Biology-driven stratification of advanced biliary tract cancer treated with nab-paclitaxel plus gemcitabine–cisplatin: A prospective observational cohort study

Abstract

Background and Aims:

Nab-paclitaxel plus gemcitabine–cisplatin (Gem/Cis/nab-P) had promising efficacy in phase II trials for advanced biliary tract cancer (BTC) but failed to demonstrate superiority in phase III. We investigated Gem/Cis/nab-P efficacy and identified molecular subgroups with clinical benefit.

Approach and Results:

This prospective observational cohort study (NCT04871321) enrolled 119 patients with advanced BTC who received Gem/Cis/nab-P from July 2021 to December 2022. Of these, 108 were included in genomic and transcriptomic analyses of pretreatment tumor samples that met quality control criteria. Among 119 patients, 39.5% had intrahepatic cholangiocarcinoma, 37.0% extrahepatic cholangiocarcinoma, and 23.5% gallbladder cancer. Most patients had metastatic disease (68.9%). At a median follow-up of 23.7 months, the median progression-free survival was 8.3 months, and the median overall survival was 19.8 months. Grade ≥3 treatment-related adverse events occurred in 70 patients (58.8%), and dose reduction was required in 99.2%. Frequent genetic alterations were TP53 (53.7%), KRAS (29.6%), and CDKN2A (20.4%), with TP53 mutations being significantly associated with worse outcomes. Transcriptomic analysis identified 4 molecular subtypes: cholangiocyte-like, stromal, metabolic, and inflammatory-proliferative. The cholangiocyte-like subtype, marked by increased cholangiocyte markers, had the most favorable prognosis. Stromal and metabolic subtypes showed moderate outcomes, characterized by a fibroblast-rich stroma with activated angiogenesis and enriched metabolic pathways, respectively. The inflammatory-proliferative subtype had the worst prognosis, with cell cycle and inflammatory activation, and an exhausted immune microenvironment.

Conclusions:

This study demonstrated the clinical activity of Gem/Cis/nab-P in advanced BTC and highlighted that biology-driven patient stratification based on genomic and transcriptomic features may provide important prognostic information.

INTRODUCTION

Biliary tract cancer (BTC) is a rare malignant tumor of the gastrointestinal system, encompassing intrahepatic cholangiocarcinoma (ICC), extrahepatic cholangiocarcinoma (ECC), and gallbladder cancer (GBC). Although surgical resection remains the only potentially curative treatment, most patients are diagnosed at an unresectable stage, resulting in dismal survival outcomes for advanced BTC.¹ The rarity and biological diversity of BTC subtypes have impeded the development of effective therapeutic strategies and clinical trials for advanced BTC. For these reasons, gemcitabine–cisplatin (Gem/Cis) remained the standard first-line therapy for BTC since 2010.² However, the addition of anti-PD-1/PD-L1 to the Gem/Cis backbone has recently demonstrated improved survival outcomes, establishing it as the new standard of care.^3,4 Nevertheless, the overall survival (OS) for advanced BTC patients is still only ~1 year.

Another triplet combination treatment, nab-paclitaxel plus gemcitabine–cisplatin (Gem/Cis/nab-P), showed encouraging results in a phase II trial, with median progression-free survival (PFS) of 11.8 months, median OS of 19.2 months, and objective response rate (ORR) of 45%, as well as treatment-enabled conversion surgery in 20% of patients.⁵ However, Gem/Cis/nab-P failed to meet the primary endpoint in the phase III trial, as it did not show significant superiority over Gem/Cis, with a median OS of 14.0 months compared with 13.6 months for Gem/Cis (p=0.41).⁶ However, a subgroup analysis from this study revealed favorable survival outcomes in patients with locally advanced BTC and GBC,⁶ suggesting the importance of identifying subpopulations that are most likely to benefit from this regimen.

We conducted a prospective observational study of Gem/Cis/nab-P in patients with locally advanced/metastatic BTC to determine the feasibility of Gem/Cis/nab-P and to identify subpopulations that may benefit from this regimen.

METHODS

Study design and patients

A total of 119 patients with advanced BTC treated with Gem/Cis/nab-P at CHA Bundang Medical Center, Seongnam, Korea, were enrolled in our prospective study between July 2021 and December 2022 (ClinicalTrials.gov identifier: NCT04871321). Patients were eligible if they were ≥19 years of age with histological or cytological confirmation of BTC (ie, ICC, ECC, or GBC) that was unresectable, metastatic, or recurrent; were scheduled to receive Gem/Cis/nab-P; and presented with at least one Response Evaluation Criteria in Solid Tumors (RECIST; version 1.1) measurable lesion as assessed by the investigators. Additional requirements included completion of in-house next-generation sequencing, an Eastern Cooperative Oncology Group (ECOG) performance status of 0–1, an anticipated survival of ≥3 months, the ability to understand study procedures, and provision of written informed consent. Patients were excluded if they had uncontrolled or significant intercurrent illness (including infection or organ failure), prior history of systemic chemotherapy for BTC, dementia or other impairments preventing informed consent, or any condition that investigators judged as incompatible with study participation.

Treatments

Gem/Cis/nab-P consisted of 800 mg/m² gemcitabine, 25 mg/m² cisplatin, and 100 mg/m² nab-paclitaxel administered on days 1 and 8 of each 21-day cycle. The nab-paclitaxel doses were reduced at the investigator’s discretion in cases of toxicity. All participants were treated until disease progression, intolerable toxicity, or patient withdrawal. This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of CHA University (no. 2021-03-045). The full study protocol is available in Supplemental Material 1, http://links.lww.com/HEP/K338.

Targeted genomic sequencing

A total of 119 formalin-fixed paraffin-embedded (FFPE) tumor tissue samples collected before Gem/Cis/nab-P treatment underwent in-house next-generation sequencing. Eleven samples were sequenced but excluded from downstream analysis because the DNA sequencing quality was too poor (uniformity ≤70%) to reliably call alteration, resulting in a total of 108 samples being included in the genomic analysis. Genomic DNA was extracted using the RecoverAll Multi-Sample RNA/DNA Isolation Workflow (Invitrogen, Carlsbad, CA, USA) according to the manufacturer’s protocol. Targeted sequencing was performed with the Oncomine Comprehensive Assay Plus (OCA Plus, Thermo Fisher Scientific, Waltham, MA, USA), which covers 501 cancer-related genes and includes assays for microsatellite instability and tumor mutational burden. Next-generation sequencing libraries were prepared and sequenced on the Ion Torrent S5 XL platform. Raw sequencing reads were aligned to the human reference genome 19 (hg19) using Torrent Suite software (version 5.18), and variant calling was performed using Ion Reporter software (version 5.18).

To evaluate clinically significant alterations, we categorized each alteration according to the AMP/ASCO/CAP Somatic Variants guidelines, assigning them to tier I, II, III, or IV.⁷ For interpretative clarity, we also annotated each alteration with its corresponding OncoKB level (version 4.25), which ranks variants by clinical actionability: Level 1 for FDA-approved biomarkers in the same cancer type, Level 2 for biomarkers supported by clinical guidelines, Level 3 for variants with clinical evidence in investigational settings or other tumor types, and Level 4 for alterations supported by preclinical or biological data.^8,9 Tier IV alterations, classified as “neutral” or “likely neutral” in OncoKB, were excluded from all analyses.

Whole transcriptome sequencing

RNA sequencing was performed using the same FFPE tumor tissue samples that were included in the genomic analysis. Total RNA was extracted using Trizol reagent (Invitrogen, Carlsbad, CA, USA) and RNA quality assessed using the Agilent TapeStation 4000 system (Agilent Technologies, Amstelveen, the Netherlands). RNA quantification was performed using an ND-2000 Spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA).

RNA libraries were prepared using the QuantSeq 3′ mRNA-Seq Library Prep Kit (Lexogen, Vienna, Austria) according to the manufacturer’s instructions. Libraries were purified and sequenced using the Illumina NextSeq 550 platform with 75 bp single-end reads.¹⁰ Raw sequencing reads were quality-filtered using BBDuk (average Q-score >20) and aligned to the hg19 using Bowtie2.¹¹ Gene-level expression per sample was quantified using Bedtools.¹² For most analyses, read counts were normalized and variance-stabilizing transformed using the DESeq2 package (version 1.34.0).¹³ Cell type fraction quantification using EcoTyper was based on counts per million mapped reads (CPM) normalization with the edgeR package (version 3.36.0).¹⁴

To delineate transcriptomic subtypes of BTC, in the ConsensusClusterPlus package (version 1.62.0) we applied unsupervised clustering to the 2000 most variably expressed genes (selected by median absolute deviation) using the partitioning-around-medoids (PAM) algorithm with Pearson correlation as the distance metric and testing k = 2–6.¹⁵ Four robust clusters (hereafter “subtypes”) were identified based on the consensus matrix, cumulative distribution function curves, and delta‑area plots. To characterize the biology of these subtypes, we performed gene‑set variation analysis (GSVA; version 1.42.0)¹⁶ on Molecular Signatures Database (MSigDB) Hallmark gene sets obtained using the msigdbr package (version 7.5.1), generating sample‑level pathway enrichment scores.¹⁷ In parallel, we estimated cell‑type fractions in EcoTyper (R 4.1.3 implementation) from CPM‑normalized counts following the workflow described in the EcoTyper GitHub repository (https://github.com/digitalcytometry/ecotyper).^18,19 Pathway-level scores and cell‑type fractions were then averaged within each subtype, z‑score normalized, and visualized using ComplexHeatmap package (version 2.10.0) in R 4.1.3 (www.r-project.org), facilitating direct comparison of biological functions across molecular subtypes of BTC.²⁰ A schematic overview of this analytical workflow is provided in Supplemental Figure S1 in Supplemental Material 2, http://links.lww.com/HEP/K339.

Assessments

Two board-certified radiologists (Chansik An and Sanghoon Jung, with 12 and 15 years of experience in abdominal imaging, respectively) independently evaluated the tumor response according to RECIST version 1.1.²¹ The best overall response was determined, and any discrepancies were resolved by consensus after discussion. During this review by radiologists, 9 cases were judged to have no appropriate, measurable target lesion suitable for objective assessment of the response, mainly due to inadequate contrast enhancement or ambiguous delineation of the lesion. These 9 patients were excluded from the ORR and disease control rate (DCR) analyses per protocol.

The safety and tolerability of the treatment were evaluated using the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 in all patients who received at least one dose of Gem/Cis/nab-P.²²

Statistical analysis

For molecular subtype comparisons, categorical variables were assessed by the Fisher exact test, whereas continuous variables were analyzed using one-way analysis of variance (ANOVA). OS was calculated from the initiation of Gem/Cis/nab-P to death from any cause or last follow‑up, and PFS was calculated from treatment start to radiographic progression or death, whichever occurred first. Kaplan–Meier curves were generated using the survival package (version 3.5–7) in R, and group differences were evaluated by the log‑rank test.²³ Multivariable associations with survival were evaluated using a Cox proportional hazards model. All results with p<0.05 were considered significant. All computations and plotting were produced in R with support from the ggplot2 package (version 3.4.4).²⁴

RESULTS

Baseline clinical characteristics of the study population

A total of 119 patients with advanced BTC were enrolled in this prospective observational study between July 2021 and December 2022. All patients were treated with Gem/Cis/nab-P as first-line systemic chemotherapy. The median age was 65 years (IQR, 59–70 y). Of these patients, 77 (64.7%) had an ECOG performance status of 1. The primary tumor sites were ICC (39.5%), ECC (37.0%), and GBC (23.5%). Most patients had metastatic disease at diagnosis (68.9%). Detailed baseline characteristics are summarized in Table 1.

TABLE 1.

Baseline clinical characteristics

Characteristics	N=119
Median age, years (IQR)	65 (59–70)
Sex, n (%)
Male	74 (62.2%)
Female	45 (37.8%)
ECOG performance status, n (%)
0	42 (35.3%)
1	77 (64.7%)
Primary site, n (%)
Intrahepatic cholangiocarcinoma	47 (39.5%)
Extrahepatic cholangiocarcinoma	44 (37.0%)
Gallbladder cancer	28 (23.5%)
Disease status, n (%)
Initially unresectable	100 (84.0%)
Recurrent	19 (16.0%)
Disease stage, n (%)
Locally advanced	37 (31.1%)
Metastatic	82 (68.9%)
Histology, n (%)
Adenocarcinoma	114 (95.8%)
Adenosquamous carcinoma	5 (4.2%)

Abbreviation: ECOG, Eastern Cooperative Oncology Group.

Efficacy

At the time of the data cut-off (January 2, 2024), the median follow-up duration was 23.7 months (range, 0.8–30.2 months). The median PFS was 8.3 months (95% CI, 7.0–11.5) with a 12-month PFS rate of 37.8% (95% CI, 29.8%–48.0%; Figure 1A). The median OS was 19.8 months [95% CI, 15.3–not reached (NR)], with 12- and 24-month OS rates of 63.0% (95% CI, 54.9%–72.3%) and 48.5% (95% CI, 39.9%–58.8%), respectively (Figure 1B).

FIGURE 1.

Kaplan–Meier curves for progression-free survival (PFS) and overall survival (OS) in patients with biliary tract cancer (BTC) treated with gemcitabine, cisplatin, and nab-paclitaxel. (A) PFS and (B) OS in the overall BTC cohort. (C) PFS (p=6.1×10⁻⁵) and (D) OS (p=0.03) stratified by primary BTC site. (E) PFS (p=2.7×10⁻⁶) and OS (p=4.9×10⁻⁴) stratified by disease stage. Abbreviations: ECC, extrahepatic cholangiocarcinoma; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma; NR, not reached.

By tumor subtype, ECC demonstrated the most favorable survival outcomes (median PFS 13.1 months; median OS: NR), followed by ICC (median PFS 7.9 months; median OS: NR) and GBC (median PFS 4.3 months; median OS 10.7 months). These differences were significant according to primary site (p=6.1×10⁻⁵ for PFS and p=0.03 for OS; Figures 1C, D). Patients with locally advanced disease had significantly longer median PFS (23.9 vs. 7.0 months) and OS (NR vs. 15.0 months) than those with metastatic disease (p=2.7×10⁻⁶ for PFS and p=4.9×10⁻⁴ for OS; Figures 1E, F). However, because ECC included a high proportion of patients with locally advanced disease (52.3%) associated with a favorable survival outcome, we conducted a survival analysis restricted to 82 patients with metastatic disease (Supplemental Figure S2 in Supplemental Material 2, http://links.lww.com/HEP/K339). In this analysis, the primary site was not significantly different for PFS (p=0.22) or OS (p=0.54).

Among 110 patients for whom tumor response was evaluable, the ORR was 30.0% (95% CI, 22.2%–39.1%) and the DCR 85.5% (95% CI, 77.7%–90.8%). The detailed ORRs and DCRs by primary site and disease stage are provided in Table 2. Twenty-two patients (18.5%) underwent conversion surgery. Notably, among the 37 patients with initially locally advanced disease, 18 (48.6%) received curative-intent surgery.

TABLE 2.

Best overall response to gemcitabine, cisplatin, and nab-paclitaxel stratified by primary tumor site and disease stage

		Primary site			Disease stage
Response	Total cohort (N=119)	ICC (n=47)	ECC (n=44)	GBC (n=28)	Locally advanced (n=37)	Metastatic (n=82)
Best overall response, n (%)
Complete response	1 (0.8%)	1 (2.1%)	0 (0.0%)	0 (0.0%)	0 (0.0%)	1 (1.2%)
Partial response	32 (26.9%)	11 (23.4%)	10 (22.7%)	11 (39.3%)	10 (27.0%)	22 (26.8%)
Stable disease	61 (51.3%)	27 (57.5%)	21 (47.7%)	13 (46.5%)	22 (59.5%)	39 (47.6%)
Progressive disease	16 (13.4%)	7 (14.9%)	7 (15.9%)	2 (7.1%)	2 (5.4%)	14 (17.1%)
Not evaluable	9 (7.6%)	1 (2.1%)	6 (13.7%)	2 (7.1%)	3 (8.1%)	6 (7.3%)
Objective response rate, n (%) [95% CI]	33 (30.0%) [22.2%–39.1%]	12 (26.1%) [15.6%–40.3%]	10 (26.3%) [15.0%–42.0%]	11 (42.3%) [25.5%–61.1%]	10 (29.4%) [16.8%–46.2%]	23 (30.3%) [21.1%–41.3%]
Disease control rate, n (%) [95% CI]	94 (85.5%) [77.7%–90.8%]	39 (84.8%) [71.8%–92.4%]	31 (81.6%) [66.6%–90.8%]	24 (92.3%) [75.9%–97.9%]	32 (94.1%) [80.9%–98.4%]	62 (81.6%) [71.4%–88.7%]

Patients who could not be assessed for first response (n=9) due to inadequate contrast enhancement or ambiguous lesion delineation were excluded from the calculation of the objective response rate and disease control rate (per protocol n=110).

Abbreviations: ECC, extrahepatic cholangiocarcinoma; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma.

Safety and tolerability

All patients received at least one treatment cycle of Gem/Cis/nab-P and were evaluable for treatment-related adverse events (AEs). Treatment-related AEs of any grade were observed in 118 (99.2%) patients, and ≥grade 3 or higher AEs in 70 (58.8%) of patients (Table 3). The most common grade 3 or higher AE was neutropenia (42.0%), followed by elevated aspartate aminotransferase (AST; 10.9%) and thrombocytopenia (10.1%). Dose reduction was required in 99.2% of patients; only 1 patient remained on the starting dose for the duration of treatment. The mean relative dose intensity (RDI) was 62.9% (95% CI, 60.9%–65.0%) for gemcitabine, 69.8% (95% CI, 67.9%–71.7%) for cisplatin, and 61.7% (95% CI, 59.3%–64.0%) for nab-paclitaxel, resulting in a composite three-drug RDI of 62.5% (95% CI, 60.3%–64.6%).

TABLE 3.

Treatment-related adverse events

	N=119
Adverse events	Any grade	Grade 3/4
Hematological toxicity, n (%)
Anemia	106 (89.1%)	11 (9.2%)
Neutropenia	82 (68.9%)	50 (42.0%)
Febrile neutropenia	6 (5.0%)	6 (5.0%)
Thrombocytopenia	86 (72.3%)	12 (10.1%)
Non-hematological toxicity, n (%)
Nausea	78 (65.5%)	2 (1.7%)
Vomiting	19 (16.0%)	0 (0.0%)
Anorexia	58 (48.7%)	4 (3.4%)
Fatigue	37 (31.1%)	3 (2.5%)
Diarrhea	15 (12.6%)	6 (5.0%)
Sensory neuropathy	30 (25.2%)	2 (1.7%)
AST increased	60 (50.4%)	13 (10.9%)
ALT increased	56 (47.1%)	6 (5.0%)
Impaired renal function	30 (25.2%)	0 (0.0%)
Hyperbilirubinemia	11 (9.2%)	0 (0.0%)

Abbreviations: ALT, alanine aminotransferase; AST, aspartate aminotransferase.

Genomic landscape of patients with BTC

Of the 119 patients with BTC, 108 with targeted next-generation sequencing data obtained before Gem/Cis/nab-P treatment were included in the genomic analysis. We focused on the biologically significant gene alterations representing single-nucleotide variants (SNVs), insertions and deletions (indels), copy number variations (amplifications/deletions), and fusions in 30 therapeutically relevant genes: APC, ARID1A, ARID1B, ATM, ATR, BAP1, BRAF, BRCA1, BRCA2, CCND1, CCNE1, CDK4, CDKN2A, EGFR, ELF3, ERBB2, FGFR2, IDH1, IDH2, KRAS, MET, NF1, NF2, NRAS, PBRM1, PIK3CA, PTEN, SMAD4, STK11, and TP53.^25–30 The most prevalent genetic aberrations in the 30 key genes were alterations in TP53 (53.7%), KRAS (29.6%), CDKN2A (20.4%), and ARID1A (16.7%) (Figure 2A). Tier 1 therapeutic targets, as defined by the AMP/ASCO/CAP Somatic Variant guidelines, were identified in 11.1% of patients, including IDH1 mutations (6.4% in ICC, 2.3% in ECC),^31,32 FGFR2 fusions (4.3% in ICC),^33,34 KRAS G12C mutations (0.9% across BTC),³⁵ and ERBB2 amplifications (2.3% in ECC, 14.3% in GBC)^36,37 (Figures 2B, C).

FIGURE 2.

Genomic landscape of biliary tract cancer (BTC). (A) Frequency of alterations in 30 key genes among patients with BTC. Genetic alterations include non-synonymous mutations [eg, single-nucleotide variant (SNV) and insertion/deletion (indel)], copy number variations (amplifications and deletions), and gene fusions. (B) Proportion of BTC patients harboring somatic variants stratified by the highest level of clinical evidence according to AMP/ASCO/CAP guidelines with corresponding OncoKB annotations. Each patient was classified based on the most clinically significant variant detected. (C) Distribution of the genetic alterations shown in (A). Patients are arranged in order of increasing number of alterations and grouped by the primary anatomical site of the BTC. (D) Kaplan–Meier curves for progression-free survival (PFS) and overall survival (OS) in patients with mutant and wild-type TP53. (E) Kaplan–Meier curves for PFS and OS in patients with mutant and wild-type KRAS. Abbreviations: ECOG PS, Eastern Cooperative Oncology Group performance status; Mut, mutation.

Moreover, TP53 mutations were associated with worse survival outcomes in both PFS (p=9.1×10⁻³) and OS (p=1.7×10⁻³) among patients treated with Gem/Cis/nab-P (Figure 2D). In contrast, KRAS mutation was not associated with PFS (p=0.85) or OS (p=0.53; Figure 2E). Although mutations in ARID1B, PIK3CA, and STK11 were linked to worse survival outcomes, their impact warrants further investigation given the limited number of cases (Supplemental Figure S3 in Supplemental Material 2, http://links.lww.com/HEP/K339).

Transcriptome-based molecular subtypes in BTC

To identify molecular subtypes in advanced BTC, we performed transcriptomic analysis on baseline tumor samples from 108 patients treated with Gem/Cis/nab-P and examined how the molecular subtypes were related to the treatment response. To minimize the impact of noise on the gene expression data, we selected the 2000 genes with the most variably expressed transcripts and performed consensus clustering on this subset (Supplemental Figure S4 in Supplemental Material 2, http://links.lww.com/HEP/K339). We identified 4 distinct molecular subtypes of BTC (Figure 3A), which we designated according to their dominant biological features: cholangiocyte-like (subtype 1, 26.8%), stromal (subtype 2, 31.5%), metabolic (subtype 3, 28.7%), and inflammatory-proliferative (subtype 4, 13.0%). Below, we summarize the clinical, genomic, and molecular characteristics of each subtype, which are detailed in the Supplemental Results in Supplemental Material 2, http://links.lww.com/HEP/K339.

FIGURE 3.

Molecular classification of biliary tract cancer (BTC) and associated clinical, pathological, and genomic features. (A) Consensus clustering matrix of 108 patients with BTC based on the top 2000 genes ranked by median absolute deviation using k (number of clusters)=4. The corresponding heatmap shows clinical characteristics and genomic alterations that are either significant (p<0.05) or classified as clinically actionable (Tier I alterations according to AMP/ASCO/CAP guidelines). (B) Kaplan–Meier curves for progression-free survival (PFS) and overall survival (OS) across the identified BTC molecular subtypes. Significance was evaluated using the log-rank test. (C) Distribution of BTC subtypes according to primary BTC site and disease stage. (D) Bar plot showing genomic alterations that differ significantly among molecular subtypes of BTC (p<0.05). (E, F) Forest plots of multivariable Cox proportional hazards regression in 107 patients with advanced BTC. One patient without a baseline CA19-9 was excluded. Significance in panels (A), (C), (D), and (E) was assessed by one-way ANOVA for continuous variables and the Fisher exact test for categorical variables; ns, not significant, *p<0.05, **p<0.01, ***p<0.001, and ****p<0.0001. Abbreviations: Amp, amplification; CA19-9, carbohydrate antigen 19-9; ECC, extrahepatic cholangiocarcinoma; ECOG, Eastern Cooperative Oncology Group; GBC, gallbladder cancer; ICC, intrahepatic cholangiocarcinoma.

Subtype 1

Subtype 1 showed the most favorable survival outcomes with Gem/Cis/nab-P treatment (Figure 3B). Although this subtype was comprised mostly of ECC (79.3%) and locally advanced BTC (69.0%) (Figure 3C), subtype-specific mutations were lacking among the 30 key genes (Figure 3D). At the transcriptomic level, we found no biologically interpretable pathways among 50 Hallmark gene sets; however, these tumors consisted largely of normal-like epithelial cells that markedly overexpressed cholangiocyte markers (KRT19, CFTR, and MUC1; Figures 4A–C). This highlights the preservation of normal biliary identity and supports the designation “cholangiocyte-like.”

FIGURE 4.

Transcriptomic characteristics of molecular subtypes in advanced biliary tract cancer. (A) Heatmap showing the subtype-specific enrichment of Hallmark gene sets. Gene-set variation analysis (GSVA) scores were averaged within each subtype and transformed into z-scores. Higher scores (red) indicate stronger pathway enrichment, whereas lower scores (blue) indicate depletion. (B) Heatmap displaying the relative abundance of inferred cell types from EcoTyper, averaged within each subtype and z-score normalized. Yellow indicates higher relative abundance, and purple indicates lower abundance. (C) Boxplots of normalized expression levels of select marker genes for cholangiocytes, cancer-associated fibroblasts (CAFs), bile acid metabolism, immune checkpoint, macrophage/T cell/regulatory T cells (Tregs), and proliferation across molecular subtypes. All statistical comparisons were performed using one-way ANOVA; ns, not significant, *p<0.05, **p<0.01, ***p<0.001, and ****p<0.0001. In (A) and (B), only Hallmark gene sets, or cell types with p<0.05, are shown; bold labels indicate features of particular importance within each subtype. Abbreviations: GSVA, gene set variation analysis; CAF, cancer-associated fibroblast; DN, downregulated; UP, upregulated; UV, ultraviolet; TGF-β, transforming growth factor beta; TGF-α, transforming growth factor alpha; NF-κB, nuclear factor kappa B; IL, interleukin; JAK, Janus kinase; STAT, signal transducer and activator of transcription; Treg, regulatory T cell; CTL, cytotoxic T lymphocyte; Th1, type 1 T helper cell; Th2, type 2 T helper cell; G2M, G2/M checkpoint; OXPHOS, oxidative phosphorylation; ECM, extracellular matrix.

Subtype 2

Subtype 2 was associated with intermediate survival outcomes (Figure 3B). At diagnosis, 79.4% of cases were metastatic (Figure 3C). These tumors indicated low tumor purity (Supplemental Figure S5 in Supplemental Material 2, http://links.lww.com/HEP/K339) with pronounced cancer-associated fibroblast (CAF)-driven angiogenic pathways (Figure 4A). This subtype was dominated by tumor-associated endothelial and CAF1 populations and indicated increased expression of CAF markers (ACTA2, FAP, and TAGLN; Figure 4B). They also exhibited an immune milieu of immature T and B cells (Figure 4C). These features define a stromal subtype.

Subtype 3

Subtype 3 also exhibited moderate PFS/OS and was predominantly ICC (71.0%), with therapeutic targets such as IDH1 mutations and FGFR2 fusions (Figures 3A–C). Notably, this subtype was driven by metabolic programs, such as the bile acid, fatty acid, and oxidative phosphorylation pathways, and was characterized by an increased presence of epithelial cells with a metabolic signature and significant upregulation of bile acid metabolism genes (CYP7A1, CYP8B1, CYP27A1, and ABCB11; Figures 4A–C). These data classify it as a metabolic subtype.

Subtype 4

Subtype 4 had the worst outcomes, independently predicting poor PFS and OS (hazard ratio=6.07 and 6.10; Figures 3B, E, F; Supplemental Table S1 in Supplemental Material 2, http://links.lww.com/HEP/K339), and it had the lowest DCR to Gem/Cis/nab-P (Supplemental Table S2 in Supplemental Material 2, http://links.lww.com/HEP/K339). These high-purity tumors frequently carried TP53 (85.7%) and ARID1B (21.4%) mutations (Figure 3D) and exhibited enrichment of cell cycle pathways and inflammatory responses, such as IL6/JAK/STAT3 and NF-κB-mediated TGF-α signaling (Figure 4A). They also had an immunosuppressive microenvironment characterized by the presence of exhausted CD8⁺/CD4⁺ T cells and M2-like macrophages (Figure 4B). At the gene level, we confirmed the activation of immune checkpoint molecules (CD274, CTLA4, and LAG3), macrophage/T cell/Treg cell markers (MRC1, GATA3, and IL2RA), and cell-cycle proliferative markers (PCNA, CDK1, and MCM2), and the inactivation of cholangiocyte markers (KRT19, CFTR, and MUC1; Figure 4C), defining an aggressive inflammatory-proliferative subtype.

Among the 4 distinct molecular subtypes, subtype 1 and subtype 4 were associated with the most favorable and the poorest survival outcomes, respectively (Figure 3B). However, because subtype 1 consisted predominantly of locally advanced BTC (69.0%; Figure 3C), which is associated with favorable survival, this uneven distribution of disease stages across subtypes may have introduced confounding effects that contributed to the favorable outcomes observed in subtype 1. To further address this potential confounding, we conducted a survival analysis restricted to 73 patients with metastatic disease at initial diagnosis (Supplemental Figure S6 in Supplemental Material 2, http://links.lww.com/HEP/K339). Subtype 1 continued to demonstrate the most favorable OS (p=1.5×10⁻⁴), whereas subtype 4 consistently had the worst outcomes for both OS (p=1.5×10⁻⁴) and PFS (p=4.4×10⁻⁷), supporting the robustness of our molecular stratification even after excluding locally advanced cases.

DISCUSSION

This is the first prospective observational study integrating comprehensive genomic and transcriptomic profiling to define molecular subtypes in BTC patients treated with the triplet regimen of Gem/Cis/nab-P. We identified 4 distinct transcriptomic subtypes (cholangiocyte-like, stromal, metabolic, and inflammatory-proliferative) with divergent clinical, genomic, and tumor microenvironment characteristics (Supplemental Figure S7 in Supplemental Material 2, http://links.lww.com/HEP/K339). Importantly, these subtypes demonstrated marked differences in their response to Gem/Cis/nab-P, supporting their clinical relevance. By linking molecular features to therapeutic outcomes in a homogeneously treated cohort, our study offers a novel, biology-driven framework for patient stratification in advanced BTC.

In this study, the observed median PFS of 8.3 months and median OS of 19.8 months suggest meaningful improvement compared with historical outcomes of Gem/Cis doublet therapy, indicating the potential clinical benefit of adding nab-paclitaxel. The SWOG S1815 trial, a randomized phase III study conducted in patients with metastatic or recurrent BTC, did not demonstrate a significant improvement in OS or PFS with the addition of nab-paclitaxel to Gem/Cis (median OS 14.0 months; median PFS 7.5 months).⁶ In contrast, our study found a longer median OS of 19.8 months, which may be attributed to differences in study design, as this was a prospective observational study conducted at a single high-volume center with active multidisciplinary care, a differently selected patient population, and proactively implemented dose modification strategies in real-world practice. Our results are also comparable to those of a phase II trial conducted at a high-volume center in the United States,⁵ which reported a median PFS of 11.8 months and a median OS of 19.2 months. The proportion of patients with locally advanced disease was also higher in our cohort (31.1%) than in that of S1815 (26%),⁶ and 48.6% of our patients with locally advanced disease underwent conversion surgery, which may have favorably influenced survival outcomes. Interestingly, GBC was associated with the poorest prognosis in our study, whereas the S1815 trial reported relatively better survival outcomes for GBC compared with other anatomical subtypes. This discrepancy highlights the inherent limitations of using anatomical subtypes alone to predict the treatment response. Accordingly, a biologically informed classification, such as the molecular classification proposed in our study, may play an important role in overcoming clinical heterogeneity and guiding more precise treatment strategies.

Treatment-related AEs were reported in nearly all patients in this study, with AEs of grade 3 or greater occurring in 58.8% of cases. Hematological toxicities were particularly common, including anemia (89.1%), thrombocytopenia (72.3%), and neutropenia (68.9%), demonstrating a pattern generally consistent with the SWOG S1815 trial.⁶ In that study, the incidence of AEs ≥grade 3 was higher in the nab-paclitaxel arm (60%) compared with the standard Gem/Cis arm (45%), and treatment-related deaths were also more frequently observed (7 vs. 1 cases). In contrast, no treatment-related deaths occurred in our study, and most AEs were manageable. A key distinction of our study was the implementation of a more proactive dose reduction strategy from the beginning of treatment. Dose reductions were applied in 99.2% of patients, with a composite three-drug RDI of 62.5%, which may have contributed to the relatively lower incidence of high-grade AEs compared with S1815. These findings highlight the importance of flexible dose adjustment tailored to individual patient characteristics to ensure treatment continuity and safety when using nab-paclitaxel–based combination therapy. Furthermore, to facilitate broader clinical adoption of this triplet regimen, additional studies are warranted to determine the optimal dosing and scheduling strategies that can maintain efficacy while minimizing toxicity.

Given the limitations of current anatomy-based classification,³⁸ precise molecular classification of BTC is essential to understanding its underlying biology. Although recent efforts have proposed molecular subtypes of BTC, most of them do not reflect the limitations of previously widely used platforms (eg, microarray) and remain restricted to BTCs arising from specific anatomical locations.^{25,27,39–42} Therefore, we present a comprehensive molecular characterization of a large BTC cohort based on RNA-sequencing and identify 4 distinct molecular subtypes. As BTC tumors often present with ambiguous anatomical localization, which can result in misclassification when relying only on anatomy-based systems, our site-independent molecular classification offers a robust and clinically valuable alternative.

The cholangiocyte-like subtype is characterized by an epithelial-dominant microenvironment with preserved cholangiocyte identity. The stromal subtype shows a microenvironment rich in CAFs and endothelial cells, with an upregulated angiogenic pathway and immature lymphoid infiltration. The metabolic subtype features epithelial cells with a metabolic signature and minimal stromal or immune engagement. The inflammatory-proliferative subtype harbors high tumor purity and an immunosuppressive milieu dominated by exhausted T cells and tumor-associated macrophages, as well as an enriched cell cycle-related pathway.

Among these subtypes, the inflammatory-proliferative subtype presented with the worst survival outcomes (Figures 3B, E, F), highlighting the potential roles of chronic inflammation and immune exhaustion in therapeutic failure. Notably, BTCs harboring TP53 mutations also had poor survival when treated with Gem/Cis/nab-P (Figure 2D), suggesting that this regimen may not benefit patients with these mutations and should be reconsidered in this genetic context. Considering this, our multicenter phase 1b/2 trial of a PD-1 inhibitor in combination with Gem/Cis and low-dose nab-paclitaxel (NCT06893380) may provide a rational strategy for overcoming immune evasion in this high-risk group. These findings highlight the clinical importance of transcriptome-based subtyping in BTC, which may provide a framework for biology-guided therapeutic stratification, particularly for Gem/Cis/nab-P treatment.

This study has several limitations. Although several potential therapeutic strategies are suggested for each molecular subtype based on pathway activity, tumor microenvironment, and marker gene expression (Figures 4A–C), further studies and clinical trials are needed to validate these findings. We aimed to compare our transcriptome-derived subtypes with the molecular classifications of BTC reported previously, but such comparisons were limited by the use of microarray-based platforms in prior studies or by the restricted accessibility of the corresponding datasets.^26,27,40,42 Nevertheless, high-resolution, next-generation RNA sequencing revealed transcriptomic subtypes with distinct features that may represent an extension of previously reported classes. However, although the identified molecular subtypes were associated with outcomes after Gem/Cis/nab-P treatment, the single-arm design precludes their establishment as predictive factors. These findings may be more likely to reflect prognostic associations and will require validation in comparative cohorts to clarify their predictive relevance. We also recognize that durvalumab plus Gem/Cis is the current standard of care for first-line treatment of advanced BTC.⁴ Further evaluation of our classification in patients treated with this regimen is warranted.

In conclusion, this prospective observational cohort study (NCT04871321) demonstrated that the triplet regimen of Gem/Cis/nab-P yields favorable survival outcomes in patients with advanced BTC, resulting in a median OS of 19.8 months, even with substantial dose reductions. This study also proposes a transcriptome-based BTC classification framework independent of anatomical sites, which may complement traditional classifications and support treatment decision-making.

Supplementary Material

hep-84-726-s001.docx ^{(40.6KB, docx)}

hep-84-726-s002.xlsx ^{(41.2KB, xlsx)}

hep-84-726-s003.docx ^{(4MB, docx)}