Primary liver cancer remains a significant global public health challenge, characterized by persistently high incidence and mortality. This review synthesizes current epidemiological data to analyze trends and etiological shifts, with particular emphasis on China, which bears over 40% of the global burden. Findings highlight a transition in dominant risk factors from viral hepatitis to metabolic dysfunction-associated steatotic liver disease (MASLD), alongside persistent threats from aflatoxin exposure and lifestyle behaviors. Evidence-based prevention strategies, including universal hepatitis B virus (HBV) vaccination, antiviral therapy expansion, aflatoxin control, and early metabolic intervention, are critical to reducing disease burden. The integration of artificial intelligence into screening and management represents a promising advancement. A multi-faceted approach combining vaccination, surveillance, lifestyle modification, and technological innovation is essential for effective global liver cancer control.

TNF and TGF-β1 expression in HL-60 cells.

Abstract Background Patients with hepatocellular carcinoma (HCC) have an unmet need for new therapies that improve survival. This phase II trial investigated the efficacy and safety of ociperlimab and tislelizumab plus BAT1706 (a bevacizumab biosimilar) in patients with first-line HCC. Methods In this phase II, multicenter, randomized, multi-arm, open-label trial, patients with advanced HCC received ociperlimab and tislelizumab plus BAT1706 (Arm A) or tislelizumab plus BAT1706 (Arm B). The primary objective was to evaluate efficacy using objective response rate (ORR) assessed by the investigator per RESIST v1.1 for Arms A and B. Results 94 patients were randomized to Arm A (N = 62) and Arm B (N = 32). Confirmed ORR (95% confidence interval) was 37.1% (25.2–50.3) for Arm A and 40.6% (23.7–59.4) for Arm B. In Arms A and B, respectively, 90.3% and 80.6% of patients experienced treatment-related treatment-emergent adverse events (TEAEs), 59.7% and 32.3% experienced Grade ≥ 3 treatment-related TEAEs and 22.6% and 9.7% experienced treatment-related TEAEs leading to treatment discontinuation. Immune-mediated adverse events were reported in 50.0% of patients in Arm A and 45.2% of patients in Arm B. Infusion-related reactions occurred in a single patient in Arm A. Conclusion In patients with advanced HCC, tislelizumab plus BAT1706 demonstrated promising ORR, while adding ociperlimab was not associated with improved efficacy. The safety profile of ociperlimab and tislelizumab plus BAT1706 was tolerable and manageable, with no new safety signals identified. Trial registration ClinicalTrials.gov: NCT04948697 (September 20, 2021).

2508 Background: Advanced hepatocellular carcinoma (HCC) remains a high unmet-medical-need malignancy with limited therapeutic options. Ori-C101 is a novel, armored, autologous GPC3-directed CAR-T cell therapy. Following promising results from early-phase trials (ChiCTR190028121; NCT05652920, BEACON study), we shall herein update outcomes from the BEACON study with focusing on long-term safety, durability of response, and survival after more than 2 years of follow-up. Methods: This is an open-label, multi-center, phase Ib dose-escalation and expansion study enrolled patients (pts) with GPC3 + advanced HCC who had progressed on ≥2 prior lines of systemic therapy (including ICIs and TKIs). A single dose of Ori-C101 was administered via hepatic arterial infusion to enhance regional cell delivery. Integrated analyses assessed safety, tolerability, PK, and efficacy (per RECIST v1.1)，aiming to determine the RP2D. Results: As of Dec 24, 2025, 19 pts received Ori-C101 infusion across 4 dose levels (DLs). All pts had BCLC stage B/C disease and 31.6% (6/19) had extrahepatic metastases. Pts were previously treated with a median of 3 lines (range 2–8) therapies. Safety: Safety remained manageable; no late-onset nor cumulative toxicities were observed through the extended follow-up period. The most common ≥G3 TEAEs (≥10.0%) were transient hematologic toxicities and hepatic laboratory abnormalities. CRS occurred in 100.0% (19/19) of pts; while ≥G3 CRS observed in 42.1% (8/19). No ICANS occurred. One pt at DL4 experienced a DLT of G4 CRS complicated by secondary DIC. Efficacy: In 18 efficacy-evaluable pts, Ori-C101 demonstrated a robust dose-dependent response. Confirmed ORR was 50.0% (9/18); DCR was 77.8% (14/18). At RP2D (DL3), the confirmed ORR and DCR were 66.7% (6/9) and 88.9% (8/9), respectively. Critically, responses were not only rapid but also remarkably durable. 88.9% (8/9) of responders achieved objective response within 1.1 months; at M3, 83.3% (5/6) of responders at the RP2D remained PR. Notably, 1 pt at DL4 achieved CR with a duration exceeding 20 months. Preliminary overall survival data indicate a substantial long-term survival benefit with a median OS of 14.4 months (range 2.6–22.0). In addition, Dose-Exposure-Responses analysis showed dose-dependent CAR-T cells expansion, pharmacodynamic effects and improved tumor response. Conclusions: Ori-C101 demonstrates a manageable safety profile and compelling, durable anti-tumor activity in GPC3 + advanced HCC. The combination of high ORR and prolonged survival benefit distinguishes Ori-C101 as a potential paradigm-shifting therapy for patients who have failed standard-of-care treatments. A phase II/III study is currently underway to further confirm the efficacy and asses the safety of Ori-C101. Clinical trial information: NCT05652920 .

Immune checkpoint inhibitor (ICI) based combined neoadjuvant/conversion therapy (NAT) is increasingly utilized in the management of initially unresectable intrahepatic cholangiocarcinoma (iCCA). However, immune-related pathologic response (irPR) features and immunotherapeutic response score (ITRS) remain poorly understood in iCCA. This study enrolled 147 iCCA patients who received NAT. irPR features (tumor infiltrating lymphocyte, tertiary lymphoid structure, lymphoid aggregate, plasma cell infiltration, granuloma, neutrophil, foamy macrophage, cholesterol cleft, hemosiderin macrophage, giant cell, neovascularization and mature fibrosis) were accessed on routinely processed hematoxylin-eosin stained post-NAT surgical resection specimen tumor bed slides. Kaplan-Meier and Cox regression analyses were used to investigate irPR features and ITRS correlations with recurrence-free survival (RFS) and overall survival (OS). The irPR features were identified in the post-NAT surgical resection specimen of iCCA. Six irPR features were included in ITRS. A binary ITRS scheme was developed, wherein patients classified as ITRS-low exhibited significantly better OS (hazard ratio [HR]:3.03; 95

Abstract Background Tumor heterogeneity presents a formidable challenge in understanding the mechanisms driving tumor progression and metastasis. The heterogeneity of hepatocellular carcinoma (HCC) in cellular level is not clear. Methods Integration analysis of single-cell RNA sequencing data and spatial transcriptomics data was performed. Multiple methods were applied to investigate the subtype of HCC tumor cells. The functional characteristics, translation factors, clinical implications and microenvironment associations of different subtypes of tumor cells were analyzed. The interaction of subtype and fibroblasts were analyzed. Results We established a heterogeneity landscape of HCC malignant cells by integrated 52 single-cell RNA sequencing data and 5 spatial transcriptomics data. We identified three subtypes in tumor cells, including ARG1+ metabolism subtype (Metab-subtype), TOP2A+ proliferation phenotype (Prol-phenotype), and S100A6+ pro-metastatic subtype (EMT-subtype). Enrichment analysis found that the three subtypes harbored different features, that is metabolism, proliferating, and epithelial-mesenchymal transition. Trajectory analysis revealed that both Metab-subtype and EMT-subtype originated from the Prol-phenotype. Translation factor analysis found that EMT-subtype showed exclusive activation of SMAD3 and TGF-β signaling pathway. HCC dominated by EMT-subtype cells harbored an unfavorable prognosis and a deserted microenvironment. We uncovered a positive loop between tumor cells and fibroblasts mediated by SPP1-CD44 and CCN2/TGF-β-TGFBR1 interaction pairs. Inhibiting CCN2 disrupted the loop, mitigated the transformation to EMT-subtype, and suppressed metastasis. Conclusion By establishing a heterogeneity landscape of malignant cells, we identified a three-subtype classification in HCC. Among them, S100A6+ tumor cells play a crucial role in metastasis. Targeting the feedback loop between tumor cells and fibroblasts is a promising anti-metastatic strategy.

Precision surgical intervention in urological oncology is inherently constrained by the limited penetration depth and lack of tactile feedback in traditional endoscopy. Near infrared II (NIR-II, 1000-1700 nm) fluorescence laparoscopy overcomes these limitations by drastically suppressing photon scattering and tissue auto-fluorescence, facilitating sub-millimeter deep-tissue navigation. However, navigating the urological system requires circumventing unique physiological barriers, notably dense adipose encapsulation and fluid-filled cavities. This review comprehensively evaluates the translational integration of high-performance InGaAs hardware architectures and advanced nanomedicines designed to conquer these specialized anatomical niches. We trace the evolution of NIR-II molecular probes from leveraging the tail fluorescence of ICG to biomarker-driven active targeting (eg, PSMA, CAIX, Nectin-4) and tumor micro-environment (TME)-responsive smart nanoplatforms. By meticulously modulating supramolecular physical chemistry, such as leveraging lipid critical packing parameter (CPP) phase transitions and mitigating aggregation-caused quenching (ACQ). These platforms achieve spatially synchronized diagnostic imaging and targeted phototheranostic ablation. Preclinically and in early human trials, this synergy enables precise primary lesion delineation, trans-adipose pelvic lymph node mapping, and neurovascular bundle (NVB) preservation across prostate, renal, and bladder malignancies. It also fosters systemic anti-tumor immunity through immunogenic cell death. Despite these breakthroughs, broad clinical translation remains impeded by formidable nanotoxicological hurdles, particularly reticuloendothelial system sequestration and the accelerated blood clearance (ABC) phenomenon, alongside robotic surgery integration challenges. To rewrite standard-of-care guidelines, future paradigms must prioritize "safe-by-design" architectures, logic-gated multi-receptor targeting to eliminate diagnostic ambiguity, and large-scale, multicenter randomized controlled trials evaluating long-term oncological outcomes.

Background: Hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality, with resistance to immunotherapy posing a major clinical challenge. Natural killer (NK) cells exhibit impaired infiltration and cytotoxicity in HCC; however, the mechanisms underlying NK cell-mediated immune evasion are still poorly understood. This study investigated how NOP2/Sun RNA methyltransferase 2 (NSUN2), a 5-methylcytosine (m 5 C) RNA methyltransferase, induces metabolic reprogramming and immunosuppression to drive HCC progression. Methods: We conducted a genome-wide CRISPR screen in HCC cells cocultured with NK cells. To delineate the downstream mechanisms, we integrated profiling of the m5C epitranscriptome, transcriptome, and chromatin landscape with metabolic characterization. The impact of NSUN2 on histone lactylation and programmed cell death 1 ligand 1 (PD-L1) transcription was further investigated. Functional assays in vitro and in vivo using syngeneic murine models and pharmacological inhibition validated these findings. Clinical relevance was assessed using patient tissues, The Cancer Genome Atlas dataset, and immunotherapy cohorts. Results: Genome-wide CRISPR screening in HCC cell–NK cell coculture models identified NSUN2 as a key suppressor of NK cell-mediated cytotoxicity. Mechanistically, NSUN2-mediated RNA m 5 C modification enhanced the messenger RNA stability and expression of glycolytic enzymes, including enolase 1 ( ENO1 ), pyruvate kinase M1/2 ( PKM ), and lactate dehydrogenase A ( LDHA ), thereby increasing lactate production. Accumulated lactate promoted histone H3 lysine 18 lactylation (H3K18la), which enhanced chromatin accessibility at the CD274 (encoding PD-L1) promoter and recruited signal transducer and activator of transcription 3 (STAT3) to drive PD-L1 expression, ultimately inhibiting NK cell-mediated cytotoxicity. Clinically, high NSUN2 expression was associated with elevated PD-L1 levels, poor prognosis, and immunotherapy resistance in patients with HCC. In vivo, NSUN2 knockout increased NK cell infiltration and suppressed tumor growth, while the STAT3 inhibitor TTI-101 combined with anti-PD-L1 therapy enhanced NK cell cytotoxicity and inhibited HCC progression. Conclusions: Our data demonstrated that NSUN2 drove immune evasion in HCC by coupling m 5 C-dependent glycolytic reprogramming with H3K18la-mediated epigenetic activation of PD-L1. These findings suggest that NSUN2 could represent a critical nexus between m 5 C RNA methylation and immunosuppression, providing a therapeutic rationale for combination immunotherapy in HCC.

LBA4000 Background: TACE, a global standard of care (SoC) for unresectable eeHCC, induces a tumor immune response. STRIDE (Single Tremelimumab [T] Regular Interval Durvalumab [D]) has shown OS benefit at 6-year follow-up and is a SoC for unresectable advanced HCC. We report preplanned analyses from EMERALD-3 (NCT05301842), which combined STRIDE ± lenvatinib (L) with TACE. Methods: Eligible pts (≥18 yr) with confirmed eeHCC were randomized 1:1:1 to STRIDE (T 300 mg + D 1500 mg on Day 1 then D 1500 mg Q4W) + L (8 or 12 mg QD) + TACE; STRIDE + TACE; or TACE until reaching 175 pts/arm. Randomization continued 1:1 until STRIDE + L + TACE and TACE reached 275 pts/arm. D and L continued for ≤36 months (mo), until disease progression, unacceptable toxicity, or withdrawn consent. Pts were stratified by region, any prior palliative embolization, and baseline tumor burden by the Up-To-Seven criteria. The primary endpoint was PFS for STRIDE + L + TACE vs TACE by a stratified Cox proportional hazards model and stratified log-rank test. Key secondary endpoints were OS (STRIDE + L + TACE vs TACE), and PFS plus OS (STRIDE + TACE vs TACE). Results: As of Feb 23, 2026, 293 pts were randomized to STRIDE + L + TACE, 175 to STRIDE + TACE, and 292 to TACE. Baseline characteristics were broadly balanced across arms. STRIDE + L + TACE showed a statistically significant improvement in PFS vs TACE (HR, 0.70; 95% CI, 0.57–0.86; p=0.0007), and a positive OS trend (HR, 0.84; 95% CI, 0.65–1.09; p=0.1814). STRIDE + TACE also improved PFS (HR, 0.71; 95% CI, 0.56–0.91) and OS (HR, 0.70; 95% CI, 0.51–0.95) vs TACE. STRIDE ± L + TACE showed higher 24-mo OS rate vs TACE (Table). The incidence of treatment-related AEs of maximum grade 3/4 was 62.7% for STRIDE + L + TACE, 48.6% for STRIDE + TACE, and 18.6% for TACE. Conclusions: STRIDE + L + TACE significantly improved PFS vs TACE. At interim analysis, with ≤45% maturity, a positive trend for OS with STRIDE ± L + TACE vs TACE was observed. STRIDE + TACE also improved PFS vs TACE. AEs were aligned with known safety profiles of individual therapies. The EMERALD-3 results support STRIDE ± L + TACE as potential new treatment option in unresectable eeHCC. Clinical trial information: NCT05301842 . STRIDE + L + TACE(n=293) TACE(n=292) STRIDE + L + TACE(n=first 175) STRIDE + TACE(n=175) TACE (n=first 175) PFS (95% CI) HR 0.70 (0.57–0.86)p = 0.0007* 0.71 (0.56–0.91) † Maturity 64%* 75% † Median, mo 13.0 (12.2–16.7)* 9.8 (8.0–11.4)* 13.1 (11.0–17.7) † 12.9 (10.2–15.9) † 8.1 (6.5–10.2) † OS (95% CI) HR 0.84 (0.65–1.09) p = 0.1814 † 0.70 (0.51–0.95) † Maturity 40% † 45% † Median, mo 39.5 (34.1–NC) † 34.7 (28.8–NC) † 39.5 (32.6–NC) † NC (37.7–NC) † 32.9 (24.1–43.2) † 24-mo rate, % 66.9 (61.0–72.2) † 61.5 (55.4–67.0) † 67.8 (60.3–74.2) † 68.0 (60.4–74.5) † 57.8 (50.1–64.9) † NC, not calculable. Based on Data cutoff 1: *Sep 2, 2025; 2: † Feb 23, 2026.

ABSTRACT Emerging evidence highlights the tumor microenvironment's (TME) role in hepatocellular carcinoma (HCC), yet how tumor‐infiltrating myeloid cells drive relapse is unclear. Using full‐length single‐cell RNA sequencing (scRNA‐seq) on samples from primary and early‐relapse HCC patients, we identified a dendritic cell subset DC3, which in relapsed tumor exhibited features of mature DCs enriched in immunoregulatory molecules (mregDCs). Mechanistically, mregDCs recruit dysfunctional CD161 + CD8 + T cells, which secrete TNF‐α, thereby activating the non‐canonical NF‐κB pathway to promote the differentiation of mature DCs into mregDCs via tumor necrosis factor receptor 2 (TNFR2). Our results from in vivo mouse models demonstrated that dual blockade of TNFR2 and PD‐L1 reduced tumor burden more effectively than anti‐PD‐L1 monotherapy in mregDC‐rich HCC. We also found strong interactions between mregDCs and FCN1+ monocytes, a myeloid‐derived suppressor cell (MDSC)‐like population. Our study characterizes an mregDC‐mediated immunosuppressive network in relapse HCC, nominating TNFR2 as a therapeutic target for myeloid‐focused HCC immunotherapy.

SUMMARY Generalizable Agentic Laboratory Intelligence for Learning, Experimentation, and Optimization (GALILEO) closes the prediction-to-intervention gap in therapeutic peptide discovery. Unlike prior AI-scientist systems, GALILEO couples an Observation-Thought-Action-Summary (OTAS) reasoning loop with robotic peptide synthesis and multimodal phenotyping. Candidate peptides are retrieved from a clinically informed peptide prior (CPP) and locally edited through auditable operations. GALILEO autonomously prioritized LRRC8C and SLC25A1 branches and used wet-lab feedback to update target beliefs, sequence policies, assay choices, and mechanism hypotheses. Peptides generated under this framework blocked LRRC8C currents, perturbed osmolyte/redox homeostasis, and promoted tumor-dependent T-cell activation, while SLC25A1 peptides disrupted citrate-export metabolism, reduced extracellular acidosis, and enhanced CD8+ T-cell function. These results establish retrieval-and-editing-based physical learning as a route for auditable, experimentally grounded interventions and transferable membrane-blocker rules.

Abstract Gut microbiome alterations are increasingly associated with hepatocellular carcinoma (HCC), highlighting the gut–liver axis as a key contributor to tumor progression and prognosis. Taxon‐based HCC microbiome studies have shown limited reproducibility because they are affected by database dependency, taxonomic ambiguity, and overlooked ecological interactions. The Two Competing Guilds (TCG) model, based on stable gut microbiome interactions, provides a structurally grounded framework for robust, generalizable biomarkers. Using shotgun metagenomic data from a newly recruited cohort of 120 surgically resectable HCC cases and 76 benign liver tumor controls, we constructed co‐abundance networks to identify stably correlated genome pairs and assembled a hepatic cancer‐TCG (HCC‐TCG) model composed of 142 genomes. Functionally, one Guild had more genes for butyrate production from carbohydrate fermentation while the other Guild was enriched in genes for virulence factors and antibiotic resistance, highlighting its potential proinflammatory roles. Classifiers trained on the abundance profiles of HCC‐TCG genomes successfully distinguished HCC from benign liver tumors (area under the receiver operating characteristic, AUROC = 0.70) and from colorectal liver metastases (CRLM) (AUROC = 0.78). In an external validation cohort, the model further discriminated against HCC from intrahepatic cholangiocarcinoma (iCCA) (AUROC = 0.72), and from healthy controls (AUROC = 0.79–0.85), demonstrating its broad applicability for tumor stratification across clinical contexts. Moreover, HCC‐TCG profiles predicted post‐resection recurrence risk and response to adjuvant therapies (AUROC up to 0.83). Importantly, external validation in two independent cohorts of advanced HCC patients treated with PD‐1/PD‐L1 inhibitors demonstrated consistent predictive performance (AUROC = 0.64–0.73), confirming the model's generalizability in nonsurgical and immunotherapy contexts. This genome‐specific, ecologically structured, and database‐independent framework identifies a conserved Guild‐based microbiome signature for HCC. Our findings demonstrate that a fixed genome‐resolved ecological structure retains transferable discriminatory signal across clinical contexts. The HCC‐TCG framework provides a genome‐specific, interaction‐based foundation for future development of non‐invasive microbiome stratification strategies requiring prospective validation.

Metabolic reprogramming within the tumor microenvironment (TME) impairs antitumor immunity and compromises the efficacy of immunotherapy. Through multiomics-based metabolic subtyping in intrahepatic cholangiocarcinoma, we identified a subgroup with the worst prognosis that demonstrates significant enrichment in both cyclooxygenase/arachidonic acid (COX/AA) metabolism and KRAS mutations. Mechanistically, KRAS mutation-mediated NF-κB pathway activation upregulates CXCL5 expression, thereby recruiting CXCR2+ polymorphonuclear myeloid-derived suppressor cells (PMN-MDSCs) into the TME. Concurrently, KRAS mutation drives prostaglandin E2 (PGE2) production in tumor cells, and PGE2, in turn, enhances AA uptake and COX-2 expression in PMN-MDSCs, establishing an amplifying loop between tumor cells and PMN-MDSCs that exacerbates PGE2 production. PGE2 accumulation potently suppresses the antitumor activity of CD8+ T cells via the prostaglandin E receptor 4 (EP4). Therapeutic targeting of the COX-2-PGE2-EP4 axis, combined with anti-PD-1 immunotherapy, demonstrates profound synergistic efficacy in both KRAS-mutant murine models and patient-derived tumor fragments harboring KRAS mutations. SIGNIFICANCE:This work identifies a PGE2-COX-2 positive feedback loop between tumor cells and PMN-MDSCs that may serve as a therapeutic vulnerability for KRAS-mutant cholangiocarcinoma.

FGFR2 fusion is a common alteration in malignancies, including intrahepatic cholangiocarcinoma (iCCA). While FGFR2 fusion has oncogenic properties, iCCA patients harboring this alteration demonstrate favorable prognosis, remaining a paradox. Here, we delineated the transcriptomic landscape of FGFR2 fusion-positive iCCA. We observed transcriptional features related to PI3K-AKT signaling in tumor cells and reduced neutrophil infiltration, particularly of the PD-L1+ neutrophil subtype. Mechanistically, FGFR2 fusion was associated with reduced H3K27ac enrichment at the CXCL3 promoter and decreased CXCL3 expression in tumor cells, which may contribute to impaired neutrophil recruitment to tumor tissues. In preclinical models, pharmacological FGFR inhibition increased CXCL3 levels and neutrophil infiltration. Combining neutrophil blockade with clinically available FGFR inhibitors enhanced antitumor activity. In conclusion, this study provides mechanistic insights into the paradox between the oncogenic properties of FGFR2 fusion and favorable clinical outcomes in FGFR2 fusion-positive iCCA, and suggests a potential strategy to optimize FGFR2-targeted therapies.

Clinical research advances slowly because its core tasks, from evidence synthesis to mechanistic validation, remain fragmented. We present MedGenesis, a clinical artificial intelligence (AI) scientist built on a world-model reasoning loop that jointly updates a Latent Hypothesis Space and a Latent Action Space under expected information gain (EIG), uncertainty reduction (UR), and a safety prior P(safe), and integrates longitudinal electronic health records (EHRs) via the Virtual Clinical Trajectory and Observation Representation (ViCTOR) for cohort retrieval, trajectory stratification, and time-to-event analysis. On two benchmarks - ClinicalResBench (1,697 expert-curated questions) and ClinicalRepBench (40 paper-reproduction tasks) - MedGenesis outperformed frontier language models and biomedical AI systems while reducing hallucination. Across 1 million patient observations spanning five clinical evidence formats, it generated traceable outputs across meta-analysis, randomized controlled trials, real-world trajectories, case-control studies, and case reports, with one wet-lab-coupled run nominating a 3-hydroxybutyrate - neutrophil axis modulating antitumor immunity. These results compress hypothesis-to-evidence cycles from years to hours, creating a continuous clinical discovery process. ### Competing Interest Statement The authors have declared no competing interest. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The study was approved by the Institutional Review Board of Zhongshan Hospital, Fudan University (No. B2023-350). Human analyses were conducted in accordance with the Declaration of Helsinki and institutional data-governance rules. De-identified donor blood samples were handled under the same approved governance framework. Animal experiments were performed under institutional animal-care procedures, with humane endpoints applied according to approved laboratory practice. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Data availability ClinicalResBench and ClinicalRepBench were constructed for this study. Benchmark metadata, task descriptions, scoring rubrics, aggregate model outputs, and processed evaluation tables will be made available upon publication through a public repository, subject to removal of copyrighted source text, restricted clinical content, and information that could enable re-identification of patients or reconstruction of institutional records. Full paper-reproduction task materials in ClinicalRepBench will be released in redacted form when redistribution of original article content is not permitted. Publicly available datasets and resources used in this study can be accessed from their original repositories, including PubMed/PMC, TCGA, GEO, GTEx, CPTAC, SEER, NHANES, ClinicalTrials.gov, ChEMBL, PubChem, FAERS, and other cited public databases. Controlled-access resources, including UK Biobank and MIMIC-IV, remain governed by their original data-use agreements and must be accessed through their respective application procedures. The UK Biobank analysis was conducted under an approved UK Biobank access application. Institutional EHR, pathology, imaging, follow-up, and outcome-registry data from Zhongshan Hospital, Fudan University, are not publicly available because they contain patient-level clinical information and are subject to institutional governance, ethics approval, and data-use restrictions. All institutional analyses were performed within the approved institutional computing environment. No patient-level institutional data left the institutional infrastructure. De-identified aggregate results, summary statistics, model-evaluation tables, and figure-level processed data will be made available where permitted by ethics approval and data-governance policy. Code availability Analysis code used to generate the figures, benchmark summaries, statistical analyses, meta-analysis outputs, survival analyses, enrichment analyses, and processed tables will be released in a public repository upon publication. The repository will include documented scripts, environment files, example inputs, and instructions sufficient to reproduce the reported aggregate analyses from released or publicly obtainable data. Code for ClinicalResBench and ClinicalRepBench evaluation, including scoring templates, rubric files, post-processing scripts, and aggregate benchmarking notebooks, will be released after removal of restricted source text and copyrighted or non-redistributable materials. Code for analyses involving controlled-access cohorts or institutional EHR data will be provided as executable templates or pseudonymized workflow scripts, but not with patient-level inputs. Components that depend on protected institutional databases, governed EHR wrappers, or internal access-control systems will not be publicly released. National Natural Science Foundation of China, 82130077, 82341008, 92459301, 82441049, 82394450 Noncommunicable Chronic Diseases-National Science and Technology Major Project, 2023ZD0502000 Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China, JYB2025XDXM508 Lingang Laboratory, LGL-8888-07 Shanghai Academy of Natural Sciences (SANS) Exploration Scholars Project New Cornerstone Science Foundation

e16291 Background: Clinical management of hepatocellular carcinoma (HCC) is becoming increasingly complex, encompassing curative resection for early-stage disease and multimodal treatment strategies across disease stages. However, preoperative risk stratification and timely evaluation of treatment response remain challenging due to the lack of reliable blood-based tools. Tumor-agnostic circulating tumor DNA (ctDNA) methylation profiling may help address these unmet clinical needs by enabling real-time, noninvasive assessment of tumor biology and therapeutic efficacy. Methods: We prospectively enrolled 174 HCC patients undergoing curative hepatectomy and collected plasma samples within 24 hours prior to surgery (Ts). In parallel, an independent neoadjuvant therapy cohort of 27 HCC patients treated with immune checkpoint inhibitor-based systemic therapy was included, in which plasma samples were obtained at baseline before treatment initiation (Tb) and prior to surgery (Ts). All plasma samples were assayed using the GutSeer. A tumor-agnostic machine learning model was developed based on 342 HCC patients and 1,415 healthy subjects to generate a methylation-based cancer score (MCS) for prognostic stratification and pathological response assessment. Results: The hepatectomy cohort was followed for a median of 26.3 months with 47 recurrences. Preoperative Ts MCS was significantly associated with advanced tumor stage, larger tumor size, and vascular invasion, indicating its ability to reflect tumor burden and underlying biological aggressiveness. Based on Ts MCS, patients were stratified into low-risk (n = 45) and high-risk (n = 129) groups, yielding a hazard ratio (HR) of 9.36 (95%CI: 2.27–38.59; p < 0.001) and a negative predictive value of 95.6%. Multivariate analysis confirmed Ts MCS as an independent predictor of recurrence (HR = 5.09, 95%CI: 1.18-21.92; p < 0.05). In the neoadjuvant therapy cohort, MCS decreased significantly from Tb to Ts (p < 0.001). The change of MCS (Tb-Ts) was strongly negatively correlated with the proportion of residual viable tumor (R = -0.70, p < 0.001), indicating MCS changes closely mirror pathological response. Consistently, MCS change accurately discriminated major pathological response (MPR, ≤50% viable tumor cells) from non-MPR (AUC = 0.859), outperforming changes of alpha-fetoprotein (AUC = 0.731) and des-gamma-carboxy prothrombin (AUC = 0.686). Conclusions: This study demonstrates that tumor-agnostic ctDNA methylation profiling enables effective preoperative risk stratification in resectable HCC. Furthermore, dynamic changes in MCS provide a precise, noninvasive biomarker for assessing pathological response to systemic therapy. These findings highlight the broad clinical utility of methylation-based liquid biopsy approaches across surgical and systemic treatment settings in HCC. Clinical trial information: NCT06178809 .

e16223 Background: Hepatocellular carcinoma (HCC) remains challenging to manage once unresectable. Although atezolizumab plus bevacizumab is standard first-line therapy, its potential to enable conversion to curative resection is not well defined. We evaluated the efficacy and safety of atezolizumab plus bevacizumab combined with on-demand transarterial chemoembolization (TACE) in a real-world cohort, with a focus on surgical conversion. Methods: We conducted a multicenter real-world study of 121 patients with unresectable HCC treated with atezolizumab plus bevacizumab ± TACE. The primary endpoint was the objective response rate (ORR), and secondary endpoints included progression-free survival (PFS), overall survival (OS), disease control rate, surgical conversion rates and safety. Results: Patients treated in the first-line setting achieved better outcomes than those receiving second-line therapy (ORR 47.3%, 53/112, vs 22.2%, 2/9), with longer median PFS (19.0 vs 8.8 months) and OS (35.3 vs 13.0 months). In the first-line cohort, patients receiving atezolizumab plus bevacizumab with TACE showed higher ORR (50.5%, 48/95, vs 29.4%, 5/17), longer median PFS (20.9 vs 8.0 months) and OS (35.3 vs 25.6 months) compared with those without TACE, with outcomes numerically exceeding historical GO30140 and IMbrave150 benchmarks. Successful R0 resection was achieved in 33.7% (32/95) of patients in the with-TACE group versus 5.9% (1/17) in the without-TACE group. Among resected patients, pCR occurred in 34.4% (11/32) and MPR in 68.8% (22/32; MPR, < 50% viable tumor cells). Among patients achieving MPR, RECIST responses included CR (n = 3), PR (n = 10), and SD (n = 9), indicating incomplete concordance between pathological and radiological assessments. A single tumor and higher CD8 + T-cell infiltration in baseline biopsy specimens were associated with a more favourable ORR. We further identified ECOG performance status, lower tumor burden, and pathological features of baseline biopsy specimens (higher CD4 + and CD8 + T-cell infiltration) as key factors associated with successful conversion to surgery. Treatment-related adverse events (TRAEs) occurred in 91.6% and 88.2% of patients in the with- and without-TACE groups, with grade 3/4 events in 54.7% and 52.9%, respectively; no grade 5 events or postoperative complications were observed. Conclusions: In this multicenter real-world cohort, on-demand TACE combined with atezolizumab plus bevacizumab enabled deep tumor responses and curative-intent resection in one-third of initially unresectable HCC patients. Baseline immune infiltration identified candidates most likely to benefit from conversion. These findings support a conversion-oriented strategy integrating locoregional and systemic therapy to bridge unresectable HCC toward surgical cure.

The clinical application of artificial intelligence in breast cancer diagnosis is constrained by poor interpretability and insufficient modeling of tumor heterogeneity during molecular subtyping. In clinical practice, radiological triage and downstream pathological assessment follow a cascaded workflow, in which triage errors directly compromise subsequent decisions. To reflect this process, we retrospectively evaluate a two-stage, workflow-oriented multimodal framework for benign–malignant triage and molecular subtype prediction. In the first stage, a Knowledge-Infused Chain-of-Thought mechanism integrates standardized BI-RADS diagnostic rules into a large multimodal model via Low-Rank Adaptation. The model generates structured, rule-consistent reasoning steps and diagnostic outputs and achieved 98.51% accuracy in benign–malignant classification under the internal evaluation protocol. In the second stage, molecular subtyping is performed using a dedicated Mixture of Experts architecture, wherein a gating network assigns malignant cases to specialized expert subnetworks for subtype-specific feature learning. To evaluate the system under a clinically realistic setting, we define an auxiliary Clinical Workflow Composite Score that accounts for error propagation and explicitly penalizes missed malignancies. Experiments on a multi-center dataset of 666 cases yielded the highest observed workflow-level performance for Ours-7B among the evaluated general-purpose and medical-domain Multimodal Large Language Models. Independent external evaluation on a 196-case hospital cohort yielded the highest workflow-level and specialist-stage point estimates among evaluated methods. These retrospective findings support further evaluation of explicit clinical rules and expert routing within workflow-aligned, auditable decision-support research. The anonymous source code repository is available at https://anonymous.4open.science/r/KI-COT-C835/.