Expression profiling and prediction of distant metastases in head and neck squamous cell carcinoma

Abstract

Background

For breast and prostate cancer, a gene expression signature of the tumour is associated with the development of distant metastases. Regarding head and neck squamous cell carcinoma (HNSCC), the only known risk factor is the presence of ⩾3 tumour‐positive lymph nodes.

Aim

To evaluate whether a HNSCC gene expression signature can discriminate between the patients with and without distant metastases.

Methods

Patients with HNSCC with and without distant metastases had >3 tumour‐positive lymph nodes, and did not differ with respect to other risk factors. Statistical analysis was carried out using Student's t test, as well as statistical analysis of microarrays (SAM), to assess the false discovery rate for each gene. These analyses were supplemented with a newly developed method that computed deviations from gaussian‐order statistics (DEGOS). To validate the platform, normal mucosa of the head and neck was included as control.

Results

2963 genes were differently expressed between HNSCC and normal mucosa (t test; p<0.01). More rigorous statistical analysis with SAM confirmed the differential expression of most genes. The comparison of genes in HNSCC with and without metastases showed 150 differently expressed genes (t test; p<0.01), none of which, however, could be confirmed using SAM or DEGOS.

Conclusions

No evidence for a metastasis signature is found, and gene expression profiling of HNSCC has seemingly no value in determining the risk of developing distant metastases. The absence of such a signature can be understood when it is realised that, for HNSCC in contrast with breast cancer, the lymph nodes are a necessary in‐between station for haematogenous spread.

Head and neck squamous cell carcinoma (HNSCC) is the fifth most common type of carcinoma worldwide.¹ Despite improvement in local control, survival has only marginally increased during the past three decades. A major negative factor in this respect is the development of metastatic disease at distant sites. Distant metastases occur in 10–20% of patients with HNSCC²,³,⁴,⁵ and the incidence may increase in the near future.⁶ Although new screening modalities have improved the detection rate of distant metastases at initial evaluation, these metastases continue to emerge during follow‐up in several patients. In all, 50% distant metastases are detected clinically within 9 months of treatment and 80% are detected within 2 years.⁷ The presence of distant metastases has dismal consequences for the patient. Adequate treatment is often not possible and life expectancy is dramatically decreased. It is important to better predict whether a patient will develop distant metastases during follow‐up, as this influences the decision on how to treat the patient for the initial HNSCC. In that case, unnecessary extensive treatment of the primary tumour can be omitted. The only option for this group of patients is effective local treatment in combination with adjuvant systemic treatment. Adjuvant systemic treatment, however, is still in the stage of development at present. To determine whether a patient will develop distant metastases during the course of the disease, an accurate marker that predicts distant metastases is urgently needed. Patients with multiple lymph node metastases, especially >3, have a relative high (up to 50%) risk of developing distant metastases.²,⁸,⁹,¹⁰ Unfortunately, this histopathological feature is still of little value for the individual patient, as <50% of these patients will develop distant metastases.

Over the past few years, gene expression profiling using microarray hybridisation, analysing thousands of genes simultaneously, has provided new insights into carcinogenesis and cancer dissemination. HNSCC has previously been studied with expression arrays by various authors as reviewed by Akervall,¹¹ and in a few reports an association was shown between gene expression changes and clinically relevant variables, such as the presence of lymph node metastasis¹² or patient survival.¹³,¹⁴. Chung et al¹⁴ reported a set of genes proposed to be predictive for lymph node metastases, which showed marked similarities with a gene set that was associated with metastatic disease in breast cancer.¹⁵ Recently, Roepman et al¹⁶ also identified a set of genes associated with the occurrence of lymph node metastases. Although an indication was found that this gene set holds promise of predicting the process of lymphatic metastatic disease, no analysis of survival, which is known to be strongly related to the development of distant metastases, was presented.

In breast cancer, van‘t Veer et al¹⁵ found an expression signature that correlated with the presence of distant metastases and that can possibly be applied to predict the occurrence of such a metastasis, although in a later study¹⁷ it was shown that the classifying gene set is strongly dependent on the way the data are analysed. Ramaswamy et al¹⁸ identified a gene expression signature of adenocarcinoma metastases (breast cancer, prostate cancer, medulloblastoma and large B cell lymphoma), which was present in some primary tumours and could therefore be used to predict which tumours have metastatic potential. On the basis of such studies, it was interpreted that some tumours already harbour a metastatic signature at an early stage, and that this persists in all or most cells in the primary tumour with metastatic potential during their lifetime.¹⁹ The implication is that, in some tumour types, such as breast cancer, cells with metastatic potential are able to directly disseminate from the primary tumour to distant sites, where they progress to overt metastases without previous passage through the lymph nodes.²⁰ For HNSCC, the situation is possibly different. As there is a strong association between the presence of positive lymph nodes and the presence of distant metastases,² it is conceivable that the lymph node is a necessary passing station for the ultimately haematogenously spreading cells.²⁰ We analysed whether gene expression profiling can predict the development of distant metastases in patients with HNSCC with >3 lymph node metastases, a patient group with a 50% risk of developing distant metastases.⁸,⁹,¹⁰ To avoid the influence of confounding factors as much as possible, strict criteria were applied to the selection of patients with HNSCC. So, both groups consisted of patients with >3 tumour‐positive lymph nodes, and, to exclude the possibility that a negative result was due to the choice of platform or analytical methods, a panel of normal mucosal specimens was added.

Participants and methods

Patients and samples

The source of this study was a panel of liquid nitrogen‐stored HNSCC specimens that were collected during the past 10 years. We selected tumours from patients who either developed distant metastases during follow‐up (case group) or who remained disease free for a minimum follow‐up period of 3 years (control group; table 1). Further selection criteria were ⩾4 tumour‐positive lymph nodes, histologically tumour‐negative surgical margins, localisation in the oral cavity, oropharynx, hypopharynx or larynx, and no recurrent disease other than distant metastases. This selection was carried out to obtain a case–control with identical clinical features on a groupwide basis and a similar risk of developing distant metastases. The habits of smoking and drinking alcohol were scored, and, when information was available, tobacco use was calculated in pack‐years (number of packs per day, 25 cigarettes per pack, multiplied by years of active smoking) and alcohol consumption as unit‐years (glasses of beer, wine or liquor per day, multiplied by years of consumption).

Table 1 Patient and tumour characteristics.

Patient code	Carcinoma				Patient
	Localisation	TNM	Extra‐capsular spread	No of tumour‐positive lymph nodes	Age at diagnosis (years)	Sex	Follow‐up	Smoking		Alcohol drinking
								Yes/no	Pack‐years	Yes/no	Unit‐years
M1	Oral cavity	T3N2b	Yes	5	71	Male	DM lung, DOD	Yes	25	Yes	400
M2	Oral cavity	T3N2c	Yes	11	67	Female	DM liver, DOD	Yes	Unknown	Yes	unknown
M3	Larynx	T4N3	Yes	27	52	Male	DM bone, DOD	Yes	47	Yes	35
M4	Oral cavity	T4N2b	Yes	7	65	Male	DM lung, DOD	Yes	Unknown	Yes	Unknown
M5	Oropharynx	T3N2b	Yes	5	52	Female	DM lung, DOD	Yes	35	Yes	280
M6	Oral cavity	T2N2b	No	7	48	Male	DM pericardium, DOD	Yes	30	Yes	40
M7	Oral cavity	T4N2b	No	5	74	Male	DM bone, DOD	Yes	37	No	0
M8	Oral cavity	T4N2c	Yes	5	58	Male	DM liver, DOD	Yes	50	Yes	400
M9	Oral cavity	T4N2c	Yes	4	65	Male	DM lung, bone, DOD	Yes	Unknown	Yes	Unknown
M10	Hypopharynx	T1N3	Yes	5	54	Male	DM bone, DOD	No	0	No	0
M11	Oropharynx	T2N2b	Yes	4	55	Female	DM bone, liver, DOD	Yes	30	Yes	Unknown
NM1	Larynx	T4N2b	Yes	6	48	Female	df 68 months	Yes	Unknown	Yes	Unknown
NM2	Oropharynx	T2N2b	Yes	8	48	Male	df 44 months	Yes	23	Yes	250
NM3	Oral cavity	T2N2b	Yes	6	54	Male	df 59 months	Yes	48	Yes	560
NM4	Oropharynx	T2N2b	Yes	4	52	Male	df 67 months	Yes	25	Yes	136
NM5	Oropharynx	T4N2b	Yes	5	73	Male	df 110 months	Yes	26	No	0
NM6	Hypopharynx	T3N2c	Yes	12	41	Male	df 56 months	Yes	10	Yes	150
NM7	Oral cavity	T4N2c	Yes	6	49	Female	df 58 months	Yes	28	Yes	84
NM8	Oropharynx	T2N2c	No	10	57	Female	df 47 months	Yes	40	Yes	400

df, disease free; DM, distant metastases; DOD, death from disease; HNSCC, head and neck squamous cell carcinoma; M, metastasised HNSCC; NM, non‐metastasised HNSCC; TNM, tumour‐node‐metastasis.

Several patients with tumours were not eligible because of lower risk nodal status (<4 positive lymph nodes), non‐surgical treatment, tumour‐positive surgical margins, distant metastases from other primary tumours or a too short disease‐free follow‐up period. Of the total 424 frozen tumours available for analysis, 11 were found eligible for the case group and eight for the control group; also eight normal mucosa specimens, obtained from the uvula of healthy controls without cancer were added for comparison and validation of the platform and analysis.

The Institutional Review Board of the VU University Medical Center, Amsterdam, The Netherlands, approved the study protocol, and written informed consent was obtained from all the patients.

RNA preparation

All samples were snap frozen in liquid nitrogen and stored at −80°C. Tumour percentage was estimated on 5‐μm‐thick sections stained with haematoxylin and eosin, with a mean tumour percentage of 62% in the case group (range 20–90%) and 60% in the control group (range 10–90%). In all, 15–20 30‐μm‐thick frozen sections were prepared with a cryomicrotome and carefully transferred to a chilled 1.5‐ml tube containing RNAbee (Campro Scientific, Veenendaal, The Netherlands) for intact RNA isolation according to the manufacturer's protocol. Quality control of total RNA samples was carried out with the RNA 6000 Pico LabChip kit (Agilent Technologies, Palo Alto, California, USA) and analysed on the Agilent 2100 Bioanalyzer. As a common reference for array hybridisation, the Universal Human Reference RNA from Stratagene (La Jolla, California, USA) was used.

Synthesis and labelling of cDNA

Because of possible non‐linear amplification of small amounts of RNA, we used non‐amplified total RNA for hybridisation on the arrays. The amounts of total RNA varied from 5 to 15 μg, depending on the size of the sample. Details of the preparation of labelled cDNA are provided in a previously published protocol.²¹ Cell samples were labelled with Cy3 (Fluorolink Cy3 Monofunctional Dye; Amersham, Freiburg, Germany) or Cy5 (Fluorolink Cy5; Amersham) for common reference.

Array hybridisation and scanning

The Human Release V.1.0 oligonucleotide library, containing 18 861 60‐mer oligonucleotides representing 17 260 unique genes as designed by Compugen (San José, California, USA) was obtained from Sigma‐Genosys (Zwijndrecht, The Netherlands). Hybridisation was carried out as previously described.²¹ Spots were quantified by Imagene V.5.5.4 software (Biodiscovery, Marina del Rey, California, USA), using the default settings. Local background was subtracted to obtain the signal mean. The expression platform we used has been described previously in detail,²¹ and a good correlation between array and Taqman results was obtained for several genes regarding the level of expression intensities.

Analysis

All expression intensities were transformed to log₂ values and intensities <0 (below background) were classified as missing. Data were normalised by means of z score transformation.²² The expression intensity of each sample was calculated by subtracting the values of the Cy5 channel (reference) from those of the Cy3 channel (sample), yielding the Cy3:Cy5 ratio. The number of missing values varied per sample and had a mean of 13% of the values. To find potential classifying genes, differences in gene expression of HNSCCs with and without distant metastases were calculated with Student's t test (SPSS for Microsoft Windows). Only those genes with values of ⩾5 carcinomas per group were analysed. To confirm the findings, we additionally applied statistical analysis of microarrays (SAM) software V.1.21.²³ A q value (%) and a false discovery rate value (%) was obtained for each gene. To validate our platform and analytical tools against existing data, we also compared the expression profiles of all HNSCC samples versus the normal mucosal samples.

A second analysis was carried out on the dataset according to the principles of ordered statistics. For each gene, the computed t value was converted to its exact gaussian analogue (z value) via the p value. All these z values were ranked and compared with the theoretical values from the gaussian distribution. A robust linear regression on the central 90% of the values gave rise to a tolerance region (with α = 1/n, where n is the number of genes considered). Genes were considered confirmed if they lay below the region for negative values of z and above for positive values of z. The analysis is based on deviations from gaussian‐order statistics (DEGOS). Details on DEGOS are available on request.

Annotation analysis was carried out with software available at http://source.stanford.edu. Cluster analysis of the latter comparison was carried out with the software program Spotfire DecisionSite (Spotfire, Somerville, MA, USA). Parameter settings were standard, with no filter or data adjustment, and the hierarchical unsupervised clustering was executed for genes and samples with Pearson's correlation and complete linkage selected.

Results

Characteristics of the study population

The average age in the case group (eight men and three women) was 60 (range 48–74) years, whereas the average age in the control group (five men and three women) was 53 (range 41–73) years. The groups did not differ with respect to tobacco and alcohol consumption, number of positive nodes and presence of extracapsular spread. Table 1 gives further details.

Expression profiles in HNSCC and normal mucosal samples

Nineteen HNSCC samples, with or without metastatic disease, and eight normal mucosa samples were compared with respect to the RNA expression profile. With unsupervised hierarchical clustering of all 17 237 genes in the 27 samples, two major groups were separated, the carcinomas and the normal mucosal samples (fig 1). We found 17.2% (2963/17 237) genes with a significantly different expression (p<0.01); 1063 of these were upregulated and 1900 downregulated in HNSCC. Table 2 shows the relationship between a certain cut‐off of the p value and the consequences for the number of different genes. Table 3 gives the list of the 50 genes showing the largest difference between these groups. A more extensive list of all differentially expressed genes can be found at http://www.jclinpath.com/supplemental.

Figure 1 Hierarchical clustering of 19 head and neck squamous cell carcinoma (HNSCC) specimens and eight normal mucosas of healthy controls. The dendrogram is based on the similarity of the 17 237 genes (left) and the 27 tissue samples (above). The clustering method involved complete linkage and correlation as the similarity measure. Empty values were replaced by the column average. The eight normal mucosas cluster as a separate group, as visualised at the right. non m, non‐metastasised HNSCC; meta, metastasised HNSCC.

Table 2 Relation between p value and number of differentially expressed genes.

Upper cut‐off level p value	Differently expressed genes (n)
	Increased expression	Decreased expression
HNSCC v normal mucosa
0.01	1063	1900
0.001	563	931
<0.001	310	483
Metastasised v non‐metastasised HNSCC
0.01	82	68
0.001	7	6
<0.001	0	0

HNSCC, head and neck squamous cell carcinoma.

Table 3 Fifty most significantly different genes between normal mucosa and head and neck squamous cell carcinoma.

Rank‐ number	HUGO‐ identifier	GenBank accession number	Description	Number		p‐ Value Student t	SAM
				HNSCC	Normal mucosa		q Value (%)	Local fdr (%)
Increased expression in HNSCC
1	LOC492304	AK025719	Putative insulin‐like growth factor II‐associated protein	19	7	1	0	0.26
2	KIAA0261	D87450	KIAA0261	19	8	1	0	0
3	HEIR1	D28449	Inhibitor of DNA binding 3	18	8	1	0	0
4	C20orf3	AB033767	Chromosome 20 open reading frame 3	19	8	1	0	0
5	Unknown	NM_005332	Haemoglobin ζ	19	8	1	0	0.23
6	DDOST	NM_005216	Dolichyl‐diphospho‐oligosaccharide‐protein glycosyltransferase	19	8	1	0	0
7	HBG2	NM_000184	Haemoglobin γ‐G	17	8	1	0	0.32
8	MSN	NM_002444	Moesin	19	8	1	0	0
9	KIAA1922	AF119868	KIAA1922 protein	19	8	1	0	0
10	OPN	NM_000582	Secreted phosphoprotein 1 (osteopontin)	19	6	1	0	0
11	LGALS1	NM_002305	Lectin, galactoside‐binding, soluble, 1 (galectin 1)	18	8	1	0	0
12	Unknown	NM_004052	Data not found	17	8	1	0	0
13	IBP3	NM_000598	Insulin‐like growth factor binding protein 3	19	8	1	0	0
14	KIAA0092	NM_014679	Translokin	17	8	1	0	0
15	MFHAS1	NM_004225	Malignant fibrous histiocytoma amplified sequence 1	19	8	1	0.09	1.55
16	PIT1	NM_005415	Solute carrier family 20 (phosphate transporter)	19	8	1	0	0
17	HBGA	M11427	Haemoglobin γ‐A	16	8	1	0	0.16
18	FNDC3B	AL157482	Fibronectin type III domain containing 3B	19	7	1	0	0
19	NSUN5	NM_018044	NOL1/NOP2/Sun domain family, member 5	19	8	1	0.09	1.91
20	IFI15	NM_005101	Interferon, α‐inducible protein (clone IFI‐15K)	19	8	1	0	0
21	H2B/H	NM_003523	Histone 1, H2be	19	8	1	0	0
22	Unknown	Z36811	Data not found	19	8	1	0	0
23	MCM6	NM_005915	MCM6 minichromosome maintenance	19	8	1	0	0
24	BST2	NM_004335	Bone marrow stromal cell antigen 2	19	8	1	0	0
25	Unknown	K02847	Data not found	19	8	1	0	0
Decreased expression in HNSCC
1	THW	AF317550	PERP, TP53 apoptosis effector	19	8	<0.001	0	0.06
2	Unknown	AK000006	Data not found	19	8	<0.001	0	0
3	C1orf10	NM_016190	Chromosome 1 open reading frame 10	19	8	<0.001	0	0
4	PMI1	NM_002435	Mannose phosphate isomerase	18	8	<0.001	0	0.06
5	CAGA	NM_002964	S100 calcium binding protein A8 (calgranulin A)	18	8	<0.001	0	0.01
6	ECM1	NM_004425	Extracellular matrix protein 1	19	8	<0.001	0	0.10
7	CL‐20	NM_001423	Epithelial membrane protein 1	17	8	<0.001	0	0
8	SERPINB2	NM_002575	Serine (or cysteine) proteinase inhibitor	19	8	<0.001	0	0.11
9	HOP	AB019573	Homeodomain‐only protein	18	6	<0.001	0	0
10	CLCA4	NM_012128	Chloride channel, calcium activated, family member 4	14	6	<0.001	0	0
11	BICD1	NM_001714	Bicaudal D homologue 1 (Drosophila)	18	8	<0.001	0	0.10
12	DKK1	NM_012242	Dickkopf homologue 1 (Xenopus laevis)	19	8	<0.001	0	0
13	SPRR1B	NM_003125	Small proline‐rich protein 1B (cornifin)	14	8	<0.001	0	0.01
14	NAGK	NM_017567	N‐acetylglucosamine kinase	19	8	<0.001	0	0
15	VAV3	NM_006113	Vav 3 oncogene	18	7	<0.001	0	0.09
16	DDX32	NM_018180	DEAH (Asp–Glu–Ala–His) box polypeptide 32	19	8	<0.001	0	0
17	BENE	U17077	BENE protein	19	8	<0.001	0	0.07
18	ZDHHC1	U90653	Zinc finger, DHHC domain containing 1	19	8	<0.001	0	0
19	ANXA1	NM_000700	Annexin A1	19	8	<0.001	0	0
20	Unknown	NM_015961	Data not found	18	8	<0.001	0	0
21	PAFAH	NM_005084	Phospholipase A2, group VII	19	8	<0.001	0	0
22	DAL1	NM_012307	Erythrocyte membrane protein band 4.1‐like 3	19	8	<0.001	0	0
23	M/NEI	M93056	Serine (or cysteine) proteinase inhibitor	17	8	<0.001	0	0
24	PADI1	AB033768	Peptidyl arginine deiminase, type I	19	8	<0.001	0	0.10
25	SCEL	NM_003843	Sciellin	17	8	<0.001	0	0.08

Nineteen HNSCC (11 with and eight without metastases) specimens were compared with 8 normal mucosa specimens using two‐sided Student's t test (a p value of 1 reflects a value between 0.999 and 1). Regarding SAM, the q value is shown, which is comparable to the p value of the t test. Local fdr is the false discovery rate in percentage. Further details on SAM can be found in the article by Cheadle et al²².

fdr, false discovery rate; HNSCC, head and neck squamous cell carcinoma; SAM, statistical analysis of microarrays.

SAM was carried out to assess the chance of a falsely positive gene identification. Table 3 shows details of the analysis of the 50 most different genes. In addition, of the top 100 downregulated genes, all had q value (comparable to the standard p value) <0.11%, and false discovery rate <0.1%. Two of the top 100 upregulated genes had a q value >2.5%, and 23 had a false discovery rate >5% (which was <15% for 22 of these).

Expression profiles in HNSCC with and without distant metastases

The expression profiles of 11 HNSCC specimens with and eight HNSCC specimens without distant metastatic disease were compared. When designing this study we took care to exclude all possible confounding factors. So the groups with and without distant metastases were at a similar group level with respect to sex, age, tumour‐node‐metastases stage, the number of tumour‐positive lymph nodes, and the smoking and alcohol drinking behaviour of the patients (table 1).

It was not possible to discriminate the metastasising tumours from the non‐metastasising ones by unsupervised clustering of all genes. Nevertheless, we found 150 of the 17 240 (0.8%; p<0.01) genes differently expressed on comparing profiles of carcinomas with and without distant metastases (additional information is available at http://www.jclinpath.com/supplemental); 82 genes showed a lower and 68 showed a higher expression in metastasised tumours. With lower p values, fewer genes were found to be different (table 2). Table 4 shows the gene sets with the most differential expression. We carried out additional analyses to exclude the possibility that a gene could be different simply as a result of chance. More rigorous testing was carried out using SAM, which resulted in finding no gene that fulfilled the program's criteria of being significantly different. q and false discovery rate values were >80%. This suggests the rejection of all genes uncovered by Student's t test owing to the high likelihood of being false positive. Also, another statistical approach (DEGOS) did not support the notion that the t test revealed truly different genes.

Table 4 Fifty most significantly different genes between metastasised and non‐metastasised head and neck squamous cell carcinoma.

Rank number	HUGO identifier	GenBank accession number	Description	HNSCC (n)		p Value Student‘s t
				Non‐meta	Meta
Downregulated in metastasised HNSCC
1	PCDH9	AF085861	Protocadherin 9	6	7	<0.001
2	GALR1	NM_001480	Galanin receptor 1	8	9	<0.001
3	COL4A4	NM_000092	Collagen, type IV, α4	8	10	<0.001
4	Unknown	U18909	Data not found	8	10	<0.001
5	TRIM62	NM_018207	Tripartite motif‐containing 62	8	9	<0.001
6	RERE	AK024214	Arginine‐glutamic acid dipeptide (RE) repeats	6	8	<0.001
7	Unknown	AF090927	Data not found	8	9	<0.001
8	SULT1A1	NM_001055	Sulfotransferase family, cytosolic, 1A, phenol‐preferring	8	11	<0.001
9	AGTPBP1	NM_015239	ATP/GTP binding protein 1	6	11	<0.001
10	CHST5	NM_012126	Carbohydrate (N‐acetylglucosamine 6‐O) sulfotransferase 5	8	10	<0.001
11	PDPK1	NM_002613	3‐Phosphoinositide dependent protein kinase‐1	6	8	<0.001
12	Unknown	AL137637	mRNA; cDNA DKFZp434J035 (from clone DKFZp434J035)	5	8	<0.001
13	TRAF5	AB000509	TNF receptor‐associated factor 5	8	10	<0.001
14	Unknown	NM_014097	Data not found	8	11
15	Unknown	AK025573	Data not found	5	6	0.001
16	IDUA	NM_000203	Iduronidase, α‐L‐	8	10	0.001
17	TLR4	NM_003266	Toll‐like receptor 4	8	11	0.001
18	MUC3A	M55406	Mucin 3A, intestinal	5	11	0.001
19	Unknown	U48728	Data not found	6	11	0.001
20	Unknown	NM_014684	Data not found	8	11	0.001
21	DLL1	NM_005618	δ‐like 1 (Drosophila)	6	11	0.002
22	CCR9	NM_006641	Chemokine (C–C motif) receptor 9	7	10	0.002
23	PLCE1	NM_016341	Phospholipase C, epsilon 1	8	10	0.002
24	SEC22L1	AK023270	SEC22 vesicle trafficking protein‐like 1 (S cerevisiae)	7	8	0.002
25	Unknown	AF136408	Data not found	7	8	0.002
Upregulated in metastasised HNSCC
1	DFNA5	NM_004403	Deafness, autosomal dominant 5	7	11	1
2	MAP1B	NM_005909	Microtubule‐associated protein 1B	6	11	1
3	PRNP	X82545	Prion protein (p27‐30) (Creutzfeld‐Jakob disease)	8	11	1
4	RAB6A	AL049984	RAB6A, member RAS oncogene family	6	6	1
5	VGCNL1	AK002089	Voltage gated channel like 1	8	10	1
6	Unknown	AK022068	CDNA FLJ12006 fis, clone HEMBB1001585	5	11	1
7	GARP	NM_005512	Leucine rich repeat containing 32	7	11	1
8	FLJ20313	NM_017762	Myotubularin‐related protein 10	8	11	1
9	FLJ20397	NM_017802	Hypothetical protein FLJ20397	8	10	0.999
10	MEFV	NM_000243	Mediterranean fever	8	11	0.999
11	PMAIP1	D90070	Phorbol‐12‐myristate‐13‐acetate‐induced protein 1	8	11	0.999
12	UBE2I	NM_003345	Ubiquitin‐conjugating enzyme E2I (UBC9 homolog, yeast)	8	11	0.999
13	RPS2	NM_016281	Ribosomal protein S2	8	11	0.999
14	DPYSL3	NM_001387	Dihydropyrimidinase‐like 3	7	11	0.999
15	SNX5	NM_014426	Sorting nexin 5	8	11	0.999
16	FLJ12666	AK022728	Chromosome 1 open reading frame 108	8	11	0.999
17	FLJ34870	AK022384	FLJ34870 protein	7	5	0.999
18	PCDH7	NM_002589	BH‐protocadherin (brain‐heart)	8	11	0.999
19	KIAA1036	NM_014909	KIAA1036	7	9	0.999
20	GNA12	NM_007353	Guanine nucleotide binding protein (G protein) α 12	8	11	0.999
21	FTSJ1	NM_012280	FtsJ homologue 1 (E coli)	8	10	0.999
22	SAV1	AK023071	Salvador homolog 1 (Drosophila)	8	11	0.999
23	RANBP17	L08438	RAN binding protein 17	5	9	0.999
24	UBA2	NM_005499	SUMO‐1 activating enzyme subunit 2	8	11	0.999
25	SERPINA1	M26123	Serine (or cysteine) proteinase inhibitor	8	11	0.999

In all, 11 HNSCC specimens with and eight without metastases were compared using two‐sided Student's t test (a p value of 1 reflects a value between 0.999 and 1). None of the genes could be confirmed using SAM.²²

HNSCC, head and neck squamous cell carcinoma; SAM, statistical analysis of microarrays.

Discussion

HNSCC shows heterogeneity with respect to metastatic behaviour, and the current set of clinical markers is not sufficiently accurate to predict which patient is most at risk of distant metastases. This study was designed to find a set of genes that was differently expressed between metastasising and non‐metastasising HNSCCs. Patients were carefully selected for the presence of >3 lymph node metastases, a factor associated with a relatively high risk of distant metastases. When lymph node metastases in the neck are diagnosed, the chance of survival is halved. We also initially tried to select a case and a control group without lymph node metastases, but we could not find a single case in 424 tumours, strongly supporting the importance of the lymph node compartment for metastasising HNSCC.

To exclude that our platform or our analytical methods might influence the outcome, we compared the expression profile in normal and tumour tissue and discovered many genes to be differently expressed. We have purposely chosen normal tissue from patients without cancer as tumour‐adjacent normal tissue bears the risk of being genetically aberrant.²⁴ Several genes were found to be different between cancerous and normal tissue, and this list includes those genes associated with signal transduction, cell structure, cell cycle, transcription, cell–cell adhesion, cell–matrix interaction and apoptosis. Other reports also found differentially expressed genes, although with lower numbers.²⁵,²⁶,²⁷,²⁸,²⁹,³⁰ Some highly different genes are shared between these reports (eg MSN, SCEL, SPARC, collagens and cytokeratins), and, in addition, similar cellular processes have been reported to be associated. This present set of genes with differential expression has a relatively strong effect, as unsupervised clustering of all available genes (filtered for expression level) generated a dendrogram that separated out the normal tissues and the carcinomas. With more stringent methods such as SAM and DEGOS, most of the highly differentially expressed genes could be confirmed. Thus, it is possible with the presently used expression platform and analysis to generate relevant information, in line with previously published data.

We have found a panel of 150 genes that had a differential expression between tumours with >3 lymph node metastases that either did or did not give rise to distant metastases (p<0.01 with Student's t test), but this test is not optimal as it does not exclude the possibility that these differences have occurred by chance. This number of 150 differentially expressed genes is roughly what would be expected, if randomly distributed values for each gene were assumed. We identified no relevant gene using SAM, an established and rather stringent method that uses permutations to increase the power of significance analysis. Either a gene was not different or there was a high likelihood of it being falsely different. A second method, DEGOS, which was recently developed in our centre to overcome some intrinsic analysis problems with SAM, also did not provide evidence that the genes were unambiguously differentially expressed. Our study does not provide evidence for a metastatic signature, and indicates that expression profiling has seemingly no additional value in predicting the development of distant metastases. The current method to assess the risk of distant metastases, the examination of lymph nodes for the presence of cancer, unfortunately cannot be improved.

Importantly, relatively small patient groups have been used in our study. Nevertheless, when comparing the eight normal mucosas with eight randomly chosen HNSCC specimens, SAM was able to find most of the genes that were also different when data of all 19 HNSCCs were used. This indicates that a large difference in expression is hardly influenced by the number of specimens of available in this study. Also when studying breast cancer, we were able to find a metastatic signature, even when only seven metastatic tumours were compared with seven non‐metastatic carcinomas.³¹ Our results are in line with those of Cromer et al²⁶ who studied 15 metastasising and 11 non‐metastasising hypopharyngeal carcinomas. Although these authors report a signature of 164 differentially expressed genes, they also concluded that it was too early to state whether a useful signature exists. There is the possibility that analysis of much larger numbers of patients with HNSCC will lead to the discovery of a distant metastases expression signature. Its existence will probably be based on several genes that show a small difference in expression between metastasising and non‐metastasising HNSCC, or it will be valid only on a subgroup of HNSCC. If a signature with such characteristics exists, it is doubtful whether it will have much value for the individual patient.

We have studied tumours from various locations in the head and neck area, and there was a relative over‐representation of tumours of the oral cavity in the metastasising group of tumours. It is unclear at this moment whether the expression profile of HNSCCs differs between subsites and whether this influenced the outcome of our study. The numbers of tumours analysed in this study are too small to correct for this potential bias and for the fact that there was no subsite categorisation of the HNSCC after global unsupervised clustering using the information of all genes (fig 1).

In conclusion, we have used microarray expression analysis and explored its potential for diagnostic purposes in HNSCC. No evidence for a distant metastasis signature was found, indicating that expression profiling has seemingly no value in predicting the development of distant metastases. HNSCC may differ in this respect from other tumour types such as breast and prostate cancer. The fact that lymph nodes are a necessary in‐between station for haematogenous spread may explain the absence of a distant metastasis‐associated HNSCC expression profile.

Abbreviations

DEGOS - deviations from gaussian‐order statistics

HNSCC - head and neck squamous cell carcinoma

SAM - statistical analysis of microarrays