Multiomics and machine learning in cholangiocarcinoma: A paradigm shift toward molecular subtyping and precision therapy

Cholangiocarcinoma (CCA) is one of the most intractable malignancies in hepatobiliary oncology, characterized by marked molecular heterogeneity, strong therapeutic resistance, and extremely poor clinical prognosis.1 For advanced patients receiving first-line gemcitabine plus cisplatin therapy, the median overall survival is ~1 year. In addition, targeted drugs are only applicable to a small minority of patients harboring FGFR2 fusions or IDH1 mutations, and responses are often short-lived.2,3 On one hand, CCA research has mostly been classified based on anatomical location, which masks molecular-level heterogeneity. On the other hand, previous multiomics studies have often been limited to a single subtype and lacked in-depth integration of multi-dimensional data, making it difficult to explore potential therapeutic targets and the basis for precision therapy. Against this backdrop, the study by Mun and colleagues offers a transformative vision for how we might reclassify, understand, and ultimately treat this devastating disease.4 They profiled tumors across all anatomical subtypes and addressed the issue of CCA molecular heterogeneity through multiomics integration combined with machine learning: The study integrated whole-exome sequencing, transcriptome, proteome, and phosphoproteome data from 43 CCA patients. By using MultiOmics Factor Analysis (MOFA), 15 heterogeneity dimensions were identified. Subsequently, through Hierarchical All-against-All (HAllA) clustering, 3 molecular subtypes independent of anatomical location were defined for the first time, namely the immune-regulatory subtype (cluster 1, enriched in immune-related pathways), the metabolic subtype (cluster 2, defined solely by protein/phosphoprotein characteristics and with the worst prognosis), and the gene regulation/cell fate subtype (cluster 3, enriched in gene regulatory pathways). Although traditional anatomical classification can guide surgical strategies and determine trial enrollment criteria, it fails to capture the profound molecular heterogeneity within tumors and among different categories. In contrast, molecular subtypes are associated with prognosis and treatment response but not with anatomical location, demonstrating that molecular subtypes are more capable of achieving prognostic stratification and treatment response assessment. It offers a more clinically relevant stratification, linking molecular subtype to survival outcomes and therapeutic sensitivity. For an extended period, cholangiocarcinoma research has relied on anatomical location-based classification, which divides the cancer into 3 subtypes: intrahepatic (iCCA), perihilar (pCCA), and distal (dCCA). In contrast, transcriptomic analysis has led to the proposal of additional stratification approaches. Sia et al. proposed classifying intrahepatic cholangiocarcinoma (iCCA) into 2 subtypes: “inflammatory” and “proliferative,” which was later refined into metabolic subtype and proliferative subtype.5 This study, for the first time, defined 3 molecular subtypes that are independent of anatomical location through multiomics integration and machine learning. In CCA, molecularly matched therapy has been shown to improve patient survival compared with conventional chemotherapy, highlighting the urgency of incorporating subtype classification into clinical trial design.6 Previous multiomics studies on CCA were often restricted to a single anatomical site and primarily focused on genomics or transcriptomics, while neglecting proteomics and phosphoproteomics. Innovatively, this study integrates 4 types of data: whole-exome sequencing (genomics), transcriptomics, proteomics, and phosphoproteomics. Among these findings, the delineation of the metabolic subtype and the discovery of high TNK1 kinase activity are entirely dependent on proteomic and phosphoproteomic analyses. Notably, TNK1 showed neither mutations nor overexpression at the RNA level, and its therapeutic relevance could only be uncovered through functional proteomics, underscoring why genomic or transcriptomic studies alone would have completely missed this vulnerability. These results highlight that non-genetically driven but targetable mechanisms can be revealed only when proteomics and phosphoproteomics are incorporated, demonstrating that the “reductionist” approach of singleomics cannot fully decipher complex cancers like CCA. In traditional research, machine learning was mostly utilized as a data processing tool; however, this study has elevated it to a pivotal bridge linking basic research with clinical applications. On one hand, by employing machine learning algorithms such as HAllA clustering, it achieves effective dimensionality reduction of high-dimensional multiomics data and accurate subtype classification. On the other hand, through classifier training, it successfully “maps” external patient cohorts and patient-derived xenograft (PDX) models to the newly defined molecular subtypes, providing a standardized tool for the functional verification of subtype-specific therapeutic hypotheses. The discovery of the TNK1 target has opened up a new avenue for subtype-specific targeted therapy in CCA. Studies have demonstrated that TNK1 kinase exhibits high activity in the metabolic subtype of CCA, and the TNK1 inhibitor TP-5801 can significantly inhibit tumor growth in PDX models of this specific subtype, while showing no efficacy in other subtypes. This brings new hope to patients with the metabolic subtype, who were previously untreatable. Notably, TNK1 shows neither mutations nor overexpression at the RNA level; it was only through functional proteomics that TNK1 was identified as a therapeutic target. This represents a breakthrough of far-reaching significance. It indicates that TNK1 may emerge as a druggable vulnerability, but exclusively for tumor subtypes with specific biological characteristics. Subtype-specific differences in chemotherapy sensitivity have driven the transformation of CCA treatment strategies toward “personalization.” Research findings reveal that the immunomodulatory subtype is most sensitive to the first-line chemotherapy regimen (gemcitabine plus cisplatin), whereas the metabolic subtype exhibits extremely low sensitivity. This conclusion directly challenges the current clinical practice of “one-size-fits-all chemotherapy.” In future clinical practice, priority should be given to molecular subtype testing for CCA patients: the gemcitabine plus cisplatin regimen should be continued for patients with the immunomodulatory subtype, while alternative treatment approaches should be explored for patients with the metabolic and gene regulatory subtypes, so as to avoid the toxic side effects caused by ineffective chemotherapy.7 The multiomics–machine learning strategy provides a replicable paradigm for research on other refractory cancers. Performing proteogenomic analysis on cholangiocarcinoma enables the definition of clinically relevant subtypes and the identification of actionable pathways for therapeutic intervention.8 The technical framework of this study is not only applicable to CCA but also translatable to other refractory cancers characterized by high heterogeneity and strong drug resistance, such as liver cancer and pancreatic cancer. By applying the “multiomics integration and machine learning” strategy to hepatocellular carcinoma (HCC), pancreatic cancer, and other malignancies, hidden molecular subtypes may be uncovered.9 Meanwhile, multiomics features associated with lymph node metastasis can accurately identify patient populations with a high likelihood of occult systemic lesions. Such patients have a low probability of benefiting from aggressive locoregional treatments. This key finding not only provides a basis for screening eligible patients for surgery and systemic therapy, as well as planning treatment sequences, but also further facilitates the optimization of surgical and locoregional treatment strategies: by integrating multiomics features related to lymph node metastasis or occult systemic spread, it is possible to both screen out patients who are not suitable for surgery or transplantation. While these findings are encouraging, caution remains imperative. Despite its comprehensiveness, the study is still in the preclinical stage, and whether TNK1 inhibitors can demonstrate durable efficacy in clinical trials requires further investigation. Moreover, the biological function of TNK1 is complex; in different contexts, it has been reported to exhibit both pro-tumor and anti-tumor effects. Therefore, the use of TNK1 inhibitors in humans may carry unknown risks and side effects. In addition, the rarity of cholangiocarcinoma means that recruiting a sufficient number of patients stratified by molecular subtypes for clinical trials will be an enormous challenge. Practicality is also a realistic concern: the in-depth multiomics analysis conducted in this study is currently not feasible in routine clinical practice. Mass spectrometry–based proteomics and phosphoproteomics remain largely research tools, although advances in their throughput and cost-effectiveness may soon facilitate their clinical application.10 Importantly, because the metabolic subtype and its key target TNK1 are invisible at the genomic and transcriptomic levels but emerge clearly from proteomic and phosphoproteomic data, it will be essential to uncover the upstream regulatory networks that drive this subtype and sustain TNK1 activity. Furthermore, the absence of single-cell resolution data leaves unanswered questions about tumor microenvironmental interactions, intratumoral heterogeneity, and lineage-specific signaling, which will be critical for refining subtype classification and therapeutic targeting. Instead of basing their work on the anatomical origin of cholangiocarcinoma, Mun and colleagues redefined the disease according to its biological behaviors. By integrating multiomics and machine learning, they identified subtypes capable of predicting prognosis and treatment response, discovered the novel target TNK1, and prompted a reevaluation of both clinical trial design and patient treatment approaches. This paradigm shift—moving from where the tumor arises to what drives it—marks an essential step toward precision medicine in cholangiocarcinoma.