung malignancies are diagnosed based on biopsy tissue and it is especially challenging to distinguish pulmonary enteric adenocarcinoma from metastatic colorectal cancer in this setting due to the limited amount of tissue available. However, it is crucial to differentiate between primary and metastatic disease prior to surgery as this has direct impact on the surgical technique [28]. The Cancer Genome Atlas dataset included one tumor that met the criteria for pulmonary enteric adenocarcinoma (TCGA-55-A4DF). This sample was also included in the test cases, so a total of 15 pulmonary enteric adenocarcinomas were investigated.

Clinicopathological details regarding the included cases are summarized in Supplementary Table S1. All pulmonary enteric adenocarcinoma patients underwent endoscopic examination as well as computed tomography (CT) of the head, chest, and the abdomen. Six patients (PEAD 1, PEAD 3, PEAD 5, PEAD 8, PEAD 10, and PEAD 11) also had positron emission tomography computed tomography (PET/CT) performed. There were no indications for an extrapulmonary primary tumor. For validation purpose, we also analyzed four cases of pulmonary metastases of colorectal cancers (MCC 1–4) as well as four primary colorectal adenocarcinoma (CRAD 1–4) samples.

Histological reevaluation and immunohistochemistry

To evaluate if the pulmonary enteric adenocarcinoma samples could also be differentiated from metastatic colorectal cancer (“primary pulmonary” vs. “metastatic”) by means of conventional histomorphology and immunohistochemistry, histology was reevaluated by five senior pathologists blinded to clinical information.

Immunohistochemical staining for ALK, CDX2, CK7, CK20, and TTF-1 was performed on the Leica BOND-MAX and Ventana BenchMark XT automated slide stainer according to the manufacturer’s instructions. Antibodies and their according manufacturers as well as concentrations are shown in Supplementary Table S2.

Fluorescence in situ hybridization

Fluorescence in situ hybridization was done according to previously described protocols using the Vysis LSI ALK dual color probe (Abbott Molecular, USA), the MET/CEP7 dual color probe (Abbott Molecular, USA) and the RF POSEIDON ROS1 Break probe (Kreatech, The Netherlands) [29]. For each sample, signals were evaluated in at least 50 non-overlapping tumor cells. Cases were classified as ALK or ROS1 positive if at least 15% of all evaluated tumor cell nuclei displayed break apart signal patterns [30]. Cases with ≥5 MET signals per tumor cell or a MET/CEP7 ratio ≥2 were regarded as MET positive [31, 32].

DNA extraction

Semi-automated DNA extraction from formalin-fixed paraffin embedded samples was performed according the manufacturer’s instructions (Maxwell RSC FFPE Plus DNA Purification Kit, Custom, AX4920, Promega, USA).

Next-generation sequencing

Next-generation sequencing was performed using the Ion AmpliSeq Colon Lung Cancer Research Panel v2 according to manufacturer’s instructions and as described previously [33]. This panel covers 22 relevant genes associated with lung and colorectal cancer (KRAS, EGFR, BRAF, PIK3CA, AKT1, ERBB2, PTEN, NRAS, STK11, MAP2K1, ALK, DDR2, CTNNB1, MET, TP53, SMAD4, FBX7, FGFR3, NOTCH1, ERBB4, FGFR1, and FGFR2). For all samples the optimal amount of 10 ng DNA input was used.

Methylation analysis

The Infinium HD FFPE DNA Restore Kit was used for DNA restoration and methylation analysis was performed using the Illumina Infinium MethylationEPIC BeadChip, each according to protocols supplied by the manufacturer.

Raw and unprocessed methylation data for pulmonary enteric adenocarcinoma and metastatic colorectal cancer samples is available at the Gene Expression Omnibus repository under the accession number GSE116699.

Copy number analysis

Copy number analysis was based on raw methylation array data. Individual genome-wide copy number plots were generated using the conumee package [34]. Copy number plots were screened for focal amplifications and deletions, defined as a negative or positive deviation of the mean line of more than 0.4. The region of interest was visualized using the Integrative Genomics Viewer (Broad Institute, Version 2.4.13) to confirm the aberration and to identify potentially cancer-relevant genes at these loci.

Summary copy number plots showing the rate of chromosomal aberrations over multiple samples were generated using a customized version of the conumee package by D.S. The summary copy number plots for all pulmonary enteric adenocarcinoma specimens were then compared to primary lung adenocarcinoma and colorectal adenocarcinoma summary copy number plots. A chromosomal aberration was considered as characteristic for an entity if it occurred in at least 50% of samples while being present in less than 25% of the comparison entity.

Statistical analysis

Statistical analysis was performed using RStudio version 1.1.444 based on the statistical language R version 3.4.2 [35, 36].

Methylation data was processed using the minfi package based on a workflow suggested by Maksimovic et al. [37, 38]. As a first step we filtered for samples with poor overall quality, defined as a mean detection p value < 0.05. Normalization was done using the functional normalization algorithm (funnorm) [39]. Next, we excluded CpG sites that (a) are located on the sex chromosomes, (b) are associated with single nucleotide polymorphisms (SNPs), (c) have previously been reported as cross-reactive, (d) are not covered by both the Illumina Infinium Human Methylation 450 K BeadChip and the Infinium MethylationEPIC BeadChip or (e) had a detection p value > 0.01 in more than 10% of samples [40]. For further analysis, M-values were generated using the shiftBetas function from the Harman package with shiftBy = 1e−04 to avoid infinite M-values [41]. The 10,000 CpG sites with the highest standard deviation across all samples from the reference cohort were selected for further analysis.

T-distributed stochastic neighbor embedding analyses were generated using the Rtsne package with default parameters, 2000 iterations and a perplexity of 20 [42].

Random forest models were generated using the caret package with standard parameters [43]. The optimal ‘mtry’ value was determined using the trainControl function from the caret package and was set at 140.

The generated random forest classifier was then included in a simple R script that can predict the tissue type (pulmonary, colorectal or upper gastrointestinal) solely based on raw IDAT files as well as a sample annotation sheet. In case of missing values for required CpG sites, the missing data is replaced with the mean M-value across all samples analyzed in this study to minimize any possible confounding effect on the classification.

Results

Pulmonary enteric adenocarcinomas cannot be reliably distinguished from metastatic colorectal cancer by histomorphological and immunohistochemical investigation

All pulmonary enteric adenocarcinoma cases of the series were reinvestigated for immunohistochemical marker expression (Table 1 and Fig. 1). As expected, all pulmonary enteric adenocarcinoma samples (15 of 15, 100%) were positive for CDX2, while eight of 15 tumors (53%) also expressed CK20. CK7 expression was present in eleven of 15 pulmonary enteric adenocarcinomas (73%), while only two of 15 pulmonary enteric adenocarcinomas (13%) showed focal immunoreactivity against TTF-1. Four of four metastatic colorectal cancers (100%) and four of four primary colorectal specimens (100%) expressed CDX2 and CK20, but one of four metastatic colorectal cancer samples (13%) was also positive for CK7.

Table 1 Results from diagnostic immunohistochemistry of pulmonary enteric adenocarcinoma (PEAD), metastatic colorectal cancer (MCC) and primary colorectal adenocarcinoma (CRAD) samples
Full size table
Fig. 1
figure 1

Example of a pulmonary enteric adenocarcinoma with intestinal differentiation on hematoxylin and eosin stain (a). The tumor is negative for TTF-1 (b) and stains positive for CK7 (c) and CDX2 (d). Some cells show faint immunoreactivity against CK20 (e)

Full size image

All pulmonary enteric adenocarcinomas and metastatic colorectal cancer samples were investigated by five senior pathologists that were asked to identify the tissue origin of the tumors based on morphology and immunoprofile. The results from their classification are displayed in Table 2. All pulmonary enteric adenocarcinoma samples were falsely classified as a metastatic lesion by at least one pathologist. Additionally, there were also discordant results regarding the classification of the metastatic colorectal cancer specimens. Particularly, one colorectal metastasis was classified as a tumor with pulmonary origin by all investigators.

Table 2 Results from histological reevaluation. Five senior pathologists were asked to classify 14 pulmonary enteric adenocarcinoma (PEAD) and four metastatic colorectal cancer (MCC) samples as primary pulmonary or metastatic colorectal tumors based on morphology and immunoprofile. For pulmonary enteric adenocarcinoma, there was no broad consent among the investigators regarding the classification
Full size table

Molecular characterization of pulmonary enteric adenocarcinomas

KRAS mutations were the most frequent alterations and were observed in nine of 15 samples (60%), followed by TP53 mutations in five of 15 samples (33%). There were no cases with ALK, ROS1 or MET alterations. Detailed results for molecular characterization are summarized in Table 3.

Table 3 Results from immunohistochemistry (IHC, fluorescence in situ hybridization (FISH), next-generation sequencing and methylation array-based genome-wide copy number analysis. Of note, there was a high rate of KRAS mutations (60%) among the pulmonary enteric adenocarcinoma (PEAD) samples
Full size table

A summary copy number profile of the entire pulmonary enteric adenocarcinoma cohort is displayed in Fig. 2. An overlay of this graph with the summary copy number profiles of primary pulmonary and colorectal adenocarcinomas is depicted in Supplementary Figure S1. In total, pulmonary enteric adenocarcinomas demonstrated high rates of chromosomal aberrations. The most frequent alterations were partial loss of chromosome 3p (66%) and chromosome 1q (53%) as well as gain of chromosome 1p (53%) and chromosome 20q (53%). The detailed regions are listed in Supplementary Table 2. The general copy number profiles of pulmonary enteric adenocarcinomas were reminiscent of primary pulmonary cancers with comparable rates of e.g. gain of chromosome 1p or loss of chromosome 13q. The exception to this was a high frequency (53%) of gain of chromosome 20q that is not typically seen in pulmonary adenocarcinomas (about 20%) but is a frequent event in colorectal cancer (about 65%). Focal amplifications (e.g. ERBB2, MYC) or homozygous deletions (e.g. CDKN2A/B) are also summarized in Table 3. An exemplary copy number plot of one pulmonary enteric adenocarcinoma sample with focal MYC and ERBB2 amplification is displayed in Supplementary Figure S2.

Fig. 2
figure 2

Summary copy number plot for all 15 pulmonary enteric adenocarcinoma samples. The figure shows the frequency of chromosomal aberrations at the respective loci. Gains are displayed above and losses below the baseline, respectively

Full size image

DNA methylation profiling for pulmonary enteric adenocarcinoma classification

Unsupervised t-distributed stochastic neighbor embedding of the whole dataset (n = 1303) revealed three main clusters, representing primary pulmonary, colorectal, and upper gastrointestinal adenocarcinoma samples (Fig. 3). There were some overlaps among the pulmonary and upper gastrointestinal cancer clusters as well as the colorectal and upper gastrointestinal clusters. However, no pulmonary adenocarcinoma cases grouped with the colorectal cancer cluster or vice versa. The 15 pulmonary enteric adenocarcinoma samples reliably grouped with the primary pulmonary adenocarcinomas and the four metastatic colorectal cancer samples clustered with the primary colorectal adenocarcinomas. Both pulmonary enteric adenocarcinomas and colorectal metastases did not form a separate methylation cluster but rather intermixed with primary pulmonary or colorectal samples, respectively. With a proportion of 77% (10 of 14 cases with histomorphological data available), mucinous lung adenocarcinomas were overrepresented among the non-small cell lung cancer cases that accumulated in the upper gastrointestinal adenocarcinoma cluster. The rate of mucinous adenocarcinomas in the main lung cancer cluster was 6% (27 of 454 cases with histomorphological data available), indicating that the mucinous differentiation results in a closer epigenetic relation to upper gastrointestinal adenocarcinomas. Additionally, upper gastrointestinal cancers associated with the Epstein-Barr virus (EBV) formed a separate subgroup (Supplementary Figure S4).

Fig. 3
figure 3

Unsupervised t-distributed stochastic neighbor embedding for all cases (n = 1303). The tissue origin is annotated by color. Distinct clusters representing pulmonary adenocarcinomas, colorectal adenocarcinomas and upper gastrointestinal adenocarcinomas can be observed with small overlaps between the pulmonary and the upper gastrointestinal adenocarcinoma cluster as well as the colorectal and the upper gastrointestinal adenocarcinoma cluster. All pulmonary enteric adenocarcinomas and metastatic colorectal cancers are assigned to the correct clusters

Full size image

We then separated the publicly available cases into a reference (n = 600) and a validation cohort (n = 680). The reference cohort was used to train the machine learning algorithm to identify the correct tumor type. When applied to the validation cohort, the resulting classifier correctly classified all specimens as pulmonary, colorectal or upper gastrointestinal carcinomas. This also included the cases that fell into divergent clusters in the t-distributed stochastic neighbor embedding plot. We then applied this system to the test cases. All 15 pulmonary enteric adenocarcinoma samples were classified as carcinomas originating from the lung and all four metastatic colorectal cancer specimens as colorectal tumors.

The resulting algorithm was integrated in an R Script that is publicly available at https://github.com/aennecken/PEAD. Using raw IDAT files and a sample annotation file, the algorithm can be applied to single or multiple samples. Results are given in the form of an HTML document, including the final prediction and additional information, such as the proportion of votes for the individual diagnosis. Samples are classified as pulmonary, colorectal or upper gastrointestinal tumors. The diagnosis is given if the proportion of votes for one of the three diagnoses is at least 50%. If neither diagnosis reaches 50% of votes, an alternative primary site should be excluded. An example report for the 14 pulmonary enteric adenocarcinoma samples and the four metastatic colorectal cancer samples investigated in this study is available with the online full text version of this paper (Supplementary File S1).

Mean decrease in accuracy was used to assess the importance of the individual CpG sites for random forest decision making. The 100 CpG probes with the highest mean decrease in accuracy are listed in Supplementary Table S4. The most frequent genes that were associated with these CpG probes were CACNB2 (Calcium Voltage-Gated Channel Auxiliary Subunit Beta 2), HOXA9 (Homeobox A9), HOXD1 (Homeobox D1), HOXD8 (Homeobox D8) and RNLS (Renalase, FAD dependent amine oxidase; also refer to Supplementary Figure S5). KRT7 (Cytokeratin 7) was also represented among the 100 most relevant CpG sites.

Discussion

The main focus of this study was to find further evidence of the pulmonary origin of pulmonary enteric adenocarcinoma, to molecularly characterize this non-small cell lung cancer subtype and to generate a methylation-based procedure to reliably distinguish pulmonary enteric adenocarcinomas from pulmonary metastases of colorectal cancer as well as esophageal and gastric adenocarcinomas.

Our histological and immunohistochemical evaluation nicely illustrated the known overlap of pulmonary enteric adenocarcinomas and metastatic colorectal cancers and demonstrated the diagnostic problems in clinical workup of these cases. Indeed, not a single of our pulmonary enteric adenocarcinomas was considered as such by all five investigators. All five investigators are from the same institution, so likely the interpretation may vary even considerably more between different institutions. Interestingly, we also observed that pulmonary metastases of colorectal cancers can potentially mimic primary lung tumors, as one metastatic colorectal cancer sample was classified as a primary pulmonary adenocarcinoma by all pathologists due to heterogenous growth patterns and CK7 expression.

There is little data regarding the mutational profile of pulmonary enteric adenocarcinomas. The few existing case series mainly focused on KRAS, EGFR, and ALK alterations. To get a more comprehensive view on the molecular changes that could characterize pulmonary enteric adenocarcinomas, we used a targeted next-generation sequencing panel covering 22 genes associated with lung and colorectal cancer. In accordance with previous studies, we observed a slightly higher frequency of KRAS mutations compared to common non-small cell lung cancers [10].

We also investigated predictive markers for possible treatment strategies. However, we observed no ALK, MET, EGFR or ROS1 alterations. According to previous studies, EGFR mutations are a rare event in pulmonary enteric adenocarcinomas and there is only one study reporting a single case with ALK rearrangement [3, 14, 16, 44]. Although there is only limited data, alterations eligible for tyrosine kinase inhibitor therapy seem to be less common than in other non-small cell lung cancer subtypes.

Regarding copy number alterations, pulmonary enteric adenocarcinomas showed chromosomal aberrations that are more common in pulmonary than in colorectal adenocarcinomas. Interestingly, the frequency of gains of chromosome 20 in pulmonary enteric adenocarcinoma was more in accordance with typical colorectal cancer alterations. Furthermore, loss of chromosome 3p might be a characteristic event in pulmonary enteric adenocarcinomas.

On the basis of publicly available methylation data from a reference cohort of 600 samples derived from The Cancer Genome Atlas dataset, we were able to train a robust machine learning-based algorithm that successfully classified a validation cohort consisting of 680 pulmonary, colorectal and upper gastrointestinal adenocarcinomas samples without any misclassification. Additionally, all 15 pulmonary enteric adenocarcinoma specimens were classified as primary lung tumors. In conclusion, this study makes headway in proving that pulmonary enteric adenocarcinoma is a rare but actually existing non-small cell lung cancer subtype rather than a metastatic lesion. The fact that the methylation signature is in accordance with common lung adenocarcinomas rules out other explanations for the unusual morphological and immunohistochemical features of this tumor entity, such as a missed or a regressed colorectal cancer.

In theory, the analysis of the methylation profile of metastases might be complicated by several potential confounding factors. On the one hand, samples may be contaminated with adjacent benign tissue of the affected organ; on the other hand, the microenvironment at the metastatic site may induce changes in the methylation profile of the tumor cells that differ from the primary tumor. To address this concern, we also analyzed a set of four pulmonary metastases from colorectal carcinomas. All samples were classified as colorectal tumors by the random forest algorithm, providing evidence that the algorithm is not biased by these potential confounding factors.

In t-distributed stochastic neighbor embedding analysis, there was no evidence for a pulmonary enteric adenocarcinoma or metastatic colorectal cancer subgroup as all samples mixed with other primary pulmonary or colorectal cancers, respectively. However, interestingly, mucinous pulmonary adenocarcinomas accumulated in the upper gastrointestinal cancer cluster, suggesting that these tumors are epigenetically more related to esophageal or gastric adenocarcinomas than other primary lung cancers. Still, these tumors were consistently classified as pulmonary adenocarcinomas by the random forest classifier.

A major limitation of our study is the relatively low number of pulmonary enteric adenocarcinoma samples that were available for analysis. However, this could not be overcome due to the rare incidence of this subtype. When methylation profiling is considered in a diagnostic setting, the relatively high costs of consumables and technical equipment as well as currently not developed technical expertise have to be considered. This currently limits the use of this technology outside of specialized institutes. However, a broader establishment of DNA methylation analysis in the following years seems likely, considering the huge success and dynamics of this method in the classification of brain tumors [20]. Another limitation of this technique is the fact that it is crucial to ensure a high tumor content, as benign tissue (e.g. immune cells, fibroblasts, non-cancerous epithelium) may interfere with the methylation signature. Whereas non-tumorous areas can simply be excluded when evaluating immunohistochemical stainings, some samples might not be suitable for methylation analysis if the amount of tumor cells compared to normal cells is too low. For brain tumors, the dropout because of low tumor cell content was 4% of cases in a prospective experimental diagnostic implementation [20].

However, the methodology used in this study also has some major and revolutionary advances. Array-based and genome-wide methylation analyses deliver reproducible and detailed epigenetic tumor profiles that are relatively resistant to batch effects. Furthermore, the analysis can be performed on formalin-fixed paraffin embedded tissue. In contrast to immunohistochemistry, these tests investigate thousands of loci and genes. In combination with machine learning algorithms, these large datasets can be used to derive highly valuable information which can be used to solve specific diagnostic problems, as shown in this study, or to even classify whole tumor entities [20].

An R script with an example input file that uses the classifier generated in this study to predict the molecular subtype of inputted 450 K or EPIC methylation array data is publicly available at https://github.com/aennecken/PEAD. Although this pulmonary enteric adenocarcinoma classifier is a research tool and proof of principle and should be used with caution, in the future, this algorithm could enable pathologists to diagnose pulmonary enteric adenocarcinoma solely based on tissue analysis with numerous potential benefits for patients and physicians. In theory, the described method could even prevent unnecessary and potentially hazardous examinations and minimize radiation exposure. It could potentially shorten the time that is needed to make a definitive diagnosis, which is crucial to decide on the optimal surgical strategy (wedge or segmental resection vs. lobectomy) or the need of additional systemic therapy. Furthermore, the algorithm could be essential to distinguish metachronous metastasis from a secondary tumor in patients with a history of colorectal cancer. Another advantage of this methodology is the fact that it can also be performed on biopsy samples. The conventional interpretation of biopsy samples might be especially hard because only a small proportion of the tumor can be examined and the tumor cells may be altered due to sampling artifacts.

In summary, this study describes a robust classifier based on DNA methylation data that could in future potentially enable pathologist to make the diagnosis of pulmonary enteric adenocarcinoma on the basis of tissue analysis alone, independently from complex and time-consuming clinical examinations.