The incidence of differentiated thyroid cancer is rising; it is now the 7th most common cancer worldwide, with >800,000 new cases annually (1). Most cases are low risk with excellent outcomes. In the U.S., all-cause mortality for low-risk disease is <1% (2) and recurrence is <5% (3). Given the expectation of long-term survival, attention has shifted to avoiding overtreatment and using healthcare resources judiciously. De-escalating care in appropriately selected patients can reduce treatment-related morbidity, unnecessary resource use, and patient anxiety.

There are multiple key decision points in the management and follow-up of differentiated thyroid cancer. These include a) active surveillance vs surgery for small unifocal cancers with no lymph node involvement; b) extent of surgery; c) postoperative use of radioactive iodine (RAI); and d) surveillance intensity (Table 1). Optimal choices at each step in the care pathway depend heavily on the ability to accurately estimate prognosis in individual patients. Historically, risk estimation has relied on clinicopathologic features. The 2015 ATA guidelines stratified patients into low, intermediate, and high risk, with management recommendations according to this risk (3). The 2025 update refines this into low, low-intermediate, high-intermediate, and high risk (4). However, clinicopathologic systems explain only part of outcome variability and can lead to over- or undertreatment. Examples of this disconnect are well documented (5, 6). Active surveillance spares many patients with small, low-risk cancers from surgery, yet a minority demonstrate growth or nodal metastasis requiring delayed operation (7, 8). Similarly, although routine prophylactic central neck dissection often reveals occult nodal disease in >1/3 of patients, recurrence rates are comparable to cohorts managed without prophylactic dissection (9–11). These and other observations highlight the limits of traditional risk stratification systems: while useful for population-level guidance, they lack granularity for truly personalized care. Notably, a structurally incomplete response to initial therapy, while more common in higher risk categories, is still observed in up to 5% of low-risk patients and only ~33% of high-risk patients without distant metastases (12–15).

More precise prognostication should explain a greater share of outcome variance. Ideally, prognostication should be available early in the care pathway, where its impact is maximal. For instance, lobectomy is often chosen for localized cancers <4 cm to avoid overtreatment with total thyroidectomy, yet 20–50% subsequently require completion thyroidectomy (16–18), with a systematic review estimating 11–34% (19). This is typically due to higher-risk pathology identified only postoperatively. Completion thyroidectomy places patients at risk related to surgery, including hypothyroidism, hypoparathyroidism, and decreased quality of life, while also increasing health care burden (20, 21).

The advent of next-generation sequencing (NGS) has transformed the molecular characterization of thyroid lesions, allowing analysis of genomic and transcriptomic features, separately and in parallel. NGS-based approaches not only identify oncogenic drivers but also capture the transcriptomic programs that more closely reflect tumor phenotype (Figure 1). This forms the foundation for precision prognostication and biologically guided management of papillary thyroid cancer (PTC).

Molecular analysis, first adopted to refine malignancy risk in indeterminate fine-needle aspiration (FNA) cytology, now shows promise for improving risk stratification in confirmed cancer. Both mutational profiling (DNA) and interrogation of the transcriptome (mRNA) can augment clinicopathologic models to better tailor treatment and surveillance. If reliably obtained from FNA, these metrics could influence decisions earlier in the treatment pathway.

In this review, we explore how the evolving understanding of the transcriptome–phenotype relationship can inform individualized management of PTC, spanning preoperative, postoperative, and emerging therapeutic decision points.

The molecular landscape of PTC has been defined over decades of research, culminating in a detailed map of driver mutations and co-related features annotated by The Cancer Genome Atlas (22, 23). These efforts have demonstrated that most PTCs are driven by MAPK-pathway activation via BRAF^V600E or RAS mutations (22, 24); a minority are driven by receptor tyrosine kinase fusions including RET and NTRK fusions (25). These alterations shape phenotype, iodine avidity, and (with co-mutations) prognosis, but they incompletely explain outcome variability.

The BRAF gene encodes a key intracellular signaling protein in the MAPK/ERK pathway, regulating cellular proliferation. The BRAF^V600E mutation is common, present in 29-83% of PTC. Its presence can initiate tumorigenesis in normal thyroid follicular cells, increasing cellular proliferation, inhibiting apoptosis, and encouraging de-differentiation (25). By itself, the presence of the BRAF^V600E mutation is not prognostic (26–29). However, prognostic value improves when considered with co-mutations (e.g., BRAF^V600 + TERT, BRAF^V600 + RAS) (30, 31) and the transcriptomic state (26). Therapeutically, BRAF^V600E provides a target for systemic therapy in advanced disease (32, 33).

The three RAS isoforms (HRAS, KRAS, and NRAS) are small GTPases involved in signal transduction in the MAPK and PI3K-AKT pathways. Mutations result in constitutive activation of the MAPK and PI3K-AKT pathways, driving proliferation and growth, as well as promoting cell survival. NRAS mutations account for the majority of RAS alterations, followed by HRAS, with KRAS being least frequent. RAS- mutated PTCs tend to demonstrate follicular morphology, often behaving more like follicular thyroid cancer. Generally, RAS-mutated PTCs have higher iodine avidity and less aggressive behavior than BRAF^V600E mutated PTCs, consistent with feedback-attenuated MAPK signaling (22, 34, 35).

The TERT promoter controls transcription of the telomerase reverse transcriptase gene. Point mutations at positions –124 (C228T) and –146 (C250T) lead to upregulation of TERT transcription, leading to increased telomerase activity and maintenance of telomere length, enabling replicative immortality (36). TERT mutations are associated with aggressive forms of thyroid cancer, occurring commonly in anaplastic and poorly differentiated thyroid cancer (37, 38). Although uncommon in PTC (~10% of PTC specimens), it is consistently linked to worse outcomes, especially when coupled with BRAF or RAS mutations. Co-mutation status refines recurrence risk beyond single-gene calls (30, 37, 39).

The RET gene encodes a transmembrane receptor tyrosine kinase that normally regulates cell growth and differentiation, particularly in neural crest–derived and renal tissues. In PTC, RET is activated through gene fusions (most commonly CCDC6-RET and NCOA4-RET), which drive constitutive kinase activation and downstream MAPK signaling (40). RET fusions are found in ~5–10% in adult PTC, and more frequently in pediatric or radiation-associated PTC (41). RET is not normally expressed in thyroid follicular cells, so its activation via fusion is oncogenic and therapeutically targetable (42, 43).

NTRK refers to three homologous genes: NTRK1, NTRK2, and NTRK3, encoding the neurotrophic tyrosine kinases TRKA, TRKB and TRKC, respectively. In PTC, the most common NTRK fusions are ETV6-NTRK3, TPM3-NTRK1, and TPR-NTRK1 (33, 41, 44, 45). They occur in ~2–3% of adult PTCs but are more common in radiation-induced and pediatric cases. Functionally, NTRK fusions result in persistent activation of MAPK and PI3K–AKT signaling, driving uncontrolled proliferation, de-differentiation, and survival. Phenotypically, tumors often show solid or follicular architecture and higher iodine avidity than BRAF-like cancers. NTRK fusions are highly actionable, with dramatic and durable responses to TRK inhibitors such as larotrectinib and entrectinib (46–48).

A number of low frequency mutations and fusions have been reported in PTC. Their rarity limits stand-alone prognostic utility. However, they may add context when integrated with transcriptomic state and clinicopathology (49–51).

The presence of TP53 mutations is associated with aggressive behavior when found in PTC, and has been associated with progression to anaplastic thyroid cancer (52). EIF1AX mutations are found rarely (1-2.5%) in PTC, frequently co-existing with RAS mutations. They can also be found in benign lesions (50). The largest series reported is only 31 cases, and it was not possible to definitively describe outcomes that associated with EIF1AX mutations (53). Aberrant activation of the PI3K–AKT–mTOR pathway is known to promote cell proliferation, survival, and progression; PI3K and MAPK pathways frequently co-activate and cross-talk in advanced disease. PIK3CA mutations are more common to follicular and aggressive forms of thyroid cancer (49). PAX8-PPARG, a fusion of the PAX8 transcription factor and the PPARG adipogenesis regulator, is commonly identified in follicular adenomas, noninvasive follicular thyroid neoplasm with papillary-like nuclear features as well as follicular variant of PTC. When present in invasive thyroid cancer there are mixed reports on PAX8-PPARG on a marker of aggressiveness (54–56). ALK fusions (most commonly with STRN or EML4 in thyroid cancer) are identified in 1-3% of PTCs. There is some evidence suggesting aggressive behavior when ALK fusions are found in PTC (41, 51, 57).

Collectively, these driver events provide a foundational understanding of oncogenic initiation in thyroid cancer and have informed both preoperative diagnostic testing and targeted therapies for advanced disease. Yet, despite this detailed genomic map, mutation profiles alone fail to fully account for the biological and clinical heterogeneity of PTC, as explored in the next section (Figure 2).

Although genomic profiling has illuminated the recurrent driver events in PTC, its prognostic and predictive value is inherently limited. Most mutations represent initiating events in tumorigenesis but do not capture the downstream regulatory processes that determine phenotype, therapeutic sensitivity, or clinical outcome. For example, BRAF^V600E, while common, has inconsistent correlations with recurrence, and its high prevalence (~50%) reduces its discriminative utility. Even TERT promoter mutations, which more consistently associate with aggressive behavior, exert variable effects depending on co-mutational context. Low-frequency mutations such as TP53, EIF1AX, or PIK3CA, though mechanistically interesting, occur too rarely to inform population-level risk stratification.

Importantly, DNA-based testing provides a static snapshot of potential, not a dynamic measure of function (Figure 2). Mutations identify pathway activation capacity but not its actual transcriptional or phenotypic expression. They fail to reflect the contributions of epigenetic regulation, microRNA networks, metabolic reprogramming, and immune–stromal interactions, all of which shape tumor behavior. Intratumoral heterogeneity further complicates interpretation, as variant allele frequency can vary across tumor regions, influencing MAPK output and iodine avidity. Moreover, mutation burden does not necessarily predict signaling intensity or treatment response. For example, although BRAF^V600E mutation is commonly associated with reduced radioactive iodine (RAI) avidity, the correlation is far from perfect: a subset of BRAF^V600E-mutant tumors maintain iodine uptake (58, 59).

Collectively, these limitations illustrate that mutation profiling defines the genetic architecture of a tumor but not its functional state or biological behavior. The clinical phenotype (growth rate, differentiation, immune milieu, and therapeutic response) emerges from the integration of genetic, epigenetic, and microenvironmental factors operating at the transcriptional level. Recognizing this gap has shifted the field toward a transcriptome-based framework, where gene expression patterns reflect the realized state of the tumor rather than its latent potential. This transition from identifying oncogenic drivers to understanding the functional programs they activate marks a critical inflection point in thyroid cancer biology and underpins the emerging emphasis on the transcriptome as the key to precision prognostication.

The transcriptome–phenotype relationship provides a more accurate and comprehensive reflection of PTC biology than conventional histopathological classification or mutational profiling. Large-scale integrative studies such as TCGA have shown that gene expression–based subtypes (“BRAF-like” and “RAS-like”) better predict tumor differentiation, immune microenvironment, and therapeutic response than histologic variants (23). Gene set enrichment analysis (GSEA) further extends these insights by identifying coordinated changes across predefined groups of genes representing biological pathways or cellular functions (Figures 1, 2). Indeed, this approach was used to derive a clinically useful and biologically informative transcriptomic classifier for PTC, demonstrating that enrichment of MAPK, PI3K/AKT, metabolic, and immune signaling programs distinguishes molecular subtypes and accounts for differences in tumor differentiation, iodine avidity, and clinical behavior (26).

Single-cell transcriptomic analyses have expanded this framework, revealing distinct malignant thyrocyte phenotypes: follicular-like, partial EMT-like, and dedifferentiation-like (60). Spatial transcriptomics has also facilitated the identification of important ligand-receptor interactions (e.g., FN1–SDC4, CXCL13–CXCR5) that may be important in the transition of follicular cells to PTC cells (61). While these single-cell and spatial approaches provide invaluable mechanistic insight, they are not yet applicable in routine clinical practice. In contrast, bulk transcriptomic profiling has now reached clinical translation, with validated assays demonstrating prognostic and predictive value in guiding management of PTC (26).

Beyond mechanistic insights, transcriptomic profiling holds significant clinical utility in prognostication and treatment stratification. Gene-based assays such as ThyroSeq^® and Afirma^® have already transformed the management of indeterminate thyroid nodules by reducing unnecessary surgeries (3, 62–64). These molecular tests are fundamentally distinct in design and biological focus. For example, ThyroSeq^® is a mutation- and fusion-based panel that identifies oncogenic drivers to refine diagnostic certainty (“rule in” malignancy), whereas Afirma^® is a transcriptome-based expression classifier that differentiates benign from malignant nodules based on global gene-expression patterns (“rule out” malignancy). Both were developed for diagnostic triage in indeterminate nodules, not for prognostication once PTC is confirmed.

Newer transcriptome-derived approaches now extend molecular testing beyond diagnosis to capture biological behavior and recurrence risk (26, 65, 66), marking a conceptual shift from identifying malignancy to understanding tumor aggressiveness. Transcriptome-based prognostic classifiers are now being translated into clinical use. For example, Thyroid GuidePx^® is a validated multigene expression classifier developed to predict recurrence risk in PTC (26, 65). Unlike mutation-based panels, which primarily capture oncogenic initiation events, transcriptomic classifiers such as Thyroid GuidePx^® reflect the integrated biological state of the tumor, encompassing both tumor-intrinsic and microenvironmental signals. In validation studies, this approach demonstrated improved specificity for identifying indolent disease compared with ATA risk stratification, supporting its potential role in reducing overtreatment and informing individualized management pathways.

These insights set the stage for applying transcriptome-based classifiers to guide clinical decisions across the treatment pathway.

Although transcriptomic profiling provides biologically rich information, several limitations must be acknowledged. Some challenges are shared with genomic assays, including tumor heterogeneity, variable sample integrity, and the financial costs associated with next-generation sequencing (96). Other limitations are more specific to RNA-based approaches. RNA is less stable than DNA; it is subject to varying degrees of degradation and fragmentation before and after isolation, depending on sample storage temperature, storage time, and preservation medium (e.g., whether the tumor was snap-frozen and stored in liquid nitrogen or formalin-fixed paraffin-embedded) (97). Transcriptomic assays also generate high-dimensional data that require specialized bioinformatic processing and statistical modeling to ensure accurate quantification and interpretation. The lack of standardized procedures for transcriptomic testing compounds the complexity of transcriptomic data. Even in a centralized laboratory environment, implementing transcriptomic testing requires substantial technical expertise, rigorous assay validation, and robust quality-control processes, which may limit scalability and slow clinical adoption.

Despite these limitations, transcriptome-based classifiers have the potential to shift the entire paradigm of PTC management. This shift is overdue, as most other areas of oncology have already embraced transcriptomic profiling to guide treatment pathways. Examples of this include the OncotypeDx Breast Recurrence Score® and Mammaprint. However, multi-centre randomized controlled trials and real-world data will be required before there is broad adoption of these technologies in clinical practice. Future prospective trials should examine the clinical utility of molecular-guided treatment with reference to quantifiable outcomes including recurrence, quality of life, healthcare utilization, and cost-effectiveness.

Currently, there is at least one active prospective trial investigating the clinical utility of a molecular classifier in the postoperative phase to determine need for completion thyroidectomy and adjuvant RAI. Studies investigating the clinical utility of molecular classifiers in the preoperative phase are also planned and should yield valuable insights; guiding the extent of surgery required (i.e. active surveillance vs. hemi-thyroidectomy vs. total thyroidectomy) and potentially even selecting patients suitable for ablative approaches. Other gaps in our understanding of the molecular profiling and behaviour of PTC include the profiling of multifocal disease (do all multi-focal tumour have the same characteristics)?, profiling of micropapillary disease (is this the same disease as conventional PTC)?, and profiling of primary tumours compared to their metastases (do metastases present different profiles requiring a separate approach)?.

Besides prognostication, there are other exciting potential applications of transcriptome-based classifiers. There may be theranostic applications, where targeted molecular guided therapies are guided by molecular subtype; this has become common practice in some cancer types (e.g. hormone receptor and HER2 positive tumours in breast cancer). In the context of PTC, there are opportunities to use molecular classifiers to predict RAI sensitivity and refractoriness, redifferentiation strategies, tyrosine kinase inhibitors (general vs specific), and immunotherapy. In all, there is substantial potential for transcriptomic classifiers to integrate into and improve clinical guidelines such as ATA and NCCN guidelines.

The integration of transcriptomic insights into the management of PTC promises to refine risk stratification, personalize therapy, and reduce unnecessary interventions. Realizing this potential will require harmonizing molecular data with clinical algorithms through multicenter validation and prospective trials.