Plasma cell subtypes analyzed using artificial intelligence algorithm for predicting biochemical recurrence, immune escape potential, and immunotherapy response of prostate cancer

Abstract

Background: Plasma cells as an important component of immune microenvironment plays a crucial role in immune escape and are closely related to immune therapy response. However, its role for prostate cancer is rarely understood. In this study, we intend to investigate the value of a new plasma cell molecular subtype for predicting the biochemical recurrence, immune escape and immunotherapy response in prostate cancer.

Methods: Gene expression and clinicopathological data were collected from 481 prostate cancer patients in the Cancer Genome Atlas. Then, the immune characteristics of the patients were analyzed based on plasma cell infiltration fractions. The unsupervised clustering based machine learning algorithm was used to identify the molecular subtypes of the plasma cell. And the characteristic genes of plasma cell subtypes were screened out by three types of machine learning models to establish an artificial neural network for predicting plasma cell subtypes. Finally, the prediction artificial neural network of plasma cell infiltration subtypes was validated in an independent cohort of 449 prostate cancer patients from the Gene Expression Omnibus.

Results: The plasma cell fraction in prostate cancer was significantly decreased in tumors with high T stage, high Gleason score and lymph node metastasis. In addition, low plasma cell fraction patients had a higher risk of biochemical recurrence. Based on the differential genes of plasma cells, plasma cell infiltration status of PCa patients were divided into two independent molecular subtypes(subtype 1 and subtype 2). Subtype 1 tends to be immunosuppressive plasma cells infiltrating to the PCa region, with a higher likelihood of biochemical recurrence, more active immune microenvironment, and stronger immune escape potential, leading to a poor response to immunotherapy. Subsequently, 10 characteristic genes of plasma cell subtype were screened out by three machine learning algorithms. Finally, an artificial neural network was constructed by those 10 genes to predict the plasma cell subtype of new patients. This artificial neural network was validated in an independent validation set, and the similar results were gained.

Conclusions: Plasma cell infiltration subtypes could provide a potent prognostic predictor for prostate cancer and be an option for potential responders to prostate cancer immunotherapy.

Keywords: artificial intelligence; immune escape; immunotherapy; plasma cell; prostate cancer.

Figure 1

Workflow for analysis of plasma celle subtypes using artificial intelligence algorithms, (BRFS, Biochemical recurrence-free survival. DEGs, Differentially expressed genes; GSVA, Gene set varation analysis).

Figure 2

The clinical features and tumor microenvironment patterns associated with the plasma cell fraction in PCa patients. (A) The overview of the association between plasma cell fraction and clinical features of patiens. (B) Violin plots of plasma cell fraction in individual samples of PCa patiens, stratified by clinical features. (C) The immune subtypes of PCa patiens were categorized on the basis of the overall immune activity. (D) Correlation analysis between plasma cell fraction and the tumor purity, stromal score, immune score and ESTIMATE score evaluated by ESTIMATE algorithm. *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 3

Survival analysis and differential expression between high and low plasma cell fraction groups. (A) The higher plasma cell fraction group had better biochemical recurrence-free survival. (B) Volcano plots of DEGS between high and low plasma cell fraction groups. (C) The DEGs heatmap showed the expression levels between two groups. Red and blue represented high and low expressions, respectively. (D) The functional enrichment analyses of DEGs, KEEG enrichment analysis is on the left and GO enrichment analysis is on the right.

Figure 4

Two molecules subtypes of plasma cell with different clinical characteristics, biochemical recurrence probability, and mutant landscape. (A) The CDF curves of the consensus score from k = 2 to 10. (B) The Consensus clustering matrix when the best K = 2. (C) The heatmap of the expression patterns of 25 DEGs, with red indicating high expressions and blue indicating low expressions. (D) Comparison of clinical features between two plasma cell subtypes. (E) Kaplan–Meir survival analysis exhibited significantly worse BCR-free survival probability in plasma cell subtype 1. (F) Plasma cell fraction was higher in plasma cell subtype 1. (G) Waterfall plots showed the top 20 mutated between plasma cell subtype. *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 5

Analysis of immune escape potential and efficacy of immunotherapy between plasma cell subtypes. (A) Comparisons of the fraction of 20 immune cells in two plasma cell subtypes. (B) Heatmap illustrated the enrichment scores of 41 differentially enriched molecular pathways evaluated by GSVA analysis between plasma cell subtype 1 and 2. Yellow and blue represented high and low enrichment scores, respectively. (C) TIDE, Exclusion, Dysfunction and MSI score different plasma cell subtypes. (D) Comparisons of the proportions of nonresponders and responder to immunotherapy among different classification methods. (Left: plasma cell subtype, Right: immune subtype). *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 6

Construction and validation of ANN for plasma cell subtype prediction. (A) The average AUC curves of feature screening were obtained by the three machine learning algorithms (Randoms forest, SVM and XGboost) respectively under the 10-fold cross-validation, the dotted line represents the best number of features. (B) Venn diagrams shows 10 characteristic genes of plasma cell subtypes shared by the three machine learning models. (C) The ANN simulation diagram includes three hidden layers except the input&output layer, and the neurons in each layer are (64,8,64), respectively. Regions larger than 12 neurons are omitted due to plotting limitations. (D) The ROC curves of the ANN in distinguishing two subtypes in the train set and test set. (E) Confusion matrix and evaluation parameters in test set. (F) The GEO independent validatiob queue was predicted using ANN, and the expression of 10 characteristics genes was visualized through heat maps. (G) Kaplan-Meier survival analysis showed differences in biochemical recurrence-free probability among plasma cell subtypes in the GEO independently validated cohort. (H) There were differences in TIDE scores between the tow plasma cell subtypes in the GEO cohort.