CRC

CRC is often portrayed as the paradigmatic success story of GI oncology, with declining incidence in several high-income countries, established screening programs, and an expanding armamentarium of systemic therapies [3]. Yet this apparent maturity conceals profound conceptual and clinical limitations. CRC is no longer a single disease entity but a spectrum of biologically distinct malignancies that share an anatomical origin while diverging markedly in molecular drivers, immune contexture, metastatic behavior, and treatment sensitivity [8, 9]. The persistence of stage-based decision-making represents the first major caveat.

Tumor-node-metastasis (TNM) staging remains the dominant framework guiding adjuvant and metastatic treatment, despite its limited ability to capture biological risk. Patients with identical pathological stages may experience radically different outcomes, reflecting the influence of molecular features such as microsatellite instability (MSI)-high disease [10], chromosomal instability (CIN), CpG island methylator phenotype (CIMP), and consensus molecular subtypes (CMS) [8, 9]. While these classifications have refined prognostic discrimination, their integration into routine therapeutic algorithms remains inconsistent and, in many cases, superficial.

A second caveat lies in the widespread conflation of prognostic and predictive biomarkers. MSI-high status exemplifies a rare exception, conferring both favorable prognosis in early-stage disease and sensitivity to immune checkpoint inhibition in the metastatic setting [10]. In contrast, most molecular features in CRC provide prognostic information without reliably predicting benefit from specific therapies. This disconnect is particularly evident in the adjuvant setting, where chemotherapy decisions are frequently made in the absence of biomarkers capable of identifying true treatment beneficiaries [11].

The emergence of circulating tumor DNA (ctDNA) as a marker of MRD has sharpened this tension [12–14]. CtDNA robustly identifies patients at high risk of recurrence after curative-intent surgery, yet its ability to direct effective therapeutic intervention remains limited. This distinction was directly illustrated by the randomized phase 2/3 DYNAMIC-III trial in stage III colon cancer, in which postoperative ctDNA strongly stratified recurrence risk, whereas ctDNA-guided treatment modification did not translate straightforwardly into improved outcomes: de-escalation in ctDNA-negative patients reduced oxaliplatin exposure but did not meet the prespecified non-inferiority margin for recurrence-free survival, while chemotherapy escalation in ctDNA-positive patients failed to improve recurrence-free survival compared with standard management [15]. This exposes a fundamental unmet need: the absence of effective adjuvant strategies for biologically aggressive, MRD-positive CRC. DYNAMIC-III therefore provides a particularly clear example of the distinction between prognostic precision and therapeutic actionability: identifying residual disease and quantifying its associated risk does not, by itself, identify an intervention capable of modifying that risk. Thus, ctDNA has illuminated risk with unprecedented clarity while simultaneously revealing the inadequacy of current treatments to modify that risk.

Another underappreciated caveat concerns intratumoral and temporal heterogeneity. Primary tumors and metastases may harbor distinct molecular profiles, and selective pressures imposed by therapy drive clonal evolution over time [16]. Static baseline biomarker assessment therefore risks misrepresenting the biology of residual or recurrent disease. Longitudinal approaches, including serial ctDNA monitoring, may promise greater biological fidelity but introduce new layers of uncertainty related to assay sensitivity, sampling intervals, and clinical interpretation.

In the metastatic setting, therapeutic sequencing remains largely empiric. While targeted therapies for rat sarcoma viral oncogene homolog (RAS) wild-type, B-raf proto-oncogene, serine/threonine kinase (BRAF; v-raf murine sarcoma viral oncogene homolog B1)-mutant, and human epidermal growth factor receptor 2 (HER2)-amplified tumors have improved outcomes in selected subgroups, resistance invariably emerges [17]. In this context, clinicians often prioritize drug availability over biological rationale, and cross-trial comparisons substitute for direct evidence. The unmet need here is not additional agents per se, but rational, biology-driven sequencing strategies supported by adaptive trial designs.

Additional molecular and clinical determinants further illustrate the continuum between biological granularity and actionable treatment options. Kirsten rat sarcoma viral oncogene homolog (KRAS) G12C, present in a small subset of metastatic CRC, has become directly actionable: combined KRAS G12C and epidermal growth factor receptor (EGFR) inhibition has improved outcomes over standard late-line therapy, translating a formerly difficult-to-target driver into a clinically relevant therapeutic vulnerability while leaving optimal sequencing and durability unresolved [18]. Neurotrophic tyrosine receptor kinase (NTRK) fusions are exceedingly rare in CRC but can confer marked sensitivity to tumor-agnostic tropomyosin receptor kinase (TRK) inhibition, representing a paradigmatic example of high actionability despite very low prevalence [19]. Likewise, pathogenic DNA polymerase epsilon, catalytic subunit/DNA polymerase delta 1, catalytic subunit (POLE/POLD1) proofreading defects define a rare ultramutated subgroup with pronounced sensitivity to immune checkpoint blockade, although rarity and variant-level interpretation limit routine implementation [20]. By contrast, primary tumor sidedness has both prognostic and treatment-predictive relevance, particularly for anti-EGFR therapy in RAS wild-type metastatic CRC, but it remains a clinicopathological surrogate for a complex and incompletely resolved biological continuum rather than a molecular target itself [21]. These examples reinforce that precision oncology spans a spectrum from directly druggable alterations to contextual biomarkers whose clinical utility may exceed mechanistic resolution.

GC

GC remains one of the most biologically complex and clinically challenging malignancies in GI oncology. Despite declining incidence in several Western countries, it continues to represent a leading cause of cancer-related mortality worldwide, with a disproportionate burden in East Asia [1]. This geographic variability is not merely epidemiological but reflects differences in tumor biology, environmental exposure, and host factors. Yet, global treatment paradigms often assume biological interchangeability across regions, introducing a first caveat in the interpretation and application of clinical evidence.

Molecular classification has substantially refined the biological taxonomy of GC. The cancer genome atlas (TCGA) defined four major subtypes—Epstein-Barr virus (EBV)-positive, MSI-high, genomically stable (GS), and CIN—with distinct genomic architectures and candidate therapeutic vulnerabilities [22]. The Asian Cancer Research Group (ACRG) proposed a complementary classification with prognostic implications [23]. However, these taxonomies remain only partially treatment-defining: MSI/deficient mismatch repair (dMMR) has clear predictive relevance for immune checkpoint blockade, whereas most TCGA/ACRG categories still function primarily as biological descriptors rather than validated therapeutic selectors. Thus, increasingly granular taxonomy has outpaced its routine clinical use.

This distinction also clarifies which baseline biomarkers remain clinically inadequate. HER2, MSI/dMMR, programmed death-ligand 1 (PD-L1) combined positive score (CPS), and, more recently, claudin 18 isoform 2 (CLDN18.2) can alter treatment selection in defined settings; however, a single pretreatment sample provides only a static and spatially restricted representation of a heterogeneous and evolving disease. PD-L1 CPS is continuous, assay- and sampling-dependent, and enriches—rather than absolutely selects—for benefit from chemoimmunotherapy, as illustrated by CheckMate 649 and KEYNOTE-859 [24, 25]. By contrast, MSI-H/dMMR, although uncommon, identifies a substantially more immunotherapy-sensitive population and represents one of the strongest biologically grounded predictive biomarkers in GC [26]. The unmet need is therefore not simply additional baseline markers, but biomarkers with sufficient spatial, temporal, and treatment-specific fidelity to guide individual decisions.

HER2-positive GC both supports and challenges the central argument of this review. It supports precision oncology because HER2 overexpression/amplification is a validated therapeutic determinant and trastuzumab established a biomarker-selected first-line strategy [27]. At the same time, gastric HER2 expression is frequently heterogeneous and may be lost or change under therapeutic pressure, limiting the reliability of a single baseline assessment [28]. The activity of the HER2-directed antibody-drug conjugate (ADC) trastuzumab deruxtecan (T-DXd) after prior therapy demonstrates that improved drug design can partly overcome biological heterogeneity, but variable antigen expression, acquired resistance, and the need for reassessment remain clinically relevant limitations [29]. HER2 therefore exemplifies successful translation into treatment without complete therapeutic certainty.

CLDN18.2 provides a newer example in which biological identification has translated into clinically meaningful treatment options. In HER2-negative, CLDN18.2-positive advanced GC/gastroesophageal junction (GEJ) adenocarcinoma, adding zolbetuximab to chemotherapy improved progression-free and overall survival (OS) in the phase III SPOTLIGHT and GLOW trials [30, 31]. Nevertheless, CLDN18.2 expression is threshold-dependent and can be heterogeneous over time and across samples, raising familiar questions regarding assay cutoffs, biomarker stability, overlapping eligibility with immunotherapy, and optimal sequencing. Thus, even a newly validated target rapidly generates a second layer of precision-oncology uncertainty once it enters routine therapeutic algorithms.

Fibroblast growth factor receptor 2b (FGFR2b) illustrates an earlier point on the same translational continuum. In the randomized phase II FIGHT trial, bemarituzumab plus mFOLFOX6 showed a clinically promising signal in FGFR2b-selected disease, with greater benefit in tumors with higher FGFR2b expression, but the primary progression-free survival (PFS) analysis was not statistically significant and confirmatory phase III evidence remains necessary [32]. This exposure-response relationship is biologically compelling, yet it also emphasizes how target definition, expression threshold, toxicity, and prospective validation determine whether molecular resolution becomes a reproducible treatment rule.

The perioperative setting exposes additional uncertainties. Multimodal strategies combining chemotherapy with surgery have become standard for resectable disease, yet recurrence rates remain substantial [33, 34]. Pathological response is often used as a surrogate for long-term benefit despite imperfect correlation with survival, and treatment remains largely uniform across biologically diverse tumors. MSI/dMMR is particularly instructive: it carries favorable prognostic information in localized disease and may identify tumors with distinct sensitivity to systemic approaches, but its optimal integration into perioperative treatment selection remains less settled than its predictive role in advanced immunotherapy. The unmet need is therefore to use biological information prospectively to escalate, de-escalate, or substitute treatment according to risk.

Immunotherapy further illustrates the difference between enrichment and certainty. PD-L1 CPS has clinical utility and higher scores are generally associated with greater benefit from first-line programmed cell death protein 1 (PD-1) blockade plus chemotherapy, but clinically used thresholds differ across trials and regulatory settings and do not define mutually exclusive responder and non-responder populations [24, 25]. MSI-H/dMMR provides a stronger predictive signal, whereas EBV-associated biology remains intriguing but is not yet a routinely sufficient stand-alone treatment selector [26, 35]. Static immunohistochemistry therefore captures only part of a dynamic tumor-immune interaction shaped by site, sampling, and therapeutic pressure.

ADCs are expanding the therapeutic consequences of surface-antigen profiling. T-DXd has established proof of principle in HER2-positive disease [28], while ADCs directed against CLDN18.2 and other gastric-cancer-associated antigens are entering clinical development [36]. Their potential is conceptually important because payload delivery and bystander effects may broaden activity beyond the strict binary logic of conventional target inhibition. However, antigen heterogeneity, target loss, payload resistance, and uncertainty regarding optimal biomarker thresholds may reproduce—rather than eliminate—the gap between increasingly sophisticated biological selection and durable clinical benefit.

Another underexplored caveat concerns MRD. Unlike CRC, where ctDNA has begun to inform postoperative risk classification, its application in GC remains investigational [37]. Differences in tumor shedding, anatomical drainage, and assay sensitivity complicate translation. Nonetheless, the absence of reliable markers of residual disease perpetuates empirical postoperative management and delays the recognition of therapeutic failure. The unmet need is the development of longitudinal biomarkers that reflect true disease burden rather than inferred risk.

Finally, GC exemplifies the limitations of one-size-fits-all clinical trial design. Global trials often pool heterogeneous populations and increasingly overlapping biomarker-defined subsets, potentially obscuring differential benefit and complicating therapeutic sequencing. Future trials should accommodate co-expression, longitudinal biomarker change, and adapt patient classification as tumors evolve, aligning biological hypotheses with clinically meaningful endpoints rather than assuming that a single baseline label is sufficient to define therapeutic destiny.

EAC

EAC epitomizes the tension between therapeutic intensification and biological uncertainty in modern GI oncology. Over the past two decades, multimodal treatment strategies have significantly improved locoregional control and surgical outcomes, yet long-term survival remains suboptimal [38]. This section specifically addresses adenocarcinoma arising in the esophagus; esophageal squamous cell carcinoma (ESCC) and GEJ adenocarcinoma are discussed below as related but non-interchangeable entities. This distinction is important because histology and anatomic origin influence tumor biology, response to neoadjuvant treatment, and the interpretation of response endpoints.

Neoadjuvant therapy followed by surgery is a major curative-intent strategy for locally advanced EAC [39]. Pathological complete response (pCR) is consistently associated with a favorable prognosis at the individual-patient level; therefore, we do not consider pCR an invalid clinical endpoint. The relevant limitation is narrower: pCR is an imperfect surrogate for long-term treatment benefit and should not be interpreted as equivalent to biological eradication or as a stand-alone substitute for survival endpoints. A trial-level meta-analysis of 40 neoadjuvant studies including 55,344 patients found only moderate correlations between pCR and OS and concluded that pCR was not an ideal surrogate endpoint in randomized trials [40]. Thus, the conceptual inadequacy lies not in measuring pCR, but in overextending its meaning beyond what has been validated.

Accordingly, no single response marker should simply replace pCR for every purpose. For comparative neoadjuvant trials, disease-free survival (DFS) currently provides a more clinically grounded intermediate endpoint when mature OS is not yet available. An individual-patient-data meta-analysis of 4,518 patients with operable esophageal or GEJ carcinoma demonstrated strong individual-level correlations between DFS and OS and very high trial-level correlations between treatment effects on these endpoints (R2 0.93–0.96 across comparisons), supporting DFS as a validated surrogate for OS in trials of neoadjuvant chemotherapy or chemoradiotherapy [41]. OS nevertheless remains the definitive endpoint. At the individual-patient level, the more relevant unmet need is complementary postoperative risk assessment that integrates pathological response, nodal status and, prospectively, dynamic molecular measures of residual disease rather than replacing one imperfect marker with another.

The distinction from ESCC is particularly important. EAC and ESCC are not merely different histologies at the same anatomical site: they have distinct epidemiology, etiologic pathways and genomic architectures. ESCC also achieves higher pCR rates after neoadjuvant chemoradiotherapy than adenocarcinoma; a systematic review and meta-analysis reported pooled pCR rates of 32% for ESCC and 22% for adenocarcinoma after neoadjuvant chemoradiotherapy [42]. Integrated genomic analyses further show that ESCC resembles squamous carcinomas from other organs more closely than EAC, whereas EAC is characterized by prominent CIN and shares molecular features with chromosomally unstable gastric adenocarcinoma [43]. Consequently, response rates, surrogate-endpoint performance and biomarker hypotheses should not be assumed to transfer unchanged between ESCC and EAC.

GEJ adenocarcinoma requires a different distinction. Although distal EAC and some GEJ adenocarcinomas form a biological continuum, GEJ tumors are anatomically and biologically heterogeneous. The Siewert classification separates distal EAC (type I), true cardia carcinoma (type II), and subcardial GC infiltrating the junction (type III), entities that differ in histogenesis, clinicopathological characteristics, and surgical approach [44]. This heterogeneity is also relevant to trial interpretation because studies frequently pool distal esophageal and GEJ adenocarcinomas. Therefore, conclusions drawn for EAC should specify whether GEJ tumors were included and should not automatically be generalized across all junctional adenocarcinomas.

Molecular characterization of EAC has revealed recurrent alterations in tumor protein p53 (TP53), receptor tyrosine kinases, and cell-cycle regulators, together with substantial CIN [45]. However, these insights have yet to translate into consistently actionable stratification in the curative-intent setting. The relevant limitation is therefore not an absence of molecular structure, but the still incomplete conversion of molecular heterogeneity into prospectively validated perioperative treatment selection.

Immunotherapy has introduced new possibilities, particularly in the adjuvant and metastatic settings [46, 47]. The use of immune checkpoint inhibitors (ICIs) after neoadjuvant chemoradiotherapy and surgery has demonstrated DFS benefit in patients with residual pathological disease [47]. Yet, predictive biomarkers of benefit remain incompletely defined, and treatment is administered to a broad biomarker-unselected postoperative population. PD-L1 expression provides context-dependent information but does not resolve this uncertainty, again illustrating the limitations of static biomarkers in a tumor-immune landscape reshaped by multimodal therapy.

The concept of molecular residual disease is therefore complementary to, rather than a replacement for, pathological response. CtDNA can detect occult disease and has shown prognostic potential in esophageal cancer, but clinical integration remains investigational [48]. Challenges related to assay sensitivity, tumor shedding, and optimal sampling timepoints remain substantial, and no validated ctDNA-guided postoperative strategy currently exists. Future response assessment should therefore be layered: pCR or graded pathological regression for local treatment effect, DFS for clinically meaningful trial-level benefit, OS as the definitive endpoint, and longitudinal biomarkers such as ctDNA as investigational tools for individual residual-risk refinement.

This endpoint hierarchy also has implications for trial design. Future EAC studies should prespecify histology and GEJ inclusion criteria, avoid treating pCR as synonymous with survival benefit, and integrate pathological, time-to-event, and longitudinal molecular endpoints according to the clinical question being tested. Such a framework preserves the established prognostic value of pCR while avoiding claims that exceed its validated surrogacy.

PDAC

PDAC represents the most formidable challenge in GI oncology, with survival outcomes that have improved only marginally over decades despite escalating therapeutic intensity [49]. Its reputation as a biologically intractable disease is well deserved, yet this characterization risks obscuring critical conceptual shortcomings in how PDAC is studied and treated. The prevailing clinical narrative emphasizes intrinsic chemoresistance and late diagnosis, but a closer examination reveals deeper structural caveats that continue to limit progress.

One of the most entrenched assumptions in PDAC management is that treatment failure reflects an absence of actionable biology. In reality, PDAC is molecularly heterogeneous, encompassing classical and basal-like transcriptional subtypes, distinct metabolic dependencies, and variable immune landscapes [50]. However, these biological distinctions rarely inform clinical decision-making. Most molecular subtype data still come from retrospective analyses. Frontline cytotoxic therapy selection is strongly influenced by performance status, comorbidity, organ function, and anticipated tolerance. We agree that this fitness-based selection is biologically and clinically meaningful and should not itself be characterized as empirical. The residual empiricism lies instead in choosing among active regimens for otherwise eligible patients without validated tumor-specific predictive biomarkers. This gap underscores a central unmet need: integrating host fitness with tumor biology rather than treating either dimension as sufficient for precision.

The neoadjuvant and adjuvant settings illustrate the limitations of current paradigms. Combination chemotherapy has become standard for resectable and borderline resectable disease, yet recurrence rates remain high even after apparently curative surgery [51]. Pathological response and margin status are frequently used as indicators of treatment success, but their prognostic value is inconsistent. Moreover, the assumption that intensified perioperative therapy can overcome aggressive biology has not been convincingly validated [52]. Many patients complete arduous treatment courses only to experience early systemic relapse, revealing the inadequacy of anatomical surrogates for biological risk.

MRD represents a particularly unresolved frontier in PDAC. Unlike other GI malignancies, the application of ctDNA has been limited by low tumor shedding and technical challenges. Nonetheless, emerging data suggest that detectable postoperative ctDNA identifies patients with near-universal risk of relapse [53]. The caveat lies not in the prognostic signal, but in the absence of effective interventions for MRD-positive patients. As in CRC, molecular detection has outpaced therapeutic capability, exposing a void rather than a pathway forward.

In advanced disease, precision approaches already modify treatment in selected PDAC subsets and should not be overlooked. Germline testing is clinically relevant because pathogenic breast cancer gene 1/2 (BRCA1/2) and partner and localizer of breast cancer gene 2 (PALB2) variants identify homologous-recombination-deficient tumors with enhanced platinum sensitivity; cisplatin plus gemcitabine is active in germline BRCA/PALB2-mutated PDAC, while maintenance olaparib prolongs PFS in germline BRCA-mutated metastatic disease without progression on first-line platinum [54–56]. MSI-H/dMMR and NTRK fusions, together with other rare actionable alterations, can also support biomarker-directed or tumor-agnostic therapy, but their low prevalence limits population-level impact [57]. Moreover, availability of actionable molecular targets does not invariably translate into treatment delivery: testing may be delayed or incomplete, and regulatory, access, trial-availability, tissue, and clinical-timing constraints may prevent a theoretically matched therapy from being administered [58]. Thus, these successes qualify rather than invalidate our argument: the unmet need is to expand the fraction of patients in whom molecular information becomes timely, accessible, and therapeutically consequential.

Importantly, KRAS-directed therapy is beginning to change this landscape. In the phase III RASolute 302 trial, daraxonrasib, a RAS(ON) multiselective inhibitor, significantly improved OS [13.2 vs 6.6 months; hazard ratio (HR) 0.40] and PFS (7.3 vs 3.5 months; HR 0.45) vs chemotherapy in previously treated metastatic PDAC with RAS G12 mutations [59]. This result provides compelling proof that a dominant PDAC driver can be therapeutically exploited, while leaving major unmet needs regarding optimal sequencing, resistance, and earlier-line integration.

The tumor microenvironment constitutes another critical caveat. PDAC is characterized by a dense desmoplastic stroma that shapes drug delivery, immune exclusion, and metabolic adaptation. Numerous stromal-targeting strategies have failed in clinical trials, often after promising preclinical results [60]. These failures underscore the limitations of model systems that inadequately recapitulate human disease. They also highlight the danger of extrapolating biological mechanisms without accounting for systemic complexity and host factors.

Clinical trial design in PDAC further compounds these challenges. Many trials enroll biologically heterogeneous populations and rely on modest survival endpoints, limiting the ability to detect subgroup-specific benefits. Negative results are frequently interpreted as evidence of biological intractability rather than as signals of poorly aligned hypotheses and methodologies. Adaptive and biomarker-enriched trial designs remain the exception rather than the norm.

PDAC exposes the limits of a treatment paradigm driven by escalation rather than precision. The field’s most pressing unmet need is not another cytotoxic combination, but an integrated strategy that aligns molecular insight, dynamic biomarkers, and rational trial design. Without such a shift, therapeutic advances will continue to be incremental, and the biological lessons of PDAC will remain underexploited.

BTC

BTCs represent one of the most heterogeneous and conceptually challenging groups of malignancies in GI oncology. Encompassing intrahepatic and extrahepatic cholangiocarcinoma as well as gallbladder cancer, BTCs are frequently discussed as a single clinical entity despite profound differences in cell of origin, molecular drivers, anatomical context, and clinical behavior [61]. This taxonomic simplification constitutes a fundamental caveat that permeates clinical trial design, therapeutic decision-making, and outcome interpretation.

From a biological standpoint, BTCs display remarkable molecular diversity. Actionable alterations such as FGFR2 fusions, isocitrate dehydrogenase 1 (IDH1) mutations, BRAF mutations, HER2 amplification, and NTRK fusions occur at variable frequencies across anatomical subtypes [62]. These advances have ushered in a new era of systemic therapy, combining biomarker-selected targeted strategies with contemporary chemoimmunotherapy [63–66]. However, availability of actionable molecular targets is not binary. The strength of clinical evidence differs substantially across targets, and even a technically targetable alteration may not translate into treatment because of limited evidence, drug access, timing, patient fitness, or acquired resistance [67, 68]. In a prospective BTC cohort, actionable alterations were identified in 58.3% of patients, but only 11.1% received genotype-matched therapy [68]. Thus, the unmet need extends beyond detecting a target to establishing its clinical use and ensuring that matched treatment can actually be delivered.

Recent therapeutic advances illustrate both the potential and the methodological complexity of precision oncology in BTC. In IDH1-mutant advanced cholangiocarcinoma, the phase III ClarIDHy trial established the clinical benefit of ivosidenib after prior therapy; final OS was 10.3 vs 7.5 months with placebo, and adjustment for the substantial crossover yielded an HR for death of 0.49 [69]. For FGFR2-rearranged cholangiocarcinoma, pemigatinib moved from activity in previously treated disease to randomized first-line evaluation in FIGHT-302: PFS was 8.3 vs 6.8 months with gemcitabine-cisplatin (HR 0.58), with higher response rate and longer response duration, although OS was similar; notably, the trial closed early because the first-line standard changed during recruitment [70]. In parallel, immunotherapy has reshaped the unselected first-line setting: KEYNOTE-966 showed an overall-survival benefit from adding pembrolizumab to gemcitabine-cisplatin (12.7 vs 10.9 months; HR 0.83) [71], while updated TOPAZ-1 confirmed sustained benefit with durvalumab plus gemcitabine-cisplatin (12.9 vs 11.3 months; HR 0.76) [72]. These developments make cholangiocarcinoma a particularly informative model in which biomarker-defined therapies and broadly applicable chemoimmunotherapy are evolving simultaneously, raising increasingly relevant questions of molecular testing, treatment sequencing, and trial design.

The first-line systemic treatment of advanced BTC exemplifies the persistence of empiricism. Platinum-based chemotherapy has long constituted the backbone of treatment, with recent additions marginally improving survival [63]. While these advances are welcome, they apply broadly and do not address the underlying biological heterogeneity of BTCs. Moreover, therapeutic decisions are often made without molecular information, despite the growing availability of genomic profiling. This disconnect reflects both logistical barriers and uncertainty regarding the optimal timing and interpretation of molecular data.

The adjuvant setting reveals additional layers of uncertainty. Surgical resection offers the only chance for cure, yet recurrence rates remain high. Adjuvant chemotherapy is frequently administered despite limited and sometimes conflicting evidence [73]. Trials in this setting often pool disparate BTCs subtypes, obscuring subtype-specific effects and diluting potential signals of benefit. The unmet need is not further intensification of therapy, but precise identification of patients who truly benefit from postoperative treatment.

MRD has emerged as a compelling but underdeveloped concept in BTC. CtDNA analysis has demonstrated prognostic value in small cohorts, identifying patients at high risk of recurrence after surgery [74]. However, clinical integration remains limited by assay variability, tumor shedding heterogeneity, and the absence of validated intervention strategies. As in other GI malignancies, MRD detection exposes a therapeutic void: risk can be identified more readily than it can be mitigated.

Immunotherapy has yielded mixed results in BTC [63, 71, 72]. While a subset of patients derives benefit, predictive biomarkers remain poorly defined. MSI-high tumors represent a small minority, and PD-L1 expression has not reliably predicted response. The immunological landscape of BTCs is shaped by both tumor-intrinsic factors and the biliary microenvironment, yet current biomarkers fail to capture this complexity. The assumption that immunotherapy can be broadly effective without refined patient selection represents a further caveat.

Clinical trial design poses perhaps the greatest barrier to progress in BTCs. The rarity and heterogeneity of these tumors necessitate collaborative, innovative approaches, yet many trials remain underpowered and biologically heterogeneous. We consider it plausible that aggregation of anatomically and molecularly distinct BTCs may obscure subtype-specific treatment effects; however, this remains a methodological hypothesis rather than an established explanation for negative trials. Negative studies should therefore not be attributed to aggregation without direct supporting evidence. Future studies should increasingly incorporate biomarker-driven patient selection, adaptive designs, and international collaboration.

HCC

HCC occupies a singular position within GI oncology, not only because of its global epidemiological burden but also due to the unique intertwining of tumor biology and chronic liver disease. Unlike other solid tumors, HCC almost invariably arises in the context of cirrhosis or chronic hepatic injury, rendering oncological decision-making inseparable from host-related factors. This duality constitutes both the defining feature of HCC and its principal conceptual caveat.

Over the past decade, systemic therapy for advanced HCC has undergone a rapid evolution, culminating in the integration of ICIs combined with anti-angiogenic agents as standard first-line treatment [75]. These advances have reshaped expectations, yet durable benefit remains confined to a subset of patients. Importantly, HCC treatment selection is not independent of tumor biology. Contemporary staging and treatment algorithms integrate tumor burden, including tumor number and size, as well as vascular invasion and extrahepatic spread, together with liver function and performance status [76]. The remaining gap is more specifically molecular: despite extensive biological characterization, validated molecular predictors rarely determine systemic treatment choice. Alpha-fetoprotein (AFP) is a notable exception at the biomarker level: an AFP concentration ≥ 400 ng/mL identified the population in whom ramucirumab improved survival in the biomarker-selected phase III REACH-2 trial [77]. By contrast, no molecular biomarker has been sufficiently validated for routine selection among current ICI-based regimens [78]. Thus, the unmet need is not to introduce tumor biology into HCC decision-making de novo, but to convert molecular and immune biology into validated predictive tools that add clinically useful discrimination beyond established tumor- and host-based criteria.

Molecular characterization of HCC has identified recurrent alterations in pathways regulating telomere maintenance, cell cycle control, wingless/integrated signaling pathway (WNT)/β-catenin signaling, and angiogenesis [79]. Despite these insights, molecular subclassification has failed to translate into routine clinical stratification. One major caveat is the assumption that tumor-intrinsic genomics alone can guide therapy in a disease profoundly shaped by its microenvironment and host organ dysfunction. The biological behavior of HCC reflects not only malignant hepatocytes but also the fibrotic, inflammatory, and vascular milieu of the cirrhotic liver.

The interpretation of clinical trial endpoints in HCC further illustrates this complexity. OS remains the gold standard, yet it is influenced by competing risks related to liver failure and portal hypertension. Radiographic response criteria, originally developed for other solid tumors, inadequately capture treatment benefit in HCC, particularly for immunotherapeutic agents [80]. Apparent disease progression may coexist with meaningful survival benefit, while radiographic stability may mask declining hepatic reserve. These ambiguities complicate both trial interpretation and clinical decision-making.

Immunotherapy resistance in HCC highlights additional conceptual blind spots. Activation of the WNT/β-catenin [81] pathway has been associated with immune exclusion and primary resistance to immune checkpoint inhibition, yet routine testing for such features is not standard practice. Furthermore, etiological heterogeneity—viral hepatitis, alcohol-related liver disease, and metabolic dysfunction—associated steatotic liver disease—modulates immune contexture and treatment response. The tendency to treat HCC as a biologically uniform disease disregards these critical modifiers.

MRD remains an underdeveloped concept in HCC. Local therapies such as resection and ablation can achieve radiographic complete responses, yet recurrence rates are high. CtDNA and other liquid biopsy approaches hold promise for detecting occult disease, but their application is complicated by low tumor fraction and background liver-derived DNA [82]. As in other GI malignancies, the ability to identify residual risk has outpaced the development of effective preventive strategies.

Clinical trial design in HCC continues to struggle with patient selection and biologically informed classification. Broad inclusion criteria may enhance accrual but obscure biologically meaningful signals, while restrictive criteria limit generalizability. Balancing these competing priorities remains challenging. Innovative designs that integrate biomarkers, etiology, and dynamic assessment of liver function are urgently needed.