Society: AASLD

Background: The LI-RADS (Liver Reporting and Data Systems) was created to standardize reporting of CT/MR imaging for liver masses in cirrhosis patients at risk of hepatocellular carcinoma (HCC). LI-RADS (LR) 3 and 4 observations are indeterminate and have a variable natural history, with 35-65% becoming HCC and others remaining stable or regressing over time. We hypothesize that by using artificial intelligence and radiomics, we can extract imaging features to better predict presence of HCC in patients with LR3/4 lesions.
Method: Leveraging the analytic morphomics image processing platform, we used a random sampling strategy to extract radiomics features on automatically registered images from different contrast phases. Samples were randomly extracted from the tumor and peritumoral areas. Random virtual nodules were computationally generated in nontumoral areas and in control patients with cirrhosis. Chart review was performed for extraction of clinical data and outcomes. Multi-modal prediction models were developed using gradient boosting tree framework with automated hyperparameter selection.
Results: We used a retrospective cohort of 391 patients with cirrhosis and a pre-treatment multi-phase CT scan (219 from University of Michigan and 172 from University of Texas Southwestern) to train and internally validate models. Mean age was 58.6 years, 71.7% were male, and cohort was diverse regarding race/ethnicity (57.3% white, 17.5% Hispanic, 17.0% Black, 6.2% other) and liver disease etiology (56.3% viral, 16.7% alcohol, 15.2% NAFLD, 11.8% other). 252 (64.5%) patients were diagnosed with HCC within 1 year of the index scan, 1 segmented nodule and 1 virtual nodule was used per patient, comprising 21 LR3, 48 LR4, and 183 LR5 nodules. For the 139 controls, 2 virtual nodule per patient was utilized with the exception of 1 LR3 nodule which did not become HCC. Model performance was validated using k-fold cross validation. We found that models that included radiomics outperformed AFP alone in all cases (AUROC 0.89 – 0.91 vs. 0.80) (Table). The best multi-modal models which included radiomics on index imaging with AFP at time of index imaging performed similarly in LR3 (AUROC 0.91) and LR4 (AUROC 0.90) observations compared to those with LR 5 (AUROC 0.92) in predicting HCC (Figure).
Conclusion: Artificial intelligence and radiomics are promising tools to accurately predict risk of HCC in indeterminate liver nodules.
Funding Sources: NIH U01CA230669; VA I01HX002548; U01 CA230694

<b>TABLE: </b>Performance of Different Models to Predict HCC Using Radiomics and Clinical Data

TABLE: Performance of Different Models to Predict HCC Using Radiomics and Clinical Data

<b>Figure</b>. AUROC of the Full Model in Prediction of HCC in LR3,LR4,vs LR 5 observations

Figure. AUROC of the Full Model in Prediction of HCC in LR3,LR4,vs LR 5 observations

Background & Aims: Squalene epoxidase (SQLE), a key rate-limiting enzyme in cholesterol biosynthesis, promotes non-alcoholic steatohepatitis-induced hepatocellular carcinoma (NASH-HCC). However, whether SQLE plays an immunomodulatory role in NASH-HCC pathogenesis remains unclear. In this study, we aim to determine the function of SQLE in the tumor immune microenvironment, and its therapeutic implication for immune checkpoint blockade therapy in NASH-HCC.

Methods: We established hepatocyte-specific Sqle knock-in and knockout mice, and they were given diethylnitrosamine (DEN) injection plus choline-deficient high-fat (CDHF) diet to induce NASH-HCC. Orthotopic NASH-HCC model was established in immunocompetent mice fed with CDHF diet and intrahepatically inoculated with mouse RIL-175 cells with or without Sqle knockout. Alterations of the immune landscape of NASH-HCC mediated by SQLE were characterized by single-cell RNA sequencing (scRNA-seq) and flow cytometry. SQLE inhibitor terbinafine was given in combination with anti-PD1 therapy in mouse models of NASH-HCC.

Results: Hepatocyte-specific Sqle knock-in mice receiving DEN with CDHF diet exhibited markedly increased tumor number and size compared to wildtype littermates. We observed a significant reduction in functional IFN-γ⁺CD8⁺ T and Granzyme B⁺CD8⁺ T cells subsets, and an increased infiltration of Arginase-1⁺ (Arg1⁺) myeloid-derived suppressor cells (MDSCs) in Sqle knock-in tumors, suggesting that the overexpression of SQLE was associated with an immunosuppressive microenvironment. Reciprocally, hepatocyte-specific Sqle knockout mice or Sqle knockout orthotopic allograft models demonstrated inhibited tumor growth, which was associated with increased cytotoxic CD8⁺ T cells and reduced MDSCs infiltration in tumors, implying reactivation of antitumor immunity. Consistent with in vivo data, culture supernatant of human NASH-HCC cells with SQLE overexpression significantly attenuated CD8⁺ T cell effector functions but enhanced MDSC immunosuppressive activity as compared to control supernatant; whereas the depletion of SQLE exerted opposite effects. Mechanistically, SQLE-driven cholesterol biosynthesis and accumulation mediates the immunosuppressive effect in NASH-HCC. Cholesterol depletion in vitro abolished the effect of SQLE overexpression on CD8⁺ T cell suppression and MDSC activation. Notably, targeting SQLE with terbinafine was found to rescue the efficacy of anti-PD-1 treatment in NASH-HCC in mice, implying SQLE as a promising immunotherapeutic target.

Conclusions: SQLE induces an impaired antitumor response in NASH-HCC via attenuating tumor-infiltrating CD8⁺ T cell effector function and augmenting immunosuppressive MDSCs, an effect dependent on cholesterol biosynthesis. SQLE is a promising target in potentiating anti-PD-1 immunotherapy for NASH-HCC.

Background & Aims: Biliary tract cancer (BTC) is one of the most lethal human malignancies. Most of BTC patients are diagnosed at an unresectable stage, and only systemic therapies are available. Patients show different responses for chemotherapy and there is no effective way to predict chemotherapeutic response. We aimed to establish BTC patient-derived organoids (PDOs) and evaluate their accuracy for personalized drug screening of BTC patients.
Methods: Organoids were derived from 72 patients with BTC. Drug screening was carried out in BTC organoids using seven chemotherapeutic drugs (gemcitabine, cisplatin, 5-fluoruracil, oxaliplatin, irinotecan, mitomycin C and paclitaxel). Treatment responses in PDOs were validated in patient-derived organoids-based xenografts (PDOXs) in mice and in primary BTC patients. Immunohistochemistry staining was performed to evaluate the histological features. RNA sequencing was performed on BTC organoids and original BTC tumor tissues.
Results: We effectively developed 61 BTC PDOs from 82 tumors (74.4%), which showed similar histological characteristics to the corresponding primary BTC tissues. BTC tumor tissues with enhanced stemness and proliferation related gene expression by RNA sequencing were easier to form organoids. As expected, BTC PDOs showed different responses to the chemotherapies of gemcitabine, cisplatin, 5-fluoruracil, oxaliplatin, irinotecan, mitomycin C and paclitaxel. These drug screening results in PDOs were further validated in PDOX mice, and in particular confirmed in 92.3% (12/13) of BTC patients with actual clinical response, indicating the feasibility of drug screening using PDOs. Moreover, we identified gene expression signatures of BTC PDOs with different drug responses and established gene-based panels to predict chemotherapy response in BTC patients. The prediction panels of drug response to 5-FU, gemcitabine and cisplatin yielded area under the curve (AUC) of 90.7% (95% confidence interval (CI): 80.2-99.9%), 86.4% (95% CI: 73.3-99.6%) and 82.1% (95% CI: 68.1-96.1%), respectively, which may be good clinical decision-making support tools for BTC.
Conclusions: We established patient-derived BTC organoids with characteristics of originated tumors, which could be used for personalized drug screening of BTC patients with high accuracy. A chemotherapy response prediction gene-panel might help to select effective drugs for individual BTC patient.

Yifan Wu

Appearances