Mass Cytometry Based Single-cell Immune Signature And Artificial Intelligence For The Prediction Of Tumor Recurrence After Liver Transplantation For Hepatocellular Carcinoma

Background:Hepatocellular carcinoma(HCC)accounts for 80%of all liver cancer.Liver transplantation(LT)is widely considered as the most effective treatment for localized HCC.However,20%-57.8%of all patients who received LT for HCC(HCC-LT)will develop tumor recurrence,and the median overall survival of these patients is only 10.6-12.2 months.Early and precise prediction of post-LT tumor recurrence will greatly benefit the implementation of personalized post-LT anti-tumor therapy,thus improve the prognosis of HCC-LT recipients with high risks of tumor recurrence.To date,the prediction of post-LT tumor recurrence was mainly based on oncological traits including tumor number,tumor diameter,tumor differentiation,vascular invasion,serum alpha fetal protein(AFP)level.These predicting factors were integrated into HCC-LT selection criteria like Milan criteria and Hangzhou criteria to select HCC patients with low risks of post-LT tumor recurrence.However,the development of personalized post-LT management requires a more accurate prediction of post-LT tumor recurrence,which cannot be achieved by preexisting predicting methods.Therefore,more precise approaches for the prediction of post-LT tumor recurrence are in great need.Aims:This research aims to develop early and precise approaches for the prediction of post-LT tumor recurrence.Our study consists of two sections.In the first section,we intended to profile the dynamic peripheral immune atlas of HCC-LT recipients using mass cytometry(CyTOF)and thereby searching for immunological biomarkers that can predict post-LT tumor recurrence.The second section was dedicated to systematically integrating the clinical information of HCC-LT recipients into a predicting model for post-LT tumor recurrence by machine learning methods.Methods:Section 1:We collected peripheral blood samples from HCC-LT recipients with tumor recurrence(TR HCC-LT recipients),HCC-LT recipients with no tumor recurrence(NR HCC-LT recipients),and non-HCC-LT recipients(NHCC-LT recipients),at Shulan Hospital between January 2018 and May 2019.Blood samples were collected before surgery,and at 3 days,1 week,2 weeks,3 weeks after surgery.Peripheral Blood Mononuclear Cells(PBMCs)were extracted from each blood sample and then analyzed by a custom-designed 42-marker CyTOF panel.CyTOF data was visualized by the t-distributed Stochastic Neighbor Embedding(t-SNE)algorithm.Unsupervised clustering algorithm X-shift was applied to automatically separate the sample pool into mutually exclusive immune subsets.Pearson correlation analysis was applied to reveal the intercommunication between different X-shift-defined immune subsets.Tumor-specific immune subsets were identified by comparing the immune features of TR HCC-LT recipients with NR HCC-LT recipients.And another CyTOF panel consists of 30 biomarkers and cytokines was designed for the functional assessment of tumor-specific immune subsets.Section 2:We retrospectively included HCC patients who received LT between January 1,2015 and March 30,2020 at Shulan Hospital and the First Affiliated Hospital of Zhejiang University School of Medicine.Patients were 70%/30%randomly separated into training cohort or test cohort.The predicting ability of 40 clinical parameters was examined in a random forest pre-build process.And the most effective predictors were included in the construction of the final model.The prediction performance of the generated model was quantified by the area under the Receiver Operating Characteristic curves(AUROC).Kaplan-Meier survival analysis was applied to compare the relapse-free survival of the tumor recurrence high-risk group and tumor recurrence low-risk group defined by the random forest model.Net reclassification index(NRI)was computed to compare the predicting ability of the machine learning model and traditional HCC-LT recipients selection criteria.Results:Section 1:Systematic single-cell CyTOF analysis revealed a dramatic change in the peripheral immune composition of LT recipients after surgery.The proportion of T cells and NK cells was significantly decreased 3 days after LT.Meanwhile,the proportion of B cells and myeloid cells significantly increased.X-shift analysis revealed an increased proportion of activated immune cells(effector memory CD8+T cells,HLA-DR+CD8+T cells)accompanied by a decreased proportion of suppressive immune cells(myeloid-derived suppressor cell)and naive immune cells(naive CD8+T cells).The immune signature of HCC-LT recipients is significantly different from NHCC-LT recipients.The proportion of CD28+γδ T cells and CD57+HLA-DR+CD8+T cells were significantly higher in TR HCC-LT recipients than NR HCC-LT recipients and NHCC-LT recipients 3 weeks after LT.CD57+HLA-DR+CD8+T cells exhibited a highly cytotoxic and proliferative phenotype with high expression levels of Perforin,Granzyme B,and Ki-67.CD28+γδ T cells secreting less IFN-γ than their CD28counterparts,indicating an immature phenotype.Subsequent correlation analysis discovered that multiple links between these immune subsets were specifically strengthened in tumor recurrence recipients.Section 2:435 HCC-LT recipients were included in this section.136 HCC-LT recipients(31.3%of the total)developed tumor recurrence after LT.Tumor recurrent HCC-LT recipients had significantly higher serum AFP level(185.65 vs 22.00,P<0.001)and larger maximum tumor length(7.78 vs 4.60,P<0.0001)than HCC-LT recipients without tumor recurrence.Maximum tumor length,Maximum tumor width,and serum AFP level were the top 3 predictors for post-LT tumor recurrence in the random forest pre-build process.The random forest model performed best with the top 6 parameters(maximum tumor length,AFP,vascular invasion,albumin,total bilirubin,and international normalized ratio)included in the model.The generated model outperformed individual clinical parameters including AFP and tumor size in predicting post-LT tumor recurrence,with AUROC of 1 in the training cohort and 0.756 in the test cohort(threshold=0.434,sensitivity:67.5%,specificity:82.2%).The tumor recurrence high-risk group defined by the machine learning model exhibited significantly lower 1-year(87.9%vs 45.1%)and 2-year(82.4%vs 24.3%)relapse-free survival rate than the low-risk group.The machine learning model performed better than Milan criteria(NRI increment:0.256)and UCSF criteria(NRI increment:0.147)in distinguishing HCC-LT recipients with high risks of tumor recurrence.Conclusions:Section 1:In-depth CyTOF analysis comprehensively depicted dynamic immune responses induced by LT.The immune signature of HCC-LT recipients was distinct from NHCC-LT recipients.Tumor-specific immune subsets identified 3 weeks after LT were potential predictors for post-LT tumor recurrence.Section 2:AFP,tumor diameters are effective predictors for post-LT tumor recurrence.The machine learning model exhibited better performance than individual clinical parameters and traditional HCC-LT selection criteria in predicting post-LT tumor recurrence.