AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer

Abstract

Due to low success rates and long cycles of traditional drug development, the clinical tendency is to apply omics techniques to reveal patient-level disease characteristics and individualized responses to treatment. However, the heterogeneous form of data and uneven distribution of targets make drug discovery and precision medicine a non-trivial task. This study takes pyroptosis therapy for triple-negative breast cancer (TNBC) as a paradigm and uses data mining of a large TNBC cohort and drug databases to establish a biofactor-regulated neural network for rapidly screening and optimizing compound pyroptosis drug pairs. Subsequently, biomimetic nanococrystals are prepared using the preferred combination of mitoxantrone and gambogic acid for rational drug delivery. The unique mechanism of obtained nanococrystals regulating pyroptosis genes through ribosomal stress and triggering pyroptosis cascade immune effects are revealed in TNBC models. In this work, a target omics-based intelligent compound drug discovery framework explores an innovative drug development paradigm, which repurposes existing drugs and enables precise treatment of refractory diseases.

Fig. 1. Schematic diagram.

a The schematic illustration of the target omics-based intelligent compound drug discovery framework. b The workflow for discovering and optimizing anti-TNBC compound pyroptosis drugs based on big data and artificial intelligence. c Construction of MG@PM and the proposed mechanism of MG@PM for promoting systemic anti-tumor immune response by evoking tumor cell pyroptosis-induced cascade effect.

Fig. 2. Identification of pyroptosis genes and drugs across TNBC cohorts.

a The workflow for screening of signature genes and corresponding drugs. b Gene Ontology (GO) enrichment analysis of 45 typical pyroptosis-associated genes. c Heat map of differentially expressed pyroptosis regulator genes in TNBC and normal breast tissues from the FUSCCTNBC cohort. d Diagrams of the correlations between the expression levels of pyroptosis regulators. The scatter plot represented the correlation between CASP1 and NLRP3, GSDMD and CASP1, TRADD, and NFKB2 (NLRP3 vs. CASP1: P < 0.0001; CASP1 vs. GSDMD: P < 0.0001; NFKB2 vs. TRADD: P < 0.0001). Data are analyzed with Spearman correlation. e The distribution diagram (forest plots) of hazard ratios across typical pyroptosis genes. f The Kaplan-Meier overall survival curve of two clusters distinguished by pyroptosis strata (risk=high vs. risk = low: P < 0.0001). Data are analyzed with Log-rank test. g Correlation between IC50 values of 35 predicted pyroptosis inducers and the expression levels of typical pyroptosis genes. * P < 0.05, ** P < 0.01, and *** P < 0.001. Data are represented as mean ± SD and analyzed with two-way ANOVA followed by multiple comparisons test. Source data are provided as a Source Data file.

Fig. 3. Screening compound drugs and predicting effects through machine learning.

a The workflow for screening effective compound drugs affecting pyroptosis signature genes by artificial intelligence. b Schematic diagram of the architecture of BFReg-NN. As an example, drug targets for MIT and GA were shown on the left, and the nine identified pyroptosis signature genes were on the right. c Accuracy analysis of the BFReg-NN model comparing logistic regression, neural network, and random forest in the task of TNBC subtype classification (n = 100 technical replicates). Central line indicates median, box indicates interquartile range, whiskers show 1.5x the IQR. Data are analyzed with one-way ANOVA followed by multiple comparisons test. d Histogram of the ranking of BFReg-NN predicting the correlation between each drug combination and TNBC RFS. The calculation was repeated ten times. e Phase contrast microscopy images of MDA-MB-231 cells treated with twelve sets of compound drugs randomly selected from the top 10% of the ranking. Scale bar = 5 μm. The experiments were repeated three times independently. f Visualization images of important targets of MIT and GA to regulate pyroptosis signature genes using Integrated Gradient prediction. The warm-toned colors represented the targets of GA, while the cool-toned colors indicated the targets of MIT. The width of the lines represented the importance score. g Diagram of drug and protein combination. i) GA and TXN, ii) GA and MCL1, iii) MIT and BTK, iv) MIT and TOP2A. Source data are provided as a Source Data file.

Fig. 4. Characterization of bionic nanococrystals.

a Cytotoxicity assay of 4T1 cells treated with a combination of MIT and GA at variable mass ratios and b the corresponding combination index (CI) was plotted with fraction affected (Fa) (n = 3 independent experiments). c Phase contrast microscopy images of 4T1 cells treated with a combination of MIT and GA at variable mass ratios. Scale bar = 5 μm. d Pyroptosis index of 4T1 cells after different treatments (n = 3 independent experiments). The experiments were repeated three times independently. Data are analyzed with one-way ANOVA followed by multiple comparisons test. e TEM images of MG, MG@PM. Scale bar = 100 nm. f The particle size and photographs of MG@PM. g Zeta potential of MG@PM (n = 3 independent experiments). Data are represented as mean ± SD and analyzed with one-way ANOVA followed by multiple comparisons test (PM vs. MG: P < 0.001; PM vs. MG@PM: P = 0.1337). h The particle size variation of MG or MG@PM in PBS during seven days (n = 3 independent experiments). i UV absorption spectra of MIT, GA, and MG. j Western blot analysis of platelet signature proteins including CD47, CD41, and P-selectin on MG@PM. k Quantitative analysis of the band intensity in image J (n = 3 independent experiments). Data are represented as mean ± SD and analyzed with unpaired student’s t tests (PM vs. MG@PM in CD47: P = 0.033; PM vs. MG@PM in CD41: P = 0.7546; PM vs. MG@PM in P-selectin: P = 0.1184). l Pharmacokinetic properties of ICG-labeled MG (IMG) and MG@PM (IMG@PM) in vivo. Inset was the blood retention of IMG and IMG@PM 24 h post-injection (n = 3 mice). Data are represented as mean ± SD and analyzed with unpaired student’s t tests. m Fluorescence imaging of 4T1 tumor-bearing mice at preset time points after administration (left) and ex vivo imaging of major organs at 24 h after administration (right). n The semi-quantification results of the fluorescence intensity of major organs (n = 3 mice). Data were represented as mean ± SD. Data are analyzed with unpaired student’s t tests (IMG vs. IMG@PM in tumor: P < 0.001). Source data are provided as a Source Data file.

Fig. 5. In vitro pyroptosis effects and mechanism of nanococrystals.

a Uptake of ICG-labeled MG (IMG) and MG@PM (IMG@PM) in 4T1 cells at 1 h and 4 h, respectively. Scale bar = 50 μm. b Viability of 4T1 cells after different treatments (n = 5 independent experiments). MIT vs. MG@PM: P < 0.001; GA vs. MG@PM: P < 0.001. Data are analyzed with two-way ANOVA followed by multiple comparisons test. c Phase contrast microscopy observation of 4T1 cell morphology under different treatments. White arrows pointed to pyroptosis cells. Scale bar = 15 μm. d LDH release from 4T1 cells after different treatments (n = 5 independent experiments). MIT vs. MG@PM: P < 0.001; MIT vs. MG: P < 0.001; GA vs. MG@PM: P < 0.001; GA vs. MG: P < 0.001. Data are analyzed with one-way ANOVA followed by multiple comparisons test. e Intracellular ATP content in 4T1 cells after different treatments (n = 5 independent experiments). MIT vs. MG@PM: P < 0.001; MIT vs. MG: P < 0.001; GA vs. MG@PM: P < 0.001; GA vs. MG: P < 0.001. Data are analyzed with one-way ANOVA followed by multiple comparisons test. f Flow cytometry assay of the ROS production in 4T1 cells after different treatments. g Intracellular GSH level in 4T1 cells after different treatments (n = 5 independent experiments). MIT vs. MG@PM: P < 0.001; MIT vs. MG: P < 0.001; GA vs. MG@PM: P = 0.031; GA vs. MG: P = 0.145. Data are analyzed with one-way ANOVA followed by multiple comparisons test. h Heat map of the differential expression of pyroptosis-related proteins in the MG@PM group compared with the control group. i Quantitative protein analysis of predicted target genes for GA and MIT (n = 3 independent experiments. Control vs. MG@PM in Txn: P < 0.0001; Control vs. MG@PM in Mcl1: P = 0.0059; Control vs. MG@PM in Top2a: P = 0.0157; Control vs. MG@PM in Erbb2: P = 0.0425). Data are analyzed with unpaired student’s t tests. j Bubble plot of the KEGG pathway analysis for differentially expressed proteins enriched in the MG@PM group compared to the control group. The colors of the nodes reflected the P-values of the designated pathways, and the sizes of the nodes indicated the number of differentially expressed proteins enriched in the pathways. Data are analyzed with two-way ANOVA followed by multiple comparisons test. k Analysis of the protein functional interaction network of the MG@PM group using the Search Tool for the Retrieval of Interacting Genes / Proteins (STRING) algorithm. Data were represented as mean ± SD. Source data are provided as a Source Data file.

Fig. 6. Antitumor and immune activation efficacy of nanococrystals in vivo.

a Schematic diagram of the therapeutic regimen of murine 4T1 tumor models. b Changes in the body weight of tumor-bearing mice during the therapeutic period (n = 5 mice). Data are analyzed with two-way ANOVA followed by multiple comparisons test. c Tumor volume change curves after different treatments (n = 5 mice). d Images of the excised orthotopic tumors in various groups on day 15. The red dashed rectangle indicated mouse death. e Pathological H&E staining and TUNEL staining of tumor slices on day 15. Scale bar = 100 μm. f The expression levels of inflammatory factors INF-γ (MG@PM: P < 0.001) and g TNF-α (MG@PM: P < 0.001) in the serum of mice treated with different formulations on days 6, 7, and 8 (n = 3 mice). Data are analyzed with two-way ANOVA followed by multiple comparisons test. h Flow cytometry analysis of the proportion of mature DC cells in tumor-draining lymph nodes (TDLN) (n = 4 mice). Data are analyzed with one-way ANOVA followed by multiple comparisons test. i Flow cytometry analysis of the ratio of CD4⁺ T cells in tumors (n = 4 mice). MIT vs. MG: P < 0.001; GA vs. MG: P = 0.0038. Data are analyzed with one-way ANOVA followed by multiple comparisons test. j Flow cytometry analysis of the ratio of CD8⁺ T cells in tumors (n = 4 mice). MIT vs. MG: P < 0.001; GA vs. MG: P = 0.0015. Data are analyzed with one-way ANOVA followed by multiple comparisons test. k Flow cytometry analysis of the proportion of Treg cells in tumors (n = 4 mice). MIT vs. MG: P < 0.001; GA vs. MG: P = 0.0016. Data were represented as mean ± SD and analyzed with one-way ANOVA followed by multiple comparisons test. Source data are provided as a Source Data file.

Fig. 7. In vivo anti-metastasis efficacy of nanococrystals and the cascade pyroptosis effect.

a Schematic illustration of dosing regimens in 4T1-luc metastatic tumor models. b In vivo bioluminescence images of mice in different groups. c Representative lung photographs and H&E staining images of lung slices in different groups. Scale bar = 500 μm. d Number of lung metastasis nodules in different groups (n = 5 mice). Anti-PD-1/Abraxane vs. MG@PM: P < 0.001. Data are analyzed with one-way ANOVA followed by multiple comparisons test. e Survival curves of mice in different groups in 60 days (n = 6 mice). PBS vs. MG@PM: P < 0.001; Anti-PD-1/Abraxane vs. MG@PM: P = 0.180. f Immunohistochemistry analysis of GSMDE at the lung site after the indicated treatments. The dashed line referred to the metastatic tumor area. Scale bar = 100 μm. g Fluorescence images of GSDME (red) and GZMB (green) distribution in metastatic lung tissues after the indicated treatments. The nuclei were stained with DAPI (blue). Scale bar = 50 μm. Data were represented as mean ± SD. Source data are provided as a Source Data file.