A Performance Comparison of Commonly Used Assays to Detect RET Fusions

Fusions involving the RET receptor tyrosine kinase are recurrent drivers in cancer (1). The RET protein, encoded by the RET (REarranged during Transfection) gene, is activated by neurotrophic factors that induce receptor dimerization, phosphorylation of the kinase domain, and signal transduction of downstream pathways in physiologic conditions (2, 3). In tumors with activating RET fusions, a partner gene, most commonly KIF5B, CCDC6, or NCOA4, is fused to the distal portion of the RET gene including the kinase domain (4–9). This chromosomal rearrangement or structural variant (SV) can generate a constitutively active chimeric protein composed of a C-terminal coil-coiled domain and an N-terminal kinase domain.

Since their initial discovery in papillary thyroid cancers, RET fusions have been identified in a diverse set of solid tumors at varying frequencies. Most notably, up to 2% of non–small cell lung cancers (NSCLCs) and 10% of thyroid cancers can harbor oncogenic RET fusions (10). In addition, RET fusions have been infrequently reported (<1%) in other common and rare tumors including colorectal cancer, pancreatic cancer, breast cancer, melanoma, soft-tissue sarcoma, salivary duct carcinoma, and ovarian cancer, among others (10–12). Despite their overall rarity, patients with advanced RET fusion–positive tumors collectively represent a large, genomically distinct cohort with suboptimal outcomes to conventional chemotherapy, immunotherapy, and multikinase inhibitors (13–16).

Recently, a new generation of highly selective RET tyrosine kinase inhibitors (TKIs) has been developed to target oncogenic RET fusions and mutations (17). In clinical trials, these agents, namely, selpercatinib and pralsetinib, have demonstrated deep and durable responses in most patients with advanced RET fusion–positive tumors (18–21). Given these dramatic results and the recent FDA approval of selpercatinib and pralsetinib, testing for RET fusions is now standard of care for patients with advanced NSCLC and thyroid cancer (22). Currently, there are multiple methods for detecting RET fusions including FISH, RT-PCR, and DNA/RNA-based next-generation sequencing (NGS). In addition, RET IHC can be used to measure RET protein expression, which may serve as a surrogate marker for RET fusions. Despite the growing number of diagnostic assays, the variability in their performance represents a significant challenge for harmonizing RET fusion testing. In prior studies, RET IHC has shown poor correlation with RET fusion status as determined by RT-PCR and FISH (23, 24). However, these reference methods have not been extensively benchmarked against newer NGS-based assays. Furthermore, DNA-based NGS is known to have a limited sensitivity for certain fusions (25–28), and its accuracy for identifying RET fusions remains largely unknown. Here, we compare the performance of DNA- and RNA-based NGS, FISH, and IHC and explore the molecular landscape of actionable RET fusions in a large pan-cancer cohort.

After approval from our institutional review board, a retrospective review of the MSK-IMPACT (DNA-based NGS) and MSK-Fusion panel (RNA-based NGS) results from January 1, 2014 to May 9, 2020 was performed. MSK-IMPACT and MSK-Fusion as well as break-apart FISH were all performed in Clinical Laboratory Improvement Amendments–approved laboratories following standard operating procedures.

Genomic DNA was extracted from formalin-fixed, paraffin-embedded tissue and submitted for MSK-IMPACT. MSK-IMPACT is a hybridization capture-based NGS assay that interrogates all exons and select introns of 468 genes to identify mutations, copy-number changes, microsatellite instability status, and select SVs in its current iteration (previous versions interrogated 341 and 410 genes; ref. 29). All three panels covered the entire coding region of RET as well as introns 9–11. Introns 7 and 8 were added in the latest panel with 468 genes.

DELLY was used to detect SVs from MSK-IMPACT as described previously (30). All SV calls involving RET were manually reviewed; only the SVs with high-quality paired-end reads (≥2) were included. In tumors with multiple RET SVs, the most canonical SV and/or the SV with the highest read support were selected. RET SVs were annotated as kinase fusions if they were predicted to form a protein fusion involving (i) a non-RET 5′ partner gene and (ii) an intact RET kinase domain (NM_020975: exons 12–18) in the 3′ partner position. Furthermore, RET SVs from MSK-IMPACT were classified into two groups. First, RET SVs that predicted RET kinase fusions with previously described fusion partners were classified as recurrent oncogenic RET fusions. This included inversion-mediated RET SVs that were reciprocal to known kinase fusions (e.g., RET-KIF5B for KIF5B-RET). Second, novel and/or atypical SVs were classified as SVs of unknown significance (SVUS) and further characterized by MSK-Fusion (see below). This included (i) SVs that predicted kinase fusions with novel fusion partners, (ii) SVs with partners mapping to intergenic regions, (iii) intragenic RET SVs not involving other genes (e.g., deletion of RET exons 3–5), and (iv) SVs predicted to form antisense fusions involving two gene transcripts with incompatible orientations. All protein fusions were annotated as out-of-frame, in-frame, and mid-exonic based on output from DELLY. For our study, SVs with mid-exonic breakpoints were annotated as fusions involving the adjacent uninterrupted exon. In patients with multiple tumors that were analyzed using MSK-IMPACT, the results from the initial tumor sample and/or the sample with paired MSK-Fusion testing were used.

Cell-free DNA (cfDNA) was sequenced using MSK-ACCESS in a subset of patients in our cohort. MSK-ACCESS is a hybridization capture-based NGS assay that is optimized for detecting select genetic variants using circulating tumor cfDNA from plasma (31). The assay targets 129 genes and covers RET exons 12–19 and introns 7–11. Manta was used for SV calling as described previously (32).

Tumors from formalin-fixed paraffin-embedded tissues were tested for RNA-level fusions using MSK-Fusion. MSK-Fusion is a custom RNA-based NGS assay that uses Anchored Multiplex PCR (via the Archer platform) as described previously (28, 33). The custom panel covers select exons in 62 genes including RET exons 8–13. All fusion transcripts calls were manually reviewed, and only those with a minimum of five unique reads as well as three unique reads with unique start sites were reported. Tumors tested included (i) cases with DNA-level SVUS involving kinase genes including RET, as detected by MSK-IMPACT, (ii) de novo solid tumors lacking MAPK driver alterations on MSK-IMPACT (e.g., lung adenocarcinoma with no mitogenic drivers), and (iii) cases with clinical evidence of progression to a targeted or hormone therapy but lacking a resistance mechanism on MSK-IMPACT (e.g., EGFR-mutated lung adenocarcinoma progressing on EGFR TKI but no detectable resistance alteration).

Cases with RET SVs on MSK-IMPACT were placed into four groups based on the results from the above testing: (i) group A: RET SVs predicted to form recurrent oncogenic RET fusions (RET kinase domain and a known 5′ partner) and therefore not tested with MSK-Fusion, (ii) group B: RET SVs predicted to form recurrent oncogenic RET fusions that were confirmed on MSK-Fusion, (iii) group C: RET SVUS that were transcribed into RET fusions on MSK-Fusion, and (iv) group D: RET SVUS that did not result in RET fusions on MSK-Fusion. Cases from groups A, B, and C were defined as RET fusion positive. Cases from group D were considered RET fusion negative. For groups B and C, data from RNA sequencing were used for fusion annotation.

Break-apart FISH was performed on a subset of RET fusion–positive (n = 48) and -negative cases (n = 25) using a commercial break-apart FISH probe for RET (ZytoVison; Supplementary Methods). For each hybridization, a minimum of 100 nonoverlapping nuclei were assessed for numbers of green and red signals. Using a prespecified cutoff, FISH was considered positive for a RET rearrangement if ≥10% tumor cells demonstrated a rearrangement pattern as defined by the separation of red (5′) and green (3′) signals ≥2 signal diameter including atypical signal patterns. FISH analysis was performed by qualified clinical cytogenetic technologists and interpreted by a board‐certified clinical cytogeneticist (U. Aypar).

RET fusion–positive cases were selected from groups A, B, and C where additional material was available. RET fusion–negative cases included group D cases (n = 8) and treatment-naïve lung adenocarcinomas that were negative for RET SVs but positive for other activating MAPK alterations by MSK-IMPACT (n = 17). The latter group was included given that RET fusions are mutually exclusive with oncogenic drivers in treatment-naïve lung cancers (7).

RET IHC was performed on 70 RET fusion–positive and 89 RET fusion–negative cases using the clone EPR2871 (ab134100; Abcam). EPR2871 is a rabbit recombinant mAb generated to a proprietary peptide in the C-terminal portion of RET and has been used in prior studies (Supplementary Methods; refs. 8, 24).

RET fusion–positive cases were selected from groups A, B, and C where additional material was available. The negative control cohort included (i) group D cases (n = 9), (ii) lung adenocarcinomas (n = 34) and papillary thyroid carcinomas (n = 3) that were negative for RET fusions by MSK-IMPACT and MSK-Fusion, and (iii) treatment-naïve papillary thyroid carcinomas (n = 33) that were negative for RET SVs and positive for BRAF p.V600E by MSK-IMPACT. The last group was included given that RET fusions are mutually exclusive with the BRAF drivers in papillary thyroid carcinomas (34). Finally, we included 10 cases that were negative for RET SVs on MSK-IMPACT but positive for high-level RET amplification as defined by >2 fold-change gain.

For each case, we performed a comprehensive semiquantitative analysis capturing: (i) the staining intensity (1+, 2+, 3+), (ii) the estimated percentage of tumor cells with expression in 10% increments for each staining intensity, and (iii) the cellular localization of staining (e.g., membranous, cytoplasmic, and nuclear). Using the percentage and the staining intensity, we calculated the H-score for all cases as follows: (% with 1+) + 2 × (% with 2+) + 3 × (% with 3+). All cases were reviewed by a board-certified pathologist (S.-R. Yang) blinded to the RET fusion results. Equivocal or borderline cases were reviewed independently and resolved by a second board-certified pathologist (J.F. Hechtman).

Statistical analyses were performed using R version 3.4.0 (R Project for Statistical Computing) and GraphPad Prism version 8. All tests of significance were two tailed, and P < 0.05 was considered statistically significant. The 95% confidence intervals (CI) were calculated using exact binomial confidence limits. Publicly available data visualization tools were used for generating the circos plot (35) and the fusion proteins (36).

Of 46,897 tumors from 41,869 patients submitted for MSK-IMPACT, we detected RET SVs in 171 unique patient samples (0.4%), including 139 SVs that were predicted to result in oncogenic RET fusions with known partners and 32 SVUS (Fig. 1A). Among the 139 RET SV predicted to be oncogenic, 105 cases were not further analyzed given the canonical nature of these fusions (group A). MSK-Fusion was performed in the remaining 34 cases, including three RET SVs that were reciprocal to known oncogenic fusions. On RNA sequencing, all 34 cases were confirmed to harbor RET fusion transcripts with the same partners as predicted by MSK-IMPACT (group B). In addition, MSK-Fusion was performed on all 32 cases with SVUS for further workup. In this cohort, oncogenic RET fusion transcripts were detected in only 12 cases (group C; 37.5%), and the remaining 20 samples (62.5%) were negative for RET fusions (group D; Supplementary Fig. S1A). In all, we identified a total of 151 unique patient samples with oncogenic RET fusions (groups A, B, and C) using a combination of DNA- and RNA-based sequencing (Fig. 1B and C). The full clinical, pathologic, and molecular data are provided in Supplementary Table S1.

Here, we compared the oncogenic RET fusions that were predicted from DNA sequencing to the RET fusion transcripts that were detected on RNA sequencing in our group B cohort (n = 34). For the three reciprocal SVs, we used the inferred pathogenic SV for our analysis. Despite sharing the same fusion partner, a total of 12 RET fusion transcripts (35.3%) showed exonic breakpoints that were slightly different from those predicted from DNA sequencing (Supplementary Table S1B). Specifically, 10 RET fusion transcripts had different breakpoints for the 5′ fusion partner, with the majority of cases differing by a single exon (n = 9, 90%). Only three RET fusion transcripts showed different 3′ RET breakpoints as predicted from DNA sequencing, and most differed by a single RET exon (n = 2, 66.7%). Furthermore, out-of-frame RET SVs were seen in three cases, and all had different 5′ and/or 3′ exonic breakpoints compared with the RET fusion transcripts detected on RNA sequencing.

While RNA sequencing detected RET kinase fusion transcripts in all group C cases (n = 12), DNA sequencing identified only five RET SVs (41.7%) that were predicted to form a kinase fusion (three in-frame and two mid-exon). In these five cases, the fusion partners and the exonic breakpoints predicted from DNA sequencing were identical to those identified on RNA sequencing. Among the seven SVs not predicted to form a protein fusion, four (57.1%) involved rearrangements with intergenic regions, two (28.6%) were antisense fusions, and one (14.3%) was an intragenic RET deletion.

In this cohort of cases that lacked RET fusion transcripts (n = 20), DNA sequencing predicted five RET protein fusions (25%), including three out-of-frame fusions, one mid-exon fusion, and one in-frame fusion. Among these, one SV was predicted to form a kinase fusion involving a 5′ partner gene and a 3′ RET kinase domain (KCNMA1 exon 1 and RET exon 9–20); however, this was predicted to be out of frame. The remaining four fusions did not include the RET kinase domain in the 3′ position and instead placed RET in the 5′ location. Interestingly, these five partner genes are known to be widely expressed in many tissues, including their tissue of origin (37). While rearrangements involving these partner genes have been reported previously (38), they have not been seen in association with classic kinase partners (ALK, ROS1, RET, NTRK1/2/3, EGFR, and MET). The remaining 15 SVs (75%) in group D were not predicted to form a protein fusion; these included three antisense fusions, two of which were intragenic RET inversions, and 12 rearrangements involving intergenic regions.

In comparing groups C and D, SVUS that resulted in RNA-level fusions (group C) lacked concurrent driver alterations (0% vs. 60%, P = 0.0006, Fisher exact test) and had higher normalized paired-end read support on MSK-IMPACT (0.059 vs. 0.018, P = 0.01, Mann–Whitney U test). Specifically, group C had a median and mean normalized paired-end read support of 0.059 and 0.084, respectively (range, 0.026–0.277) while group D had a median and mean normalized paired-end read support of 0.018 and 0.045, respectively (range, 0.002–0.182). In addition, group C cases were enriched in SVs predicted to form kinase fusions compared with group D (41.7% vs. 5%, P = 0.02, Fisher exact test). To ensure that the fusion-negative results in group D were not due to low tumor purity or poor sample quality, we also compared the tumor cellularity and the cycle threshold values from MSK-Fusion between group D and group C cases and found no significant differences (P = 0.4 and P = 0.2, respectively, Mann–Whitney U test).

Among our 151 patients with oncogenic RET fusions, the median age was 61 years (range, 3 months–90 years), and the majority were women (n = 93, 61.6%; Table 1). The most common tumor type was non–small cell lung cancer (NSCLC; n = 99, 65.6%), followed by cancers from the thyroid (n = 35, 23.2%), colorectum (n = 7, 4.6%) and breast (n = 3, 2%). The remaining tumors included three stomach adenocarcinomas (2%), a pancreatic adenocarcinoma (0.7%), an infantile fibrosarcoma (0.7%), a juvenile xanthogranuloma (0.7%), and a carcinoma of unknown primary (0.7%). Most patients had advanced stage tumors and lacked other activating MAPK alterations, supporting the oncogenic role of RET kinase fusions. In addition, nearly all tumors with non-RET driver alterations (83.3%, 5/6) came from patients who were previously treated with targeted or hormone therapies, thus implicating RET fusions as a potential resistance mechanism in these patients (39, 40). Specifically, these included four patients with lung cancer with activating EGFR mutations (two exon 19 deletions and two L858R mutations) status after EGFR TKI therapy (range, 1–5 years) who acquired a RUFY2-RET fusion, a NCOA4-RET fusion, and two CCDC6-RET fusions. There was also a patient with estrogen receptor–positive breast cancer status after aromatase inhibitor therapy for 1 year who developed a SPECC1L-RET fusion. All of these patients showed clinical evidence of disease progression at the time of RET fusion detection. The clinicopathologic characteristics of group D patients are provided in Supplementary Table S2.

Given the diversity of tumors represented in our fusion-positive cohort, we analyzed the tumor-specific prevalence of RET fusions using our MSK-IMPACT database of 41,869 cases (Supplementary Table S3). Although RET fusions were rare in our unselected population (0.36%), they were enriched in patients with NSCLC (1.7%, 99/5,920) and thyroid cancer (4.9%, 35/710), as reported previously (41). Within thyroid cancers, papillary thyroid carcinomas showed the highest prevalence of RET fusions (7.8%, 27/348), followed by poorly differentiated thyroid carcinomas (4.7%, 7/148) and anaplastic thyroid carcinomas (1%, 1/98). In NSCLCs, RET fusions were identified in 2.1% (95/4,599) of lung adenocarcinomas, 3% (3/101) of large-cell neuroendocrine carcinomas, and 2% (1/49) of pleomorphic carcinomas. The remaining tumor subtypes had frequencies that were <1%.

Among the 151 RET fusions, KIF5B was the most common partner (n = 68, 45%), followed by CCDC6 (n = 44, 29.1%), and NCOA4 (n = 20, 13.3%; Fig. 1C). Other partners included SPECC1L (n = 2), CLIP1 (n = 1), TFG (n = 1), EML4 (n = 1), C10orf118 (n = 1), RUFY1 (n = 1), RUFY2 (n = 1), RUFY3 (n = 1), KIF13A (n = 1), TRIM24 (n = 1), ERC1 (n = 1), KIAA1468 (n = 1), KIAA1217 (n = 1), GEMIN5 (n = 1), SHROOM3 (n = 1), RBPMS (n = 1), GRIPAP1 (n = 1), and MBIP (n = 1). The full clinical and molecular characteristics are summarized in Fig. 2A.

Next, we analyzed the distribution of fusion partners with respect to tumor histology. In NSCLC, KIF5B was the most common partner (n = 68, 68.7%) and not seen in other tumor types. Other variants included CCDC6 (n = 18, 18.2%), NCOA4 (n = 3, 3%), and other partners (n = 10, 10.1%). In thyroid cancer, most of the fusions involved CCDC6 (n = 22, 62.9%), followed by NCOA4 (n = 9, 25.7%) and other partners (n = 4, 11.4%). In the remaining tumors, NCOA4 comprised nearly half of the fusions (n = 8, 47%), while the rest were split between CCDC6 (n = 4, 24%) and other partners (n = 5, 29%).

Although DNA-based NGS is being increasingly used to detect kinase fusions, its sensitivity and specificity for capturing RET fusions are not well characterized. Here, we evaluated the performance of MSK-IMPACT using MSK-Fusion as the reference standard (Table 2). In addition to the 66 samples (groups B, C, and D) that were cotested with MSK-IMPACT and MSK-Fusion in our series, we surveyed an additional set of 4,459 tumors from our MSK-IMPACT database that were submitted for MSK-Fusion testing (see RNA sequencing under Materials and Methods). For this analysis, detection of a RET SV from MSK-IMPACT was considered RET positive. In this combined cohort of 4,525 cases that were cotested with MSK-IMPACT and MSK-Fusion, MSK-IMPACT demonstrated 100% sensitivity (95% CI: 92.3–100) and 99.6% specificity (95% CI: 99.3–99.7) for detecting RET fusions. In particular, no new RET fusion transcripts were identified in the additional set of 4,459 tumors, and only 20 false-positive cases (group D) were present in the entire cohort. While the high specificity was maintained across the three different tumor groups, non-lung and non-thyroid malignancies comprised the majority of the false-positive cases (55%, 11/20). Notably, these false-positive SV calls had fewer read support on MSK-IMPACT compared with those that resulted in oncogenic fusion transcripts on MSK-Fusion (Fig. 2B).

To explore the detection of RET fusions from circulating tumor cfDNA, we analyzed a subset of patients from our cohort (n = 16) who underwent plasma-based cfDNA testing using MSK-ACCESS (Supplementary Table S1C). At the time of liquid biopsy, two patients had no evidence of disease (12.5%), whereas the remaining 14 patients had metastatic disease with no prior treatment (i.e., initial presentation; n = 3, 18.7%), progression of disease (n = 5, 31.3%), ongoing response to treatment (n = 1, 6.2%), and stable disease (n = 5, 31.3%). In this cohort of 16 patients, RET SVs were detected in five patients (31.3%) with either progression of disease (3/5, 60%) or at initial presentation (2/3, 66.7%). In four patients, cfDNA testing revealed the same RET SV as in tissue, with two cases sharing the identical genomic breakpoints and the remaining two cases showing slightly different genomic breakpoints but having the same exonic breakpoints as in tissue. These RET SVs included two oncogenic group A cases (KIF5B-RET and KIF13A-RET) and two group D cases with RET SVUS that were negative on tissue-based RNA sequencing. In the remaining case, liquid biopsy from a group D patient showed the same RET translocation with the identical genomic breakpoint as in tissue but with a different orientation.

Next, we evaluated the sensitivity of break-apart FISH using a subset of our cases with RET fusions (n = 48) and showed that sensitivity for FISH was 91.7% (95% CI, 80–97.7) for detecting RET fusions (Table 2; Fig. 3). In this study, we did not include a large cohort of RET fusion–negative cases to determine overall specificity. Instead, our specificity analysis was mostly restricted to eight cases from MSK-IMPACT that had RET SVs but no RET fusion transcripts on MSK-Fusion (i.e., group D cases that were false positive by MSK-IMPACT). Interestingly, all but one also showed a RET gene rearrangement (7/8) despite the lack of a RET fusion transcript, indicating that they are also false positive by FISH. Furthermore, the true positive and false-positive cases did not show a difference in their median percentage of tumor cells with rearrangements (72.5% vs. 65%, P = 0.9, Mann–Whitney U test). Notably, most of the false-positive cases (5/8) showed atypical rearrangements including loss of 5′ or 3′ signals, while only two of 48 of true RET fusion–positive cases demonstrated these atypical patterns (P = 0.003, Fisher exact test). In addition, we submitted 17 cases without RET SVs for FISH testing, and they were all negative for RET rearrangements.

When stratified by fusion partners, FISH was 100% sensitive for both KIF5B (95% CI, 80.5–100) and CCDC6 (95% CI, 76.8–100). However, NCOA4 fusions were detected with a low sensitivity (66.7%, 95% CI, 34.9–90.1) and showed a lower percentage of rearranged tumor cells compared with KIF5B and CCDC6 fusions (P = 0.005, Mann–Whitney U test; Fig. 2B).

In addition to fusion partners, sensitivity varied among the three different tumor groups (P = 0.04), with NSCLC having high sensitivity (100%, 95% CI, 85.8–100) followed by thyroid (87.5%, 95% CI, 61.7–98.4) and the remaining tumor group (75%, 95% CI, 34.9–96.8). This variation is likely driven by the different fractions of NCOA4-RET fusions in these groups given that NCOA4 was seen in 4.2%, 37.5%, and 62.5% of NSCLC, thyroid, and other tumors, respectively.

Last, we assessed the performance of RET IHC using 70 fusion-positive and 89 fusion-negative cases. Given that RET IHC is not routinely performed and its staining characteristics are not well established, we divided our cohort into a training (n = 37, fusion-positive: 19, fusion-negative: 18) and validation set (n = 122, fusion-positive: 51, fusion-negative: 71) to determine and to evaluate the optimal staining criteria, respectively. Data from the training set showed that cytoplasmic staining in ≥1% of tumor cells provided the most optimal cut-off point, with a sensitivity and specificity of 89.5% (17/19, 95% CI, 66.9–98.7) and 88.9% (16/18, 95% CI, 65.3–98.6), respectively. Using these criteria, the validation set showed a comparable sensitivity and specificity of 86.3% (44/51, 95% CI, 73.7–94.3) and 80.3% (57/71, 95% CI, 69.1–88.8), respectively. Combining the two cohorts together (n = 159), the overall sensitivity and specificity were 87.1% (61/70, 95% CI, 77–93.9) and 82% (73/89, 95% CI, 72.5–89.4), respectively (Table 2).

Among the RET fusion–positive cases (n = 70), the tumor staining pattern was strikingly consistent, with nearly all expression being restricted to the cytoplasm as reported previously (Fig. 2C; refs. 8, 24). One case, an infantile fibrosarcoma with a SPECC1L-RET fusion, had dual cytoplasmic and membranous staining. Among the RET fusion–positive cases, nine of 70 (12.9%) lacked staining, 16 of 70 (22.9%) showed at most 1+ staining, 19 of 70 (27.1%) at most 2+ staining, and 26 of 70 (37.1%) at most 3+ staining. Among the RET fusion–negative cases (n = 89), tumor cell staining was seen in 21 cases (23.6%) including five with membranous staining that was considered IHC negative. In addition, staining intensity in the fusion-negative cohort was low, with most cases (12/21, 57.1%) showing only 1+ staining. Background staining in nontumor cells included pneumocytes with apical-only staining and macrophages with globular cytoplasmic staining.

Similar to FISH, sensitivity for IHC varied significantly with the different fusion partners (P = 0.0003, Fisher exact test). Notably, IHC detected all KIF5B fusions (95% CI, 88.8–100), and these fusions had higher H-scores compared with CCDC6 (P < 0.01) and NCOA4 (P < 0.0001; Fig. 2B). In contrast, IHC missed half of the NCOA4 fusions (95% CI, 21.1–78.9), and NCOA4 fusions showed the lowest H-score among the fusion-positive cases.

Sensitivity also varied with tumor type (P = 0.002), with lung cancer showing the highest sensitivity (97.6%, 95% CI, 87.1–99.9) followed by thyroid cancer (78.9%, 95% CI, 54.4–93.9) and other tumors (60%, 95% CI, 26.2–87.8). Furthermore, lung cancers showed higher H-scores compared with thyroid cancers (P < 0.0001) and other tumors (P < 0.0001). As expected, the sensitivity and H-scores appeared to be driven by the fraction of NCOA4 cases: 2.4%, 31.6%, and 50% for NSCLC, thyroid, and other tumors, respectively. Specificity was consistently ≥80% for all three tumor types.

It is unclear whether RET amplification is associated with RET overexpression in the absence of a RET fusion. To investigate, we included 10 tumors with RET amplification lacking RET SVs in our fusion-negative cases for IHC staining. These included two breast carcinomas, two uterine serous carcinomas, gastric carcinoma, melanoma, prostatic adenocarcinoma, uterine carcinosarcoma, ovarian carcinoma, and adenoid cystic carcinoma; majority of these tumors (60%) harbored disease-specific drivers (e.g., TMPRSS2-ERG in prostate and ERBB2 amplification in breast). Among these, the two cases with the highest fold change (breast carcinoma and ovarian carcinoma with 7.6-fold and 5-fold change, respectively) showed strong IHC expression, indicating that high-level RET amplification may be associated with RET overexpression without a RET fusion. Interestingly, the staining pattern in these RET-amplified cases was unusually heterogeneous and distinct from fusion-mediated staining. Whereas fusion-positive cases showed a more uniform staining in large groups or sheets of cells, the RET-amplified cases presented with a more mosaic and checkered distribution.

In tumors with RET fusions, RET TKI directly interacts with the chimeric RET oncoproteins to suppress tumor growth. Given the critical role of RET oncoproteins, we investigated whether RET-targeted therapy remains active in a subset of fusion-positive tumors that lack robust RET expression by IHC (H-score <50). To do so, we selected 16 fusion-positive patients with H-score of <50 and identified only four patients who received selpercatinib in this cohort (Supplementary Table S4). The remaining patients (n = 7) received standard treatments (e.g., surgery, radiation, chemotherapy, immunotherapy) while others (n = 4) were treated with multikinase inhibitors (e.g., vandetanib, lenvatinib). One patient was lost to follow-up. Among the four patients who received selpercatinib, two had low expression (NSCLC with H-score of 30 and papillary thyroid cancer with H-score of 10); while the remaining showed no expression (both papillary thyroid cancers with H-score of 0). Despite the lack of robust RET expression by IHC, all four patients with RET fusions showed responses to selpercatinib. While our finding suggests that RET-targeted therapy remains active in fusion-positive tumors with low or no expression, this observation requires confirmation using a larger, prospective cohort.

Given the approval of selpercatinib and pralsetinib, oncogenic RET fusions have emerged as a must-test biomarker in patients with advanced lung and thyroid cancers (22). In this study, we provide a comprehensive characterization of RET fusions and investigate the performance of various clinical-grade assays for detecting these highly actionable fusions.

By using a combination of DNA- and RNA-based NGS, we identified a large cohort of RET fusions across various tumor types. We confirm that the majority of RET fusions involve KIF5B, CCDC6, and NCOA4 genes, all of which are located in chromosome 10 along with RET and demonstrate that their distribution varies with tissue type. Specifically, KIF5B fusions are exclusive to NSCLC, and NCOA4 and CCDC6 fusions are enriched in thyroid cancers. Interestingly, 19 of 151 (12.6%) RET fusions did not map to KIF5B, CCDC6, and NCOA4, but instead were rearranged with a wide range of other partner genes. These less common 5′ partner genes mostly mapped to other chromosomes including 2p21 (EML4), 3q12.2 (TFG), 4q13.3 (RUFY3), 4q21.1 (SHROOM3), 5q33.2 (GEMIN5), 5q35.3 (RUFY1), 6p22.3 (KIF13A), 7q33 (TRIM24), 8p12 (RBPMS), 12p13.33 (ERC1), 12q24.31 (CLIP1), 14q13.3 (MBIP), 18q21.33 (KIAA1468), 22q11.23 (SPECC1L), and Xp11.23 (GRIPAP1). Three partner genes mapped to the same chromosome as RET including C10orf118 (10q25.3), RUFY2 (10q21.3), and KIAA1217 (10p12.2). Though uncommon, these nonclassical fusions reveal the genetic diversity of RET fusions and underscore the need for diagnostic strategies that are partner agnostic.

While our finding confirms that NSCLC and thyroid cancer comprise most tumors with RET fusions, we found similar fusions in other cancers lacking mitogenic drivers including KRAS wild-type pancreatic adenocarcinoma and NTRK wild-type infantile fibrosarcoma. Preliminary data have shown that RET-targeted therapies may have pan-cancer activity (42), suggesting that RET fusions may serve as a tumor-agnostic biomarker. Furthermore, the presence of RET fusions in driver-positive tumors after targeted therapy adds to the growing evidence that RET fusions can facilitate drug resistance (15, 43, 44). Given that patients with acquired RET fusions are also sensitive to RET inhibition, screening for RET fusions in the posttreatment setting may be clinically meaningful (45). Taken together, our finding highlights the importance of RET fusion testing not only in lung and thyroid cancers but also in other driver-negative tumors and driver-positive tumors progressing on oncogene-directed therapies.

DNA-based sequencing is being increasingly used to detect RET fusions along with other molecular targets. However, its accuracy and performance for identifying RET fusions have not been rigorously evaluated despite its known limitations with fusions. In this study, we compared the results from our DNA-based MSK-IMPACT using our RNA-based MSK-Fusion as the reference. In our cotested cohort with paired MSK-IMPACT and MSK-Fusion data, MSK-IMPACT detected RET SVs in all cases with RET fusion transcripts, highlighting its utility as a sensitive screening tool. The high sensitivity is likely due to the inclusion of probes targeting RET introns 7–11 and suggests that hybrid capture DNA-based NGS panels with similar intronic coverage may provide comparable sensitivity. Unlike introns in NTRK3 or NRG1, RET intron tiling is more feasible as intron sequences are not as repetitive or long (46, 47).

In our study, the majority of the RET SVs detected from DNA sequencing were predicted to form recurrent oncogenic RET fusions with previously characterized 5′ partners (group A). A subset of these recurrent SVs was further analyzed by RNA sequencing (group B) and shown to have kinase fusion transcripts with the same fusion partner as predicted from DNA sequencing. Interestingly, in this group B cohort, a small number of DNA-level SVs and RNA-level fusion transcripts showed slight differences in their exonic breakpoints, with most cases differing by a single exon in the 5′ fusion partner. This minor discordance in breakpoints may be due to alternative splicing, particularly for SVs that were predicted to form out-of-frame fusions. Given that these differences are small, uncommon, and seen mostly in the 5′ fusion partner, the presence of these discordant breakpoints in classically oncogenic RET SVs is likely to be clinically insignificant.

Notably, MSK-IMPACT called a small group of RET SVs that were not confirmed as RET fusions on RNA sequencing (group D). These false-positive cases were all SVUS, and most did not predict a RET kinase fusion. In fact, group D was enriched in SVs that were not predicted to form any protein fusions, including SVs with intergenic regions and antisense orientations. While there were five SVs predicted to form protein fusions in this cohort, only one was predicted to form a kinase fusion with a 5′ partner gene and 3′ RET kinase domain. The remaining four SVs contained RET exons in the 5′ location and the non-RET partner genes in 3′ position. Interestingly, a considerable number of similar SVUS (12/32, 37.5%) including those with intergenic and antisense rearrangements did result in RET kinase fusions (group C), arguing that RNA sequencing may be required for clarification in cases with SVUS (Fig. 4A). While group C cases showed features that were more supportive of a potential oncogenic fusion (e.g., higher read support, absence of other mitogenic drivers, more SVs involving a 5′ partner gene and a 3′ RET kinase domain), these characteristics do not always correlate with the fusion status. As such, we recommend RNA sequencing for RET SVs with atypical or novel features.

Liquid biopsy using circulating tumor cfDNA allows for noninvasive detection of actionable genetic variants including kinase fusions (15, 48). Here, we demonstrate that RET SVs can be detected in a subset of patients using a plasma-based cfDNA NGS assay and that these RET SVs found in plasma correspond to those detected in tissue. In our small series, RET SVs were mostly identified in patients with metastatic tumors at initial presentation or during disease progression; none were identified in patients responding to therapy, with stable disease, or without disease. These findings suggest that RET fusion detection in plasma may vary with tumor load and treatment response, similar to the detection of other genetic variants using liquid biopsies (49). In addition to identifying canonical RET fusions such as KIF5B-RET, plasma testing also revealed RET SVUS that were not expressed on tissue RNA sequencing (group D). Hence, our findings suggest that liquid biopsy using circulating DNA can reveal both (i) classic RET fusions that are immediately actionable and (ii) atypical RET SVs that may require confirmatory RNA sequencing using tissue specimens.

FISH plays a critical role in fusion testing given its single-cell resolution and rapid turnaround time. As a screening strategy, it may be employed when DNA-based NGS is unavailable. Here, we show that break-apart FISH has modest sensitivity for most RET fusion partners but limited specificity. While FISH demonstrated high sensitivity for KIF5B-RET and CCDC6-RET fusions, it missed 33.3% (4/12) of NCOA4-RET fusions due to poor signal separation between the two probes. This finding is likely due to the close proximity of NCOA4 (10q11.22) and RET (10q11.21; Supplementary Fig. S1B) and suggests that cases with subtle separation patterns should be confirmed using RNA sequencing or an alternative FISH assay with an NCOA4-specific probe (Fig. 4B; ref. 50). In addition, nearly all false-positive cases on MSK-IMPACT (group D) were also false positive by FISH, resulting in a false-positive rate of 28%. Although our study was not designed to evaluate for overall specificity, our findings are consistent with false-positive rates from prior studies (40%–50%; refs. 8, 51) and highlight the potential pitfalls of relying exclusively on DNA-level events (FISH or DNA-based NGS). Given that these false-positive cases in our series were associated with atypical loss of 5′ or 3′ signals, positive FISH results with atypical patterns may benefit from further confirmation with RNA sequencing.

Previous studies have explored the utility of RET IHC using small series of fusion-positive tumors. In all, these studies have reported a wide range of sensitivity (∼50%–100%) and specificity (∼30%–90%; refs. 8, 13, 24, 52–54) likely due to differences in clones, reference methods, and positivity thresholds. Here, we also explored the performance of RET IHC using a standard monoclonal RET antibody (EPR2871) but incorporated a few important updates. Unlike prior studies, we (i) set NGS as the composite gold standard (DNA NGS for canonical RET SVs and RNA NGS for all RET fusions), (ii) divided the cohort into discovery and validation sets, (iii) calculated H-scores to integrate both the intensity and the percentage of tumor staining, and (iv) designed the study to evaluate for partner-specific expression. Through this approach, we show that while RET IHC has modest overall performance, expression varies significantly with the fusion partner, with KIF5B-RET showing diffuse and strong staining in most cases and NCOA4-RET lacking staining in half of the cases. Because prior studies using this clone have largely focused on KIF5B-RET fusions, our findings showing reduced staining in other fusion partners require further confirmation. In addition, future studies are needed to evaluate for partner-specific expression using other RET IHC clones.

The underlying mechanism for this partner-specific expression is unclear. Given that the fusion partners lie in the 5′ region, it is possible that potential differences in promoter activity and/or protein degradation signaling may be driving the variation in RET expression and stability. Interestingly, the fusion-positive tumors with reduced or absent RET IHC expression still responded to RET-targeted therapy, suggesting that the RET oncoprotein was expressed and sensitive to inhibition. In this context, another potential mechanism could be accelerated decay of chimeric mRNA leading to lower expression of RET fusion transcripts. In clinical practice, RET IHC may be best suited for identifying lung cancers with KIF5B-RET fusions for pathologic diagnosis. However, due to the variability in expression for other RET fusions and tumor types, RET IHC has limited utility as a broad screening tool. For therapy selection, all RET IHC scores should be correlated with either FISH or NGS given that the clinical trials and approvals for RET-targeted therapies were based on genetic testing results, not on IHC overexpression. Importantly, the absence of strong staining in fusion-positive tumors should not be used to withhold treatment given the responses seen in these patients.

For optimal detection of RET fusions, we recommend a testing strategy that incorporates both a screening and confirmation assay (Fig. 4). First, we recommend DNA-based NGS, if available, as a primary screening tool given its broad sensitivity for all RET fusions. All exons and introns near the RET kinase domain should be covered. Detection of canonical RET SVs (e.g., KIF5B-RET) may be immediately actionable and not require further testing. Similarly, absence of RET SVs can be regarded as a negative result. However, any atypical RET SVUS with new fusion partners, intergenic regions, or antisense fusions should be further evaluated by RNA sequencing to evaluate for the presence of a RET fusion transcript. Dual upfront DNA and RNA sequencing may represent another option; however, it is unclear whether this approach is superior to sequential testing. If DNA-based NGS is not available, we recommend screening with RET FISH. While a rearranged pattern can be regarded as a positive result, rearranged signals with atypical features may require further confirmation given the possibility of a false-positive result (group D cases). Similarly, while the absence of a clear rearranged pattern can be generally regarded as a negative result, a subthreshold split pattern may require further confirmation given the possibility of a false-negative result (e.g., NCOA4-RET fusions). As with DNA sequencing, we recommend RNA sequencing for confirmation of atypical positive and borderline-negative FISH results.

Our study has several limitations. First, our cohort was dominated by lung and thyroid cancer along with the three most common fusion isoforms (KIF5B-RET, CCDC6-RET, and NCOA4-RET). Given this bias, our findings may not be generalizable to fusions with other partner genes and those occurring in other tumors. Second, we did not use RNA sequencing to confirm the canonical RET fusions predicted by DNA sequencing. However, our results from group B confirm that these cases express fusion transcripts with the same fusion partner as predicted from DNA sequencing, indicating that RNA confirmation may not be necessary in these cases. Similarly, RNA sequencing was not consistently used to confirm the lack of RET fusions in our fusion-negative cases for FISH and IHC. To address this, we selected treatment-naïve tumors that were negative for RET SVs and positive for other mitogenic drivers on MSK-IMPACT to further minimize the possibility of selecting cases with occult RET fusions. Finally, while our RNA sequencing assay was used as the gold standard, this assay targets only a portion of the RET transcript (exons 8–13). Though this region maps to a breakpoint hotspot, there is a possibility that rare RET kinase fusions with alternative breakpoints may be missed. Interestingly, we were able to detect a CCDC6-RET fusion with a breakpoint involving RET exon 2, suggesting that RET fusions with alternative breakpoints may still be captured using our targeted assay.

In summary, we present an integrated diagnostic atlas of oncogenic RET fusions by correlating their genomic, cytogenetic, and expression-based features. Our findings highlight the utility of RNA sequencing for confirming atypical RET rearrangements from DNA sequencing and FISH. In addition, our partner-specific analysis reveals limitations of FISH and IHC for certain fusion variants, particularly NCOA4-RET. Clinical application of these insights may improve the quality and accuracy of RET fusion testing.

S.-R. Yang reports personal fees from Invitae outside the submitted work. A. Drilon reports personal fees from Loxo/Lilly, Blueprint, Helsinn, TP Therapeutics, and Ignyta/Genentech/Roche during the conduct of the study; personal fees from Bayer, Takeda/Ariad/Millenium, AstraZeneca, Pfizer, Beigene, BergenBio, Hengrui, Exelixis, Tyra, Verastem, MORE, Abbvie, 14ner/Elevation, Remedica, Archer, Monopteros, Novartis, EMD Serono, Melendi, Liebrium, Repare, Exelixis, GSK, Teva, Taiho, and PharmaMar outside the submitted work; and reports royalties from Wolters Kluwer; other from Merck, Puma, Merus, Boehringer Ingelheim; receives CME honoraria from Medscape, OncLive, PeerVoice, Physicians Education Resources, Targeted Oncology, Research to Practice, Axis, Peerview Institute, Paradigm Medical Communications, WebMD, and MJH Life Sciences. M. Ladanyi reports personal fees from Lilly Oncology and Blueprint Medicines, and grants and personal fees from LOXO oncology outside the submitted work. J.F. Hechtman reports grants from Lilly Oncology during the conduct of the study; personal fees from Illumina outside the submitted work. No disclosures were reported by the other authors.

S.-R. Yang: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. U. Aypar: Conceptualization, data curation, validation, investigation, writing-review and editing. E.Y. Rosen: Data curation, validation, investigation, writing-review and editing. D.A. Mata: Data curation, formal analysis, visualization, writing-review and editing. R. Benayed: Resources, data curation, supervision, validation, investigation, methodology. K. Mullaney: Resources, data curation, validation, investigation, methodology. G. Jayakumaran: Software, formal analysis, investigation, methodology, writing-review and editing. Y. Zhang: Resources, supervision, project administration, writing-review and editing. D. Frosina: Data curation, validation, investigation, methodology. A. Drilon: Conceptualization, resources, supervision, validation, investigation, project administration, writing-review and editing. M. Ladanyi: Conceptualization, resources, supervision, funding acquisition, project administration. A. Jungbluth: Resources, data curation, validation, investigation, methodology, writing-review and editing. N. Rekhtman: Data curation, supervision, investigation, methodology. J.F. Hechtman: Conceptualization, resources, data curation, supervision, funding acquisition, validation, investigation, visualization, writing-original draft, project administration, writing-review and editing.

This study was funded by the NCI under the MSK Cancer Center Support Grant/Core Grant (P30 CA008748) and by a research grant from Lilly Oncology (to, J.F. Hechtman).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.