Gonadotroph tumours with a low SF-1 labelling index are more likely to recur and are associated with enrichment of the PI3K-AKT pathway

Abstract

Aims

The gonadotroph tumour (GT) is the most frequently resected pituitary neuroendocrine tumour. Whilst many symptomatic GT are successfully resected, some recur. We sought to identify histological biomarkers that may predict recurrence and explore biological mechanisms that explain this difference in behaviour.

Methods

SF-1 immunohistochemistry of 51 GT, a subset belonging to a longitudinal prospective cohort study (n=25), were reviewed. Four groups were defined: Group 1- recently diagnosed GT (n=20), Group 2- non-recurrent GT with long-term follow up (n=11), Group 3- initial resections of GT that recur (n=7), and Group 4- recurrent GT (n=13). The percentage of SF-1 immunolabeling in the lowest staining fields (SF-1 labelling index (SLI)) was assessed and RNA sequencing was performed on 5 GT with SLI<80% and 5 GT with SLI>80%.

Results

Diffuse, strong SF-1 immunolabelling was the most frequent pattern in Groups 1/2, whereas patchy SF-1 staining predominated in Groups 3/4. There was a lower median SLI in Groups 3/4 than 1/2. Overall, GT with SLI<80% recurred earlier than GT with SLI>80%. Differential expression analysis identified 89 statistically significant differentially expressed genes (FDR<0.05) including over-expression of pituitary stem cell genes (SOX2, GFRA3) and various oncogenes (e.g. BCL2, ERRB4) in patchy SF-1 GT. Gene set enrichment analysis identified significant enrichment of genes involved in the PI3K-AKT pathway.

Conclusions

We speculate that patchy SF-1 labelling in GT reflects intratumoural heterogeneity and are less differentiated tumours than diffusely staining GT. SF-1 immunolabeling patterns may have prognostic significance in GT, but confirmatory studies are needed for further validation.

Introduction

Although the historical designation of “pituitary adenoma” implied a benign behaviour to these neoplasms, it has long been recognized that some tumours cause considerable morbidity to patients, particularly through complications related to local invasion into adjacent anatomic structures [1–5]. The recent terminology of pituitary neuroendocrine tumours (PitNET) recognizes this spectrum of biological potential, however, there remains a lack of reliable histological biomarkers to predict aggressive behaviour. Some experts consider the most consistent biomarkers for predicting recurrence are the presence of tumour invasion into surrounding structures and the extent of debulking [6, 7]. However, tissue biomarkers alone, such as the Ki-67 proliferative index and p53 immunoreactivity are controversial in predicting clinical outcome [8, 9]. After decades of application of these two immunostains in PitNET, they have not proved reliable in predicting recurrence and inevitably led to the recent removal of the term ‘atypical adenoma’ in the World Health Organization (WHO) classification scheme [10].

The WHO recommended immunohistochemical assessment of the lineage specific transcription factors (steroidogenic factor-1 (SF-1), pituitary-specific POU-class homeodomain transcription factor (Pit1), and T-box family member TBX19 (TPit)) for the routine diagnostic characterization of PitNET [11, 12]. Prior to this, the classification of the tumours was reliant upon the hormonal profile and, sometimes, on the ultrastructural findings of the tumour [13, 14]. Hormone-negative adenomas were diagnosed frequently in the past and were inferred to be null cell adenomas. However recent studies, have repeatedly shown that many of these hormone negative tumours are in fact gonadotroph tumours (GT), with up to 95% of these labelling with SF-1 [15, 16]. Furthermore, the improved sensitivity of SF-1 immunostaining over follicle-stimulating hormone (FSH), luteinizing hormone (LH) and alpha-subunit immunostains has been demonstrated in a large series of over 1000 PitNETs [16] and the application of transcription factors in classifying PitNET has therefore led to improved characterization of these neoplasms [14, 17]. One possible explanation for the poor reproducibility of predictive biomarkers for recurrence, such as Ki-67, in non-functioning PitNET may have been because this was considered as one grouped category and was not further subdivided by tumour lineage. Careful histological subtyping of PitNET has already shown prognostic significance, such as segregating sparsely from densely granulated somatotroph tumours and recognizing silent corticotroph tumours within the non-functioning tumour category [18–20]. More precise characterization may therefore allow for better informed therapeutic options for patients.

The GT represents the most common PitNET diagnosed in surgical pathology and are of SF-1 lineage. The SF-1 transcription factor is expressed in the hypothalamus, adenohypophysis, adrenal cortex and gonads and is critical in the development of steroidogenic tissues [21, 22]. In the context of adrenocortical carcinoma, SF-1 has already proven useful in both diagnosing adrenocortical tumours and in predicting stage-independent prognostic outcome by assessment of the strength of immunolabeling, with increased SF-1 being associated with poorer prognosis [23–26]. In this study we show that although diffuse expression of SF-1 is seen in the majority of GT, there are a subset of tumours that exhibit patchy, weak-to-negative staining within areas of the tumour and this pattern is associated with recurrence. We then demonstrated through RNA sequencing (RNAseq) of GT with diffuse versus patchy SF-1 labelling that there are distinct transcriptomic differences between these two GT types. Patchy SF-1 labelling may therefore be a reliable histological biomarker of aggressive behaviour possibly by inferring less differentiation of the tumour.

Materials and Methods

A total of 51 GT from 46 patients received at Columbia University Irving Medical Center were included in this study. The study was approved by the Institutional Review Board (IRB) of Columbia University Irving Medical Center. The majority of subjects (n=25) gave written informed consent before participation and the remainder were part of a retrospective cohort study (n=21) approved for inclusion by the IRB.

Case selection

Tumours were designated as GT based on overwhelming predominance of immunopositivity for SF-1 and/or GATA-3 over the other transcription factors (Pit1 and TPit). Four groups of GT were defined. Group 1 comprised consecutive GT that were received between May 2018 until March 2019 (n=20). Group 2 had GT that on prospective follow up with magnetic resonance imaging (MRI) over 1.3–7.0 years, were shown not to have significant radiological growth. Group 3 were the initial resections of GT that subsequently recurred and Group 4 were resections of a recurrent GT. No GT had features of apoplexy.

Recurrence was defined as radiologically significant regrowth of tumour between the first surgery and the time of second surgery. Regrowth occurred of radiographically visible residual tumour (n=6) or no visible tumour (n=6) on initial postoperative imaging. Tumour regrowth ranged from 4 to 40 mm and occurred over 2 to 25 years after first surgery. Growth was considered clinically significant if the tumour became more proximal to the optic chiasm or impinged on it and thus warranted surgical decompression. In 2 patients (#45, 49), recurrence was defined based on recrudescence of visual field deficits due to optic chiasm compression prior to 2nd surgery.

Radiology

Radiological imaging was available for all patients in this study. Pituitary MRIs were uploaded encoded into an image analysis software (Osirix MD, v. 11.0, Bernex, Switzerland) for review by the study Neuroradiologist (AGK). Lesion size in the anterior/posterior, cranial/caudal, and left to right dimensions were measured directly from the T1 weighted post contrast images. T2 weighted coronal views were also used in determining the relationship of the lesion to the optic chiasm and assess for its compression. Knosp grade of parasellar extension of the tumour towards the cavernous sinus in relation to the intracavernous carotid artery was determined [27].

Tissue processing, histology and immunohistochemistry

All GT were surgical resection specimens that were fixed in 10% neutral buffered formalin for <24 hours prior to paraffin embedding. For twelve specimens, a portion of the tumour was also flash frozen in liquid nitrogen and banked for subsequent research. Paraffin sections (7 μm) were stained with haematoxylin and eosin or immunostained on automated immunostaining platforms for tumour markers (Table 1).

Table 1:

Details regarding primary immunohistochemical stains used in this study.

Immunohistochemical stain	Company	Catalogue number	Dilution	Primary antibody incubation time (mins)	Protocol and duration	Platform
SF-1	Thermofisher	434200	1:250	15	ER2, 40 mins	Leica Bond Instrument
Pit1	Santa Cruz	SC-393943	1:100	32	CC1, 64 mins	Roche Ventana Instrument
TPit (TBX19)	Atlas Antibodies	HPA072686	1:200	30	ER1, 30mins	Leica Bond Instrument
GATA-3	Biocare	PM 405 AA	Predilute	30	ER1, 20 mins	Leica Bond Instrument
Ki-67	Ventana	790–4286	Predilute	16	CC1, 36 mins (CC1)	Roche Ventana Instrument
SOX2	Ventana	760–4621	Predilute	32	CC1, 52 mins	Roche Ventana Instrument
DAX1	Abcam	ab196649	1:100	40	ER2, 40 mins	Leica Bond instrument

A board-certified neuropathologist (RAH) qualitatively and quantitatively assessed SF-1 immunohistochemical staining using light microscopy. The lowest staining SF-1 field of each tumour was photographed at 200x magnification and the proportion of SF-1 labelled cells were manually counted from a total of 1000 cells. The Ki-67 proliferative index was determined by first locating the highest staining fields and then counting the number of Ki-67 positive stained nuclei in a field of 1000 cells.

Statistical analysis

Statistical analyses for demographics, radiographic and pathological data were performed using GraphPad Prism version 8.2.1. The majority of the data were not normally distributed by D’Agostino & Pearson test, and therefore non-parametric tests were used in these instances (Kruskal Wallis test with post-hoc Dunn tests for median comparisons or Spearman’s rank correlation coefficient for correlative tests). The time to recurrence between groups was assessed by Kaplan-Meier and Log-rank (Mantel Cox) tests. P values < 0.05 were deemed statistically significant (*) and P values <0.01 were considered highly significant (**).

RNA extraction and quality control

Of the 51 GT, 12 had banked fresh frozen tissue available (Group 1: n=5, Group 2: n=3, Group 3: n=3, Group 4: n=1). Tissue sections and RNA extraction were performed at the Columbia University Molecular Pathology Core. One H&E stained section per frozen block was examined (RAH) to confirm that the vast majority of the section contained tumour. Two frozen tumour samples from group 1 (cases #11 and #13, Table 2) were excluded because of extensive normal pituitary gland present within the sample. For each sample, twenty 10 μm thick frozen sections were cut on a cryostat and placed into a 2 mL round bottom RNase free tube pre-cooled on dry ice. RNA was extracted by Qiagen QIAsymphony SP auto-extractor, with standard total RNA plus microRNA protocol (miRNA_CT_400_V8) and eluted with 100 μl elution buffer from the kit. RNA concentration and quality were checked with Nanodrop 8000 and Agilent Bioanalyzer 2100 system using RNA Nano chip. The mean RNA integrity numbers (RIN) of these 10 samples was 9.6 (range: 9.0– 9.9).

Table 2:

Clinicopathological features of the four GT groups

Group	Case	Age (years)	Sex	Tumour size (mm)	Knosp grade	SSE/CI	Oncocyic change	Mitotic Index (mitoses/10 HPF)	Immunohistochemistry						Follow up duration (years)	Time to recurrence (years)
									SF-1	PIT1	TPit	SOX2	GATA3	Ki-67 index (%)
1	1*	29	M	46	3	Y/Y	N	1	0.91	-	-	-	+++	1.6	N/A
	2	44	M	25	2	Y/Y	Y	0	0.85	-	-	R	+++	0.72
	3	61	M	21	1	Y/Y	N	0	0.93	-	-	-	+++	0.8
	4	56	M	22	3	Y/Y	N	0	0.83	- (E)	-	VR	NA	5.0
	5	68	M	31	2	Y/Y	N	1	0.77	-	-	-	++	0.9
	6	51	W	28	3	Y/Y	N	0	0.80	-	-	Subset	+++	1.4
	7*	47	M	33.7	2	Y/Y	N	2	0.82	-	-	VR	+	3.7
	8	72	M	52	3	Y/Y	N	2	0.68	-	-	NA	+++	1.7
	9	47	M	31	1	Y/Y	Y	0	0.90	-	- (E)	-	+++	0.8
	10*	53	M	35	2	Y/Y	N	1	0.08	- (E)	-	Subset	+++	10.1
	11	72	M	30	2	Y/Y	Y	0	0.90	-	-	R	+++	1.5
	12	43	M	26	1	Y/N	Y	1	0.67	- (E)	NA	R	NA	<1
	13	54	M	18	2	Y/	N	0	0.94	-	-	Subset	+++	<1
	14	47	W	23	3	Y/	Y	1	0.65	-	-	Follicular	+++	2
	15	56	W	19	1	Y/Y	N	0	0.95	-	-	Subset	+++	<1
	16	44	M	30	3	Y/Y	N	0	0.09	- (E)	- (E)	R	++	<1
	17	46	M	21	1	Y/	N	1	0.95	-	- (E)	Subset	+++	<1
	18	54	M	21	1	Y/Y	N	2	0.63	-	-	-	++	<1
	19	62	M	30	2	Y/Y	N	1	0.86	-	-	R	+++	2.2
	20	55	M	20	3	Y/Y	N	2	0.81	-	-	Subset	+++	3.2
2	21*	52	W	30	1	Y/Y	Y	1	0.92	-	-	-	-	5	5.6
	22	58	W	14	1	Y/N	N	0	0.90	- (E)	-	R	+++	0.6	6.1
	23*	62	W	33.4	3	Y/Y	N	0	0.90	-	-	-	+++	2	6.2
	24*	32	M	23	1	Y/Y	N	1	0.84	-	-	-	++	1.8	1.7
	25	61	M	22	NA	Y/?	N	0	0.84	NA	NA	R	+++	6	6.8
	26	74	M	20	1	Y/	N	0	0.87	-	-	-	+++	1.9	6.3
	27	60	M	30	4	Y/Y	Y	1	0.59	- (E)	-	-	++	1.1	2.3
	28	40	W	34	1	Y/Y	N	1	0.90	NA	-	Follicular (subset)	+/−	2.1	8.6
	29	74	M	20	1	Y/Abut	N	0	0.89	-	- (E)	-	+++	0.8	7.0
	30	53	M	19	1	Y/Abut	N	0	0.49	-	-	-	+++	2.6	6.8
	31	68	M	18.3	2	Y/Abut	Y	0	0.94	-	-	NA	+++	0.9	5.1
3	32*	35	M	34	3	Y/Y	N	3	0.16	- (E)	+	Subset	+	1.8		3.2
	33	52	W	29	2	Y/Y	N	2	0.10	- (E)	-	R	++	1.4		2.9
	34	38	M	29	3	Y/Y	Y	7	0.50	- (E)	-	R (follicle)	+	4.4		3.8
	35	61	W	28	3	Y/Y	N	1	0.01	- (E)	-	-	++	1.6		6.0
	36	38	M	38	3	Y/Y	N	0	0.73	NA				1.5		5.0
	37*	70	M	60	4	Y/Y	N	3	0.34	-	-	-	++	1.3		3.7
	38	53	M	30	4	Y/Y	N	1	0.74	- (E)	-	NA	+	3.1		6.1
4	39	64	M	41	3	Y/Abut	N	0	0.89	-	-	-	+++	1		8.0
	40	74	M	39	NA	Y/Y	N	0	0.45	NA				1.1		3.2
	41*	68	W	27	2	Y/Y	N	1	0.61	-	-	-	+++	0.6		9.1
	42	70	W	30	4	Y/N	N	0	0.052	-	NA	-	-	1		6.0
	43	55	W	15.8	1	Y/N	N	0	0.021	- (E)	- (E)	Subset (focal)	++	1.6		2.9
	44	39	M	14	3	Y/N	N	0	0.42	- (E)	-	R	+++	1		3.2
	45*	60	M	23	3	Y/N	N	0	0.38	NA				0.6		3
	46	42	M	23	3	Y/N	N	2	0.56	-	-	R	+++	2.8		3.8
	47	62	W	25	3	Y/N	N	2	0.85	-	-	-	+++	2		4.0
	48	44	M	11	3	Y/N	N	0	0.22	-	NA	-	+	2.8		5.0
	49	82	M	29	2	Y/Y	N	1	0.34	NA				0.9		18
	50	56	M	18	2	Y/Y	N	1	0.58	- (E)	-	-	+++	1.1		Unknown
	51	75	M	27	4	Y/N	N	0	0.88	-	-	Subset	++	1		25

The SF-1 column refers to the lowest proportion of SF-1 positive nuclei (i.e. SF-1 labelling index, (SLI)). PIT1, TPit, SOX2 and GATA-3 columns refer to the semi-quantitative immunohistochemical labelling of those proteins; the degree of immunolabeling was graded as either negative (−), focally positive (+), positive immunolabeling in the majority of the tumour (++) or diffusely positive (+++). In some instances, entrapped non-neoplastic, residual adenohypophyseal cells were noted, referred to as (E). The distribution of SOX2-labeled cells were either negative (−), very rare (VR), rare (R), positive in a subset of cells but without a clear pattern of staining (Subset) or were arranged around colloid-filled follicle (follicular). The Ki-67 labelling index denotes the highest percentage of tumour cells labelling in the stained section. Other abbreviations:

^*: RNAseq performed on this tumour, M: man, W: woman, SSE: suprasellar extension, CI: chiasmal impingement, Y: yes, N: no

Library preparation, RNA sequencing and analysis

All samples were prepared simultaneously to avoid bias related to batched library preparation. All procedures were performed by the Columbia Genome Center. In brief, mRNA was isolated using poly-A tail pulldown and cDNA libraries constructed using Illumina TruSeq reagents. Libraries were sequenced using Illumina NovaSeq 6000 at Columbia Genome Center. Samples were multiplexed in each lane, yielding approximately 40 million paired-end 100bp reads for each sample.

RTA (Illumina) was used for base calling and bcl2fastq2 (version 2.20) for converting BCL to fastq format, coupled with adaptor trimming. Then, a pseudoalignment to a kallisto index was created from transcriptomes (Human: GRCh38) using kallisto (0.44.0). Differentially expressed genes were examined under various conditions using Sleuth, an R package designed to compute transcript and gene-level differential expression from kallisto abundance files.

Gene set enrichment analysis

Analysis of gene set enrichment was performed using Webgestalt [28]. Gene names were entered using the pathway functional database, KEGG, and entering the top 89 genes and log₂ fold change that were significantly differentially expressed with a false discovery rate <0.05.

STRING protein interaction analysis

To analyse protein interactions, we entered the top 89 differentially expressed genes (adjusted P <0.05) or top 138 differentially expressed genes (adjusted P <0.1) and applied K-means clustering using the STRING network database, 2019 [29].

RNA validation using NanoString nCounter® expression quantification

Rapid multiplex RNA analysis was performed on extracted RNA of the previously sequenced GT using a custom-designed Nanostring nCounter codeset and the NanoString digital molecular counting platform at Columbia Human Immune Monitoring Core (RRID: SCR_016740). The analysis panel contained paired probes for 30 genes, which hybridize with target RNA molecules in the samples and, after purification and immobilization steps, generate barcoded fluorescent signals for direct molecular counting on the NanoString instrument. The counts of 25 genes of interest were then normalized using 5 housekeeping genes and control probes spiked into each sample and analysed using the NanoString nSolver software v.4.0 (Seattle, WA) and GraphPad Prism v.8.2.1.

Results

Demographics

Of the 51 tumours, 38 were resected from male patients. The mean patient age at the time of resection was 55.6 years (standard deviation: 12.5 years) and the mean tumour size of the initial resections was 28.0 mm (standard deviation: 9.4 mm). All tumours had suprasellar extension with the majority of tumours compressing the optic chiasm. There were no significant differences of pre-operative tumour size between groups except between groups 3 and 4; group 4 having a significantly smaller tumour size than group 3 because these patients had prior tumour resection (P=0.025, Mann-Whitney U test). The results for each case, including the demographic information and details of each tumour, are provided in Table 2.

Confluent absence of SF-1 labelling is more frequently seen in GT that recur

The typical appearance of GT within the cohort comprised sheets and large nests of neoplastic cells with amphophilic cytoplasm and a pseudopapillary growth pattern (Fig. 1A, G, Suppl Fig. 1A–B). Occasional eosinophilic colloid-laden follicles were seen within the tumour (Fig. 1D), as sometimes was oncocytic change (Table 2). Clear morphological differences between tumours that recurred and those that did not were not clearly apparent on H&E sections. SF-1 immunoreactivity was diffuse and strong in many GT (n=19 of groups 1 and 2, Fig. 1B, 1E) and GATA-3 immunoreactivity was present in the vast majority of tumours, often with strong, diffuse nuclear staining, supporting that these were GT (Fig. 1C) [30, 31]. As outlined in Table 2, the vast majority of GT were either entirely negative for TPit and PIT1 or had occasional entrapped cells that labelled for either antigen. However, SF-1 immunostaining demonstrated a distinct pattern in tumours that subsequently recurred. In 18 of the 20 tumours that recurred (groups 3–4), there were areas of confluent absence of SF-1 immunolabeling in contrast to group 2 which showed mostly diffuse expression of SF-1 (P=0.0007, n=31, two-sided Fisher’s exact test). Frequently, there were areas of perivascular retention of SF-1 immunolabeling with abrupt absence of labelling within the centres of tumour nests (Figs. 1H, 1K). Therefore, when comparing GT with patchy SF-1 versus diffuse SF-1 labelling, tumours with patchy staining were significantly more likely to recur than those that have diffuse staining with an odds ratio of 24.0 (CI: 3.3–132, P=0.0007, two-sided Fisher’s exact test).

Figure 1:

Histological and immunohistochemical features of selected GT. A-C: Case #1 (Group 1) shows a monomorphic tumour with diffuse, strong SF-1 and GATA3 immunoreactivity. D-F: Case #28 (Group 2) was known not to recur within 8.6 years. The tumour is histologically monomorphic (D) and shows diffuse, strong SF-1 expression throughout the tumour (E). F: Unlike case #1, this tumour has rare neoplastic nuclei that weakly label for GATA3, shown at higher magnification in inset (F’). G-I: Case #34 (Group 3) recurred within 3.8 years after initial resection. SF-1 shows immunopositivity in a perivascular distribution with abrupt negative staining in the nests (H). A similar patchy staining is noted with a GATA3 immunostain (I). J-M: Case #35 (Group 3) recurred within 6.0 years. Clear cell change is noted in the centre of nests whereas tumour cells with amphophilic cytoplasm are noted in a perivascular distribution. Patchy SF-1 (K) and GATA3 (L) staining is seen in the tumour. M: Immunocytologic variability of SF-1 staining within GT is shown and graded from 1–4. N: A 2 × 2 contingency table of Groups 2–4 showing that most recurrent tumours had patchy SF-1 labelling unlike in the non-recurrent group. Scale bar: 100 μm.

Immunocytological variability in SF-1 nuclear labelling is shown in Fig. 1M. Diffuse (1) and granular (2) staining were considered positive staining for the purposes of SF-1 assessment whereas single intranuclear dots (3) and negative staining (4) were not. Patchy SF-1 staining did show subtle morphological differences within several tumours, most notably in case #16 (Suppl Fig. 1A–E). In this GT, there were clearly areas of diffuse, strong SF-1 labelling and morphologically these neoplastic cells were composed of cuboidal, amphophilic cells (Suppl Fig. 1B–C). However, in areas of patchy SF-1 staining, the neoplastic cells were columnar with pale staining cytoplasm and nuclear polarization away from vessels (Suppl Fig. 1D–E).

Oncocytic change was found in a subset of tumours belonging to groups 1–3 and were not enriched in either group. Weak-to-absent SF-1 nuclear labelling was sometimes seen in oncocytic tumour cells, although not always, such as in cases #2 and #9. Furthermore, although variation of GATA3 immunolabelling was noted in a subset of GT with patchy SF-1 labelling, such as cases #34 and #35 (Fig. 1I, 1L), this was inconsistent; in some tumours with strong, diffuse SF-1 staining, there were very few neoplastic cells positive for GATA3 (e.g. case #28, Fig. 1E–F) as has been described previously [30, 31].

GT that recur have fewer SF-1 immunolabeled nuclei than non-recurrent GT

The lowest staining fields of SF-1 were photographed at an original magnification of 200x. Diffuse or granular staining nuclei (Fig. 1I) were manually counted and the proportion of labelled nuclei in a total of 1000 cells was calculated (SF-1 labelling index (SLI)). The median ± interquartile range SLI was 0.83 ± 0.24, 0.89 ± 0.067, 0.34 ± 0.63, and 0.45 ± 0.45 for groups 1–4, respectively (Fig. 2A). These median SLI values varied significantly between the groups (H=17.54, P=0.0005, Kruskal Wallis test). Significant reductions in SLI were evident in recurrent groups 3 and 4 when contrasted with the non-recurrent tumours of group 2 (P=0.0099, P=0.016, respectively, post-hoc Dunn test). There was no significant difference in the SLI between groups 3 and 4 nor between groups 1 and 2 (P=0.99, post-hoc Dunn test).

Figure 2:

A: Violin plots comparing the SF-1 labelling index of GT between groups. Dashed horizontal lines reflect the median values. Following a Kruskal Wallis test, a post-hoc Dunn test was performed between groups to identify significant differences; a single asterisk (*) indicates P<0.05 and a double asterisk (**) indicates P<0.01. Representative photomicrographs of SF-1 immunolabeled GT at 200x original magnification are shown above each corresponding group. B: Kaplan Meier recurrence-free survival curves of groups 2 and 3 according to SLI, Ki-67 proliferative index, and mitotic index parameters. Ticks indicate censoring due to loss of follow-up after that specific time point. C: Receiver Operating Characteristic Curve comparing the SLI, Ki-67 proliferative indices and mitotic index as predictors of recurrence.

In groups 1–3, which represented GT that were initially resected, there was an inverse correlation between the tumour size and the SLI (r=−0.33, P=0.044, n=38, Spearman’s rank correlation, Supplemental Fig. 2). Radiographically larger tumours were therefore associated with a lower SLI. Likewise, in groups 1–3, the Knosp grade was inversely correlated with the SLI (r=−0.48, P=0.0028, n=37, Spearman’s rank correlation, Supplemental Fig. 3).

The SLI is more reliable than Ki-67 in predicting recurrence in GT

We compared the Ki-67 labelling index, mitotic index and the SLI between group 2 and groups 3/4 and the time to recurrence. We first explored the time to recurrence following initial resection using an arbitrary SLI of above or below 80% in GT from groups 2 and 3 (Fig. 2B). Initial resections of GT with a SLI < 80% had a median time to recurrence of 4.4 years (n=9) whereas tumours with SLI >80% did not show evidence of recurrence; these patients had a median follow up time of 6.3 years (n=9). A SLI <80% was therefore associated with a significantly shorter time to recurrence than SLI > 80% (P=0.0007, n=18, Log Rank (Mantel-Cox) test, Fig. 2B). An elevated mitotic rate was associated with a shorter time to recurrence (3.45 years versus 6.12 years, P=0.0024, Log Rank (Mantel-Cox) test), however, the Ki-67 proliferative index was not associated with a significantly shorter time to recurrence in either assessment (P=0.89, groups 2 vs. 3, Log-rank (Mantel Cox) test).

Since there was no significant difference in SLI between groups 3 and 4 and they had a similar staining pattern as their initial resection, we then constructed a receiver operating characteristic (ROC) curve using values taken from groups 2, 3 and 4 (Fig. 2C) to assess whether the extent of SF-1 labelling was a better diagnostic test than the Ki-67 labelling index as a binary classifier. Indeed, the area under the curve (AUC) was greater using SLI (AUC=0.89) than mitotic index (AUC=0.65) or Ki-67 proliferative index (AUC=0.63), supporting the use of SLI in clinical practice to predict recurrence of GT.

Significant transcriptomic differences exist between high SLI and low SLI gonadotroph tumours

To further explore the differences between GT with high SLI (SLI >80%) and GT with low SLI (SLI <80%), we performed RNAseq analysis to assess differential gene expression on five GT with high SLI and five GT with low SLI. Three of these tumours were recently resected with minimal follow up times. After allowing for a false discovery rate (FDR) of less than 0.05, 89 genes were significantly differentially expressed (Fig. 3A–B) between the 2 groups (full list provided in supplemental table 1). Upregulation of the stem cell gene, SOX2 was noted in low SF-1 tumours, as were genes involved in adenohypophyseal development, such as DLK1 and RBFOX3. Whilst the gene encoding SF-1, NR5A1, was not differentially expressed, a related gene that encodes for DAX1 was, NR0B1 (log₂fold change −2.50, adjusted P=0.021). To our surprise, upregulation of genes related to prolactin and growth hormone was more frequently seen in the low SLI gonadotroph tumours.

Figure 3:

Gene expression profiling data obtained following RNA sequencing. A: Heatmap of unsupervised hierarchical cluster analysis of the fifty most significantly differentially expressed genes after variance stabilizing transformation of high SLI (>80%, n=5) and low SLI (<80%, n=5) GT allows clear discrimination between the two groups. B: A Volcano plot showing the 89 significant, differentially expressed genes (blue) between the two SLI groups. Significance was determined by using the Wald test and Benjamini-Hochberg False Discovery Rate (FDR) that was set to 0.05. Genes with higher expression in low SLI GT are plotted to the right of the graph, whilst genes with lower expression in low SLI GT are plotted to the left of the graph. C: Gene set enrichment analysis showing significant enrichment of the PI3K-AKT signalling pathway in addition to other pathways involved in cytokine-cytokine receptor and neuroactive ligand-receptor interactions. D-I: Concordance of DAX1 immunohistochemistry with SF-1 immunolabeling high SLI (D-E) and low SLI tumours (H-I). D-E: Case #21 had diffuse SF-1 and DAX1 expression. F-G: Case #37 showed strong perivascular labelling for SF-1 with absence of staining within central nests, which was a similar pattern of staining with DAX1. H-I: Case #10 is a low SLI tumour with a disordered mixture of SF-1 positive cells; the DAX1 shows a similar pattern of labelling. Scale bar: 50 μm.

Several overexpressed oncogenic genes were also noted in GT with low SLI. Of note, over-expression of BCL2 was seen in GT with low SLI (log₂fold change 1.79, adjusted P=0.015) and upregulation of ERRB4 (log2fold change 5.36, adjusted P<0.001) was also observed. Gene set enrichment analysis of the top 89 differentially expressed genes demonstrated significant enrichment of the PI3K-AKT pathway, for which ERRB4 is a gene in that pathway (P=0.0043, Fig. 3C). In addition, there was enrichment for cytokine-cytokine receptor and neuroactive ligand receptor interactions, whereas metabolic pathways were decreased in low SLI tumours. A full list of significant functional enrichments is provided in supplemental table 2. Many of the genes that were differentially expressed are functionally related and clustered together with STRING network analyses (Suppl Fig. 4).

To validate our RNAseq findings, we performed targeted sequencing against 30 genes, many of which were differentially expressed by RNAseq, using the Nanostring sequencing platform on the original 10 RNA extracted samples. There was a highly significant, strong, correlation between the log₂fold change of the Nanostring RNA counts with the RNAseq log₂fold changes for these 30 genes (r=0.90, P<0.0001, Spearman’s rank correlation, Suppl fig. 5). We also found segregation of the high and low SLI tumours with this limited 25 targeted gene panel (Suppl fig. 6). Then, given the reduction in DAX1 gene expression, we immunostained FFPE sections of these ten tumours for DAX1. Consistent with our RNAseq findings, we found a similar pattern of DAX1 staining to that of SF-1 immunolabelling in the majority of cases (7 of 10 cases), with low SLI tumours showing patchy DAX1 immunolabeling (Fig. 3D–I). However, immunohistochemical stains against SOX2 showed variable staining amounts and patterns across groups (Suppl Fig. 7). In some instances, SOX2 was found around colloid-laden follicles (Suppl Fig. 7C) and sometimes occasional cells were seen interspersed within the tumour, however, the abundance of SOX2-labelled cells did not appear to correspond with the SLI of the tumour, with some low SLI tumours appearing negative for SOX2 (Suppl Fig. 7I). Although there were significant increases in SOX2 transcripts in low SLI tumours, these changes therefore did not manifest at the protein level by immunohistochemistry.

Discussion

To our knowledge, this is the largest study of GT to date that incorporates clinical, histological, immunohistochemical and molecular analyses to examine differences between indolent and more aggressive GT. Specifically, we have shown that patchy SF-1 immunolabeling and low SLI are highly enriched features in GT that recur. We found an inverse correlation between the SLI and both the radiological size of the tumour and the Knosp grade. Higher Knosp grades are associated with worse outcome following pituitary surgery and therefore suggests that GT with low SLI may be more biologically aggressive [27, 32]. Indeed, GT with SLI <80% were faster to recur and require reoperation than those with a SLI >80%. RNA sequencing studies confirmed significant biological differences between GT with low versus high SLI and showed signatures that lend further support to the hypothesis that these tumours have a more aggressive biological potential.

Comparison of the transcriptome of high SLI and low SLI GT revealed upregulation of ERRB4, PRL, GH and NGFR genes, which are genes involved in the PI3K-AKT pathway (Fig. 3C). This pathway has been discussed in various contexts in the pituitary literature [33] and targeting ERRB4 in cancers with ERRB4 inhibitors, such as ibrutinib, has shown efficacy in some settings and may therefore be a therapeutic option for aggressive GT [34]. Surprisingly, the gene that encodes SF-1, NR5A1, was not differentially expressed between the two groups. Possibilities to explain the differences in SF-1 labelling include post-translational modifications of the SF-1 protein or possibly an increased rate of protein turnover in low SF-1 tumours. However, reductions in DAX-1 expression were noted, which is another orphan nuclear receptor protein involved in pituitary gonadotropin production and expressed in GT [35, 36]. Consistent with this, the pattern of DAX1 immunohistochemical staining corresponded with the SF-1 immunolabeling of the tumour. One conjecture is that DAX1 may be interacting with the nuclear receptor of SF1 and thereby influencing the staining pattern [37, 38]. Another possible explanation for why there is perivascular retention of SF-1 but absence of labelling in the centre of nests in low SLI GT may relate to sensitivity to hypoxia. This might be supported by the fact that larger tumours that are more difficult to resect and have longer periods of ischaemia are more likely to have low SLI. However, we did not find hypoxic genes, such as the hypoxia inducible factors (HIF) or vascular endothelial growth factors (VEGF), to be upregulated in low SLI tumours. Furthermore, to our knowledge, clear changes in SF-1 expression under hypoxic states have yet to be formally established [39].

Apart from a significant increase in pituitary stem cell genes, SOX2 and GFRA3, which could suggest an increase in a stem cell population, no significant differences in POU1F1 or TBX19 genes were detected that encode for PIT1 and TPit, respectively [40]. The increased expression of SOX2 coincided with increased expression of genes involved in pituitary development (e.g. DLK1, TBX2, NGFR) and neuronal differentiation (e.g. RBFOX3, PRPH) [41–45]. The latter was particularly interesting, since gangliocytomas do rarely occur as sellar tumours and there is evidence that combined sellar gangliocytoma/PitNET can express PIT1 in both components suggesting that gangliocytomas may arise from PitNET [46]. Expression of neuronal genes have also been seen in transcriptomic studies in the MENX mouse model that resembles gonadotroph tumours [47]. Another interesting finding was that these tumours were non-functioning tumours, yet demonstrated upregulation of hormonal genes, such as growth hormone and prolactin. SF-1 labelling has been seen rarely in somatotroph and corticotroph tumours and may not be entirely specific for the gonadotroph lineage, however, the majority of our cohort was confirmed to be gonadotroph by also immunolabeling with GATA-3 [48]. Why this hormonal switch is occurring in these more aggressive tumours is unclear but may reflect a loss of differentiation within the tumour. Certainly, the patchy SF-1 labelling seen in these low SLI tumours raises the possibility that the variability of staining reflects a poorer differentiation within the tumour and increased cellular heterogeneity [22]. Importantly, these negative areas of SF-1 did not represent areas with tumour cells of a different lineage (TPit, Pit1) or SOX2+ stem cell and therefore were not double/triple PitNET. Given that the vast majority were GATA3+, we therefore assume that the cells without significant SF-1 labelling (grades 3–4, Fig. 1I) are of gonadotroph lineage. A recent large-scale pangenomic assessment of PitNET identified similar transcriptomes between null cell and gonadotroph tumours [48]. Since null cell tumours are believed to have an aggressive behaviour and show neither SF-1, TPit or Pit1 immunolabeling, patchy SF-1 in GT may suggest a behaviour akin to null cell tumours [49, 50].

Aside from certain high-risk PitNET subtypes, such as the silent corticotroph and sparsely granulated somatotroph tumour, current methodologies to predict recurrence have largely ignored cellular lineage and mainly relied on clinical and pathological assessments of PitNET irrespective of cellular differentiation/ lineage [6, 7]. In this study, when the Ki-67 proliferative index is set at 3%, as has been recommended in the past, there was no association with recurrence. However, the SLI, and to some extent mitotic index, appeared to be more robust predictors for recurrence in our study [51].

Limitations

One limiting factor in this study is the short follow up times that were available for the non-recurrent group (group 2). As can be seen in groups 3 and 4, sometimes the time to recurrence can be many years, such as case #51, which ‘recurred’ 25 years later. However, at this time point, it is difficult to ascertain whether this represents recurrence of a previously resected tumour or a metachronous event. Another limitation of this study is the limited number of well-characterized fresh frozen samples available for RNAseq. Whilst there are striking differences in RNA profiles between the two groups, the low sample size likely increased the type II error rate, such that many more genes that are differentially expressed are not detected. Additional RNAseq and validation studies are needed to explore additional molecular differences between the two groups and to identify potential drug targets.

The inverse correlation between tumour size and SF-1 labelling index raises the possibility of confounding since larger tumours are more likely to recur and increased tumour volume may limit timely formalin penetration with inadvertent lowering of SF-1 antigenicity. However, since these tumours are usually received via transsphenoidal hypophysectomy, the specimens are fragmented and thus formalin penetration is unlikely to be affected. Moreover, these specimens were placed into formalin intraoperatively and submitted for processing within 24 hours. In addition, the differential gene expression and sometimes different morphologies observed within these tumours argues against low SF-1 labelling being a formalin fixation artefact. Thus, our study provides multiple lines of evidence that there is a difference in the biological potential of GT with patchy SF-1 labelling.

Lastly, there is limited knowledge of transcriptome profiles in gonadotroph tumours to compare with our results. Falch et al. compared the transcriptome of fast growing GT with slow growing GT [52]. The only gene differentially expressed in both studies was LGALS3BP. The lack of extensive overlap may well reflect markedly different methodologies in selecting and categorizing tumours.

Conclusion

We herein show that GT with patchy SF-1 staining are more likely to recur and sooner than GT with diffuse, strong SF-1 labelling. We further demonstrate marked differences in differential gene expression, enrichment of genes involved in the PI3K-AKT pathway and relative upregulation of stem cell genes. We speculate that GT with patchy SF-1 reflects an increased intratumoural heterogeneity and poorer differentiation than diffusely SF-1 staining GT. Future studies validating this work and exploring cellular heterogeneity within these tumours at a single cell resolution may reveal novel prognostic and therapeutic targets for patients with aggressive PitNET.