Historically, recurring histotype-specific translocations have been reported in hematological malignancies and soft tissue tumors (STTs) and a growing number of gene fusions have been identified also in different types of carcinomas [1]. Such fusions not only drive tumorigenesis but also represent a powerful tool for diagnosis and potential targets for personalized therapy, with several drugs now constituting standard-of-care across malignancies. Fluorescence in situ hybridization (FISH) and reverse-transcription polymerase chain reaction (RT-PCR) are currently recognized as the standard methods for detecting gene fusions in formalin-fixed paraffin-embedded (FFPE) specimens, since they are affordable and widely distributed. Nevertheless, these technologies have a number of limitations, above all if the partner gene is unknown: since they test either single gene break apart (FISH) or specific fusions (FISH and RT-PCR), whenever several fusions of interest are present the analysis would result in time- and cost-consuming. Immunohistochemistry (IHC) has been recently introduced as a surrogate diagnostic test, being able to reliably detect fusion events due to chimeric protein overexpression [2–4], but has been approved for clinical use only in peculiar settings [5].

Although FISH, IHC and/or RT-PCR are still the most frequently used diagnostic adjunct (i.e. EWS RNA binding protein 1 (EWSR1)-friend leukemia virus integration 1 (FLI1) fusion in Ewing sarcoma), the introduction of next-generation sequencing (NGS) technologies has dramatically improved the scenario, particularly in cases requiring multiple testing, where NGS is cost-effective in identifying known and novel fusions genes [6, 7]. NGS analysis requires peculiar skills and dedicated platforms, which are currently limited to reference labs albeit available at ever-lower costs.

On the other hand, the “Achille’s heel” of NGS workflow for diagnostic purposes is represented by issues related to the quality of DNA/RNA and by computational analysis. Several pre-analytical variables can in fact negatively affect the library quality, including the handling of the surgical specimens and RNA fragmentation related to formalin-fixation (as it happens, moreover, for RT-PCR). Both issues negatively impact nucleic acids quality, introducing unpredictable biases that ultimately lead to low-coverage and/or low-quality reads [8]. Beyond nucleic acids quality, bioinformatic analysis of fusion transcripts is also challenging since all existing methods suffer, at least in part, from poor prediction accuracy. To improve the predictive power, it has been suggested that the results from at least two algorithms should be evaluated [9]. However, combining more approaches also presents disadvantages, since it is computationally expensive and, while improving sensitivity, can affect specificity.

Herein, we report the results obtained in a consecutive series of FFPE tumor specimens analyzed by NGS RNA-sequencing (RNAseq) for the identification of fusion transcripts relevant for diagnosis and/or treatments. The RNA was processed using a commercial assay, the FusionPlex^© solid tumor kit (for sarcomas or for carcinomas), and analyzed using the manufacturer’s software package ArcherDx^© Analysis suite (ADx). In order to assess the efficacy of integrating multiple fusion-detecting algorithms in a real-world dataset, we used two alternative tools to ADx, i.e. Arriba (ARR) [9] and spliced transcripts alignment to reference (STAR)- fusion (SFU) [10], which have been both recently recognized as highly accurate in fusions detection [11].

In this study, we report a cohort of solid tumor samples investigated for diagnostic purposes by RNAseq-based methodology in the context of the Institutional Molecular Tumor Board at the IRCCS Istituto Nazionale Tumori in Milan. The tumor samples have been analyzed by two different computational approaches, aimed at gene fusions detection for precision medicine. The first has been implemented using the ADx tool, provided by the manufacturer of the assay, and was run under stringent parameters to reduce false positive results. The second one was run to allow fusions detection also under “unfavorable” conditions (e.g., lower quality or lower number of reads), which are common in the case of FFPE samples. To this aim, two among the most commonly used tools for the investigation of chimeric transcript (ARR and SFU) have been chosen. Moreover, two different panels were used for cDNA libraries construction and have been analyzed herein: the FusionPlex^© sarcoma panel was applied in a number of STTs mainly for diagnostic purposes, while the FusionPlex^© lung panel was specifically used in clinical routine for the identification of molecular targets for precision medicine. To our knowledge, this represents the first comparison of different bioinformatics pipelines for the detection of fusion events for personalized treatment in a large real-world tumor cohort in the context of institutional Molecular Tumor Board. RNAseq data were furthermore compared with FISH, whenever available.

Overall, our results proved that the analysis performed by the manufacturer’s informatic pipeline on the sarcoma panel data was even more accurate and powerful than FISH (thanks to the possibility of capturing the partner loci of the tumor-associated genes), without any loss of specificity. On the other hand, the ADx pipeline on the lung panel highlighted that samples with low coverage are at higher risk of loss of sensitivity. For this reason, we deemed that the fundamental for the clinical practice to reinforce the output is obtained by a commercial pipeline (ADx) with those obtained by robust, publicly available, tools (such as ARR and SFU). It is nonetheless fundamental to underline that all the procedures described in the present study needed careful check by human resources, either by automated bioinformatics pipeline or by manual curation (i.e. visualizing outputs using tools like Integrated Genome Viewer in order to confirm the effectiveness of the identified fusion-associated reads).

To discuss our findings more in detail, all the fusions identified by ADx in sarcoma panel were in keeping with FISH data (when available), or IHC, or at least coherent with patient clinical features. Interestingly, three cryptic fusions (all missed by FISH) involving SS18 and CIC were identified by ADx, suggesting an increased sensitivity of NGS. These results in particular strongly support the idea that the ADx constitutes a reliable tool for the detection of STT translocations in FFPE samples. Furthermore, our data suggest that the sensitivity was independent of sample quality, even whenever the quality metrics of the libraries felt just below the cut-off values according to the manufacturer’s criteria. This is in all likelihood granted by the intra-tumor molecular homogeneity [30] in the presence of high tumor cellularity in STT samples, which overall allows unraveling gene fusions even in more critical conditions. The ADx high sensitivity coupled with the independence from reads count supports the speculation that also the 14 negative STT samples were, actually, true negative. In line with this idea, all the performed FISH did not identify any rearrangement, supporting the high specificity of the pipeline. Interestingly, a large ALK deletion (Δ2-18) was evidenced by ADx in an EWSR1-TFCP2 positive epithelioid rhabdomyosarcoma (in which IHC showed ALK overexpression and FISH the presence of an unbalanced ALK translocation characterized by the loss of the 5’ centromeric probe and the maintenance of 3’ telomeric probe. Similar ALK large deletions coupled with its over-expression were identified in a large series of TFCP2-translocated epithelioid rhabdomyosarcomas which, different from our case, did not show ALK gene rearrangement by FISH [31, 32]. Interestingly, unbalanced ALK translocation together with large deletion (Δ2-17) and ALK IHC over-expression was reported in a patient with systemic anaplastic lymphoma [33]. Combining our evidence with published data, we hypothesized that the observed ALK over-expression could be more likely due to genomic rearrangement at the ALK gene locus rather than gene fusion.

Previous reports, in the literature, indicated the occurrence of false negative results (e.g., 10% effective fusions were not identified in a set of 81 sarcoma samples analyzed by and ADx documented by Racanelli et al. [12]). For this reason, we tried to check the presence of chimeric transcripts, undetected by ADx, by applying two additional bioinformatic tools, ARR [9] and SFU [10]. To explain briefly the difference among the tools, ADx detects gene fusions by annotating the de-novo assembled RNA with basic local alignment search tool (BLAST) (ADx Analysis User Manual). This strategy is based on the assembly of short RNA sequences that are reciprocally contiguous in a full transcript (the “de novo assembly” step) and by their alignment (BLAST analysis) to the reference genome. Only the most abundant transcripts will be assembled [34]. To potentially increase the overall reliability, the data were re-analyzed with ARR and SFU. Both the tools are designed on STAR-seq aligners [35] and based on the “align-then-assemble” approach, which first aligns short RNA sequences reads to the genome (aware of the possible splicing events), and then reconstructs the transcripts. Differently from de novo assembly, this method can detect also less abundant transcripts [34]. Both ARR and SFU were reported to be among the most accurate (and fastest) methods for fusion detection on cancer transcriptome [11]. Surprisingly, among the 14 fusions identified by ADx, only 8/14 (57%) and 1/14 (7%) were confirmed by ARR and SFU respectively (only 1 fusion was shared between all methods); conversely, the major part of the new chimeric transcripts identified by ARR and SFU were unknown/unpublished. In contrast with previously reported data [11], our results showed a lower sensitivity of both methods in comparison with ADx. These findings prevent the use of these pipelines in diagnostic activity. However, in one sample not evaluable by ADx, despite characterized by high coverage, both ARR and SFU methods identified a fusion (RREB1-MKL2) that was missed by ADx and confirmed by FISH and RT-PCR. This result therefore suggests that, in peculiar cases, ARR and SFU can contribute to the detection of ADx-missed fusions, and prompts to further validations (FISH or PCR-based).

The FusionPlex^® lung panel has been applied for identifying a molecular target in a series of 162 FFPE samples obtained from 159 patients with solid cancers (mainly lung adenocarcinoma, but also various types of carcinoma). In this series, ADx identified 29 (18.2%) events, including 21 (13%) fusion transcripts and 8 (5%) exon skipping. In line with the manufacturers’ guidelines that suggest 5 × 10⁵ minimal reads count, a decreased sensitivity was observed in samples with low coverage. We are well aware of this read counts limitation: in clinical practice, however, this amount is not always achievable in FFPE samples from CNB or in cytology specimens that frequently present limited tumor area. Notably, in our series, a drop in ADx sensitivity was observed below 1.75 × 10⁵ reads. Below this threshold, the positive ADx outputs need to be confirmed by FISH and/or RT-PCR, while negative ones should be considered as not reliable. This conclusion is in line with data reported by Heydt and colleagues [36] on 18 fusion-positive cases, where ADx demonstrated high specificity with decreased sensitivity in samples with low-quality RNA. A similar conclusion was drawn in a series of FFPE ROS1-rearranged samples [37]. Indeed, not only pre-analytical and analytical conditions, but also the biology of carcinomas could contribute to increasing the percentage of false-negative, since it is well known that carcinoma, differently from STT [30], often express chimeric transcripts at low levels [1]. Other evidence is present in the literature that supports this, e.g., the comparison of RT-PCR with an NGS pipeline for the detection of EML4-ALK fusions in non-small cell lung cancer FFPE samples [38].

In addition to decreased sensitivity in samples with low coverage, ADx missed two ROS1 and one RET un-balanced translocations. These cases warrant a separate discussion. All of them showed “atypical” ROS1 and RET FISH patterns, with one fusion signal and one isolated 3’ signal without the corresponding 5’ signal. To our knowledge, to date, it is not yet clear if this pattern might be coupled with increased ROS1 or RET expression (despite the ROS1 IHC negative data). Although, according to ROS1 and RET FISH guidelines, such a pattern should be considered positive [27, 28], those events can be due to deletions involving the 5’ portion of the genes locus but cannot be automatically associated with the presence of gene fusion. It is worth mentioning that also both ARR and SFU output the absence of the two ROS1 and one RET translocations. The reason for this discrepancy between RNAseq, FISH, and IHC deserves further investigation.

Finally, as described for sarcoma panel, lung panel-ADx data were re-analyzed by ARR and SFU with higher overall inter-pipeline agreement than that observed in STT samples. In particular, ARR confirmed 18/21 (85%) ADx fusions and 90/120 ADx negative samples whereas, similarly to what was observed in STT, SFU confirmed only 7/21 (34%) translocations. In FusionPlex^© lung panel, therefore, ARR sensitivity was comparable to ADx one, leaving open the question of why ARR exhibited a better performance when applied with lung panels than sarcoma panel.

Despite the limited number of cases included in our cohort and the lack of a complete FISH validation, our data, obtained in a real-life series of FFPE samples, strongly support the use of NGS for the detection of fusion transcripts also in the presence of low-quality material. On the other hand, in samples that express chimeric genes at very low level (i.e. in lung carcinoma), we recommend to support the obtained results with other approaches, such as FISH and/or RT-PCR, provided that ADx, to date, represents a valuable and accurate pipeline for fusion detection and that the integration of more bioinformatic tools should be exploited to reinforce the workflow, particularly in the analysis of low-quality samples. Overall, our findings highlight the increasing need of structured bioinformatics support within public hospitals and standard diagnostic procedures.