Developing angiogenesis-related prognostic biomarkers and therapeutic strategies in bladder cancer using deep learning and machine learning

Abstract

Bladder cancer (BLCA) is a prevalent urological malignancy that exhibits a high degree of tumor heterogeneity and morbidity. Tumor angiogenesis, a vital hallmark of cancer, greatly influences the tumor microenvironment (TME). The emergence of anti-angiogenic drugs has provided a new turning point in cancer treatment. An integrated machine learning system was constructed to build the angiogenesis-related gene signatures (ARGS). ARGS was used to assess TME status in BLCA. Pharmacophore construction was employed to construct pharmacophore features of highly cytotoxic drug payload combinations for antibody-drug conjugates (ADCs). In addition, we developed a natural compound using artificial intelligence-driven drug design technology. This compound exhibits anti-angiogenic effects in BLCA and serves as a highly cytotoxic drug payload for ADCs. Multi-dimensional machine learning was used to screen biomarkers for evaluating the post-treatment effects of drug therapy in BLCA. The ARGS consists of 12 angiogenesis-related genes associated with prognostic risk in BLCA. The ARGS divides BLCA patients into high-risk and low-risk groups. Significant TME remodeling was identified in the high-risk BLCA cohort and demonstrated a strong association with tumor angiogenesis. Expression levels of key immune checkpoint markers significantly differed between BLCA risk groups. Saikosaponin D (SSD) shows promising potential as a novel ADC drug for anti-angiogenic treatment in BLCA. Multi-dimensional machine learning results indicate that MYH11 is the most likely biomarker for evaluating the post-treatment effects of SSD therapy. SSD may potentially treat tumors by regulating angiogenesis in BLCA. The detection of MYH11 can be used to assess the therapeutic effectiveness of SSD in BLCA.

Keywords: Antibody-drug conjugates; Artificial intelligence-driven drug design; Bladder cancer; Machine learning; Pharmacophore; Protein homology modeling; Saikosaponin D.

Fig. 1

Workflow diagram of this study.

Fig. 2

Functional characterization, screening, and modeling of DEGs in BLCA for ARGS model construction. A, B Heatmaps illustrating differential expression levels of DEGs between BLCA samples and normal bladder tissues; Volcano plots depicting expression fold-changes of DEGs. C, D Functional enrichment analysis of DEGs using GO and KEGG. E, F Unicox and LASSO regression were used to screen prognostic risk genes for BLCA, which were then utilized to construct the ARGS model. G Stratification of BLCA patients into high-risk and low-risk subgroups based on prognostic genes, with comparative analysis of gene expression patterns between subgroups.

Fig. 3

ARGS model was used to evaluate the prognosis and correlation analysis of BLCA patients. (A) ARGS model was utilized to distinguish high-risk and low-risk subgroups of BLCA patients. (B) The AUC values for clinical characteristics risk demonstrated good concordance between the two models. (C) PCA and t-SNE analysis confirming distinct sample stratification and ARGS model efficacy. (D) KM survival curves demonstrating significant OS disparity between ARGS-defined risk subgroups. (E) Correlation heatmap of ARGS risk scores with clinicopathological characteristics.

Fig. 4

Development of the ARGS model using ensemble learning and identification of angiogenesis-associated independent prognostic factors. A Ensemble learning validation of ARGS prognostic power. B, C Nomogram performance evaluation. D Forest plot of multivariate Cox regression showing ARGS as independent risk factor.

Fig. 5

Functional enrichment analysis of BLCA samples derived from the ARGS model for high- and low-risk groups. A, B GSVA, GO, and KEGG analysis of prognostic genes. C Functional enrichment analysis of transcription factors associated with angiogenesis risk genes.

Fig. 6

The contribution of ARGS model to TME remodeling in BLCA and prediction of immunotherapy efficacy. (A) Immune functional activation status in angiogenesis-related high-immune and low-immune subtypes of BLCA. (B) The t-SNE sample distribution plots for different immune level subgroups. (C) The degree of TME reshaping, as well as the differences in immune scores and stromal scores, among different immune subgroups. (D) Evaluation of immune cell infiltration levels and their association with angiogenesis risk using seven immune cell algorithms. (E) The heatmap depicting the association between angiogenesis risk and the activation/blockade status of immune checkpoints. (F) The contribution of angiogenesis risk to the activation/inhibition of immune-related functions in the TME. (G) Boxplots illustrating the association between angiogenesis risk and C1-C6 immune subtypes in BLCA. (H) Calculating IPS to predict the efficacy of immunotherapy in different subgroups of BLCA patients. ^*P < 0.05, ^**P < 0.01, ^***P < 0.001.

Fig. 7

ARGS model evaluation of chemotherapy drug sensitivity in high-risk BLCA patients and screening of potential anti angiogenic therapy targets based on machine learning. (A) Drug library screening for sensitive chemotherapy drugs in high-risk group patients. (B) The small molecule compound STOCK1N-35,696 shows the most potential for anti-angiogenic treatment in BLCA. (C) The BATCH algorithm was used to remove batch effects from samples in different datasets. D–G Integration of ML for the screening of biomarkers to evaluate the therapeutic efficacy of anti-angiogenic treatment in BLCA.

Fig. 8

Confirmation of biomarkers for evaluating the therapeutic efficacy of anti-angiogenic treatment and their functional analysis. (A) PCA analysis to examine the dispersion of samples after batch removal using the BATCH algorithm. (B) Venn diagram illustrating the genes with the highest potential to serve as biomarkers for evaluating treatment efficacy. C, D ROC models and AUC for the training and validation cohorts, as well as the model genes. E Violin plots depicting the expression level of model genes in the training and validation cohorts. F Functional enrichment analysis of biomarkers in BLCA using GSVA. *P < 0.05, **P < 0.01, ***P < 0.001.

Fig. 9

Screening and validation of novel anti-angiogenesis ADCs using AIDD technology. (A) Known pharmacophore receptor feature models for ADCs. (B) Homology modeling of the protein FN1’s tertiary structure and analysis of the existence of amino acid residue conformations as well as residue Chi1-Chi2 analysis. C, D Analysis of the rationality of the FN1 protein tertiary structure and Ramachandran plot of the protein model. E ADME property analysis of potential natural compounds targeting FN1 and anti-angiogenesis effects.