Ethical statementAll the animal experiments were performed in accordance with the guidelines evaluated and approved by Institutional Animal Care and Use Committee (IACUC), Fudan University School of Pharmacy (Shanghai, China).MaterialsAll chemical agents were of analytical grade. Mitoxantrone (MIT) was purchased from Meilunbio, Co., Ltd (Dalian, China). Gambogic acid (GA), Indocyanine green (ICG), and Hoechst 33258 were obtained from Aladdin Reagent Co. Ltd. (Shanghai, China). Cell counting kit-8 (CCK-8), Calcein-AM/PI, Annexin-V-FITC/PI, JC-1, and DAPI were from KeyGen Biotech (Nanjing, China). DCFH-DA was acquired from Sigma-Aldrich (St. Louis, MO, USA). Abraxane (nanoparticle albumin-bound paclitaxel) was purchased from Celgene Co. Ltd. (New Jersey, USA). Anti-CD47 antibody (ab108415), anti-CD41 antibody (ab134131), and anti-CD62P (ab255822) were purchased from Abcam (Cambridge, UK). Anti-PD-1 antibody was purchased from Bio X Cell Biotechnology (New Hampshire, USA). Dulbecco’s modified eagle medium (DMEM), certified fetal bovine serum (FBS), phosphate Buffered Saline (PBS), penicillin-streptomycin stock solutions, and trypsin-EDTA (0.25%) were obtained from Invitrogen Co., (Carlsbad, CA, USA). All the other chemical solvents and agents were acquired from Sinopharm Chemical Reagent Co., Ltd (Shanghai, China).Cell cultureThe tumor cell line MDA-MB-231, 4T1, and 4T1-Luc were obtained from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China) and cultured in DMEM with the addition of 10% FBS (v/v), 100 mg/mL of streptomycin, and 100 U/mL of penicillin. Cells were incubated at 37°C in a humidified environment with 5% CO2. Cells were tested monthly and found to be negative for mycoplasma contamination.Experimental animalsFemale Balb/C mice of 7-week-old (20 ± 1 g) and female Sprague-Dawley (SD) rats of 10-week-old (200 ± 10 g) were both obtained from SLAC Animal Ltd. (Shanghai, China) and raised under standard housing conditions in the Department of Experimental Animals, Fudan University (Shanghai, China). Female mice were chosen because the majority of breast cancers is seen in female patients. According to the guidelines of the ethics committee, the maximal tumor size permitted was 1500 mm3. Mice were euthanized when the tumor burden exceeded this threshold. Due to the blood volume requirements, female SD rats were used for the LC-MS/MS analysis.Exploration of the drug candidatesThe gene expression of 360 TNBC patients and 88 normal breast tissues was pre-processed by Transcripts Per Million (TPM) normalization. Given 45 pyroptosis genes, we used the R package “survival” to analyze the survival of TNBC patients according to the gene expression of these genes. Then we selected the key nine pyroptosis genes as target omics associated with TNBC with the highest Akaike Information Criterion (AIC), indicating the best trade-off between the survival’s goodness of fit and the model’s simplicity65. According to the nine key pyroptosis genes, we identified 3804 drug candidates by chemical–gene interactions in the Comparative Toxicogenomics Database (http://ctdbase.org/downloads/). To filter these drug candidates, we obtained the RNA composite expression of the pyroptosis regulators and the corresponding compound activity data (DTP NCI-60) from the CellMiner database (https://discover.nci.nih.gov/cellminer/home.do). We conducted a correlation analysis between the average z-score of compound activity and the RNA composite expression of the pyroptosis regulators. Then we identified 133 drug candidates significantly related to at least one pyroptosis gene (P < 0.05). To further narrow down the list of drug candidates, we analyzed the targets of these compounds using the SwissTargetPrediction database (http://www.swisstargetprediction.ch/) and PPI networks obtained from the STRING database (https://string-db.org/)66. We qualified 35 drug candidates with at least one target associated with pyroptosis genes whose distance was smaller than or equal to 2 of the PPI value.Introduction of BFReg-NNBFReg-NN could define the neural network architecture by existing biological knowledge. For example, we divided biological factors into distinct levels, L = {Gene, Protein, Pathway,…}. The modulatory relationships of internal factors were presented as a matrix set A = {AGene, AProtein, APathway,…,AL}. AGene could be a genetic relationship determined by the Gene Regulatory Network (GRN), while AProtein was defined by PPI. Since APathway was a hypergraph, every edge could link over two nodes, called hyperedges. The factors in a hyperedge propagated the information to directly affect others. Values in Al were binary, representing the presence of relations. We also identified the binary mapping matrixes M = {M1, M2, …, ML−1} from level l to its upper-level l+1 as the interaction between levels. We programmed M1 as a mapping from genes to proteins and M2 to be a straightforward mapping across levels of proteins and pathways. Both A and M determined the structure of the neural network, which included neurons and connections between neurons.BFReg-NN used gene expression data \(x\) as input. We encoded each gene individually using the embedded layer, where \({H}_{i}^{{{\mathrm{0,0}}}}={emb}({x}_{i})\). At the intra-level \(l\), we aimed to help each biological factor interact with others by \({A}_{l}\). Thus, we used graph neural networks and message-passing mechanisms to update the embedding \({H}^{l}\), where \({\widehat{H}}_{i}^{l}={{{\rm{update}}}}({\sum}_{j\in {A}_{l}(i)}{{{\rm{meassage}}}}({H}_{i}^{l},{H}_{j}^{l}),{H}_{i}^{l})\). The message function was to produce messages from factor j to factor i, where \({A}_{l}(i)\) decided which factors were neighbors for factor i. The updating function was to renew the embedding of factor i with the acquired message and the prior hidden embedding. The embedding was learned level by level. As \({M}_{l}\) was inter-layer relationship between l and l + 1, we utilized the shielded deep neural network at the next layer to update the original expression \({H}_{i}^{l+{{\mathrm{1,0}}}}\), where \({H}^{l+1}={{{\rm{activation}}}}(({M}_{l}\odot {W}^{l}){\widehat{H}}^{l,}+{b}^{l})\). The element-wise multiplication \({M}_{l}\odot {W}^{l}\) ensured that absent relations were not applied to update. \({W}^{l}\) and \({b}^{l}\) were parameters that can be learned in deep neural networks.We further enhanced BFReg-NN by adding extra edges in A. The available biological knowledge reflected the hidden relationships between factors detected by biotechnology. Nevertheless, owing to technical limitations, certain knowledge remained challenging to be detected. Hence, we set up the interaction in two ways. One was generalized conditioning sustained by available knowledge A. The other was a partial interaction concealed at the non-existent edge of A, predicting fresh biological knowledge. We used \({A}_{l}\) to restrain the learnable matrix \({A^{\prime} }_{l}\) rather than the binary matrix \({A}_{l}\) employed in the basic model to find fresh knowledge. For non-existent edges, we reweighed it by a small value \(\alpha=0.005\) due to it being less convincing. As a result, the edge intensities depending on both types of knowledge were revised as follows:$${A^{\prime}}_{l}=\left\{\begin{array}{c}{\beta }_{i,j}^{l},\, {{{\rm{universal}}}}\; {{{\rm{regulation}}}}\; {{{\rm{that}}}}\; {{{\rm{already}}}}\; {{{\rm{exists}}}}\; {{{\rm{in}}}}\, {A}_{l}\\ \alpha {\beta }_{i,j}^{l},\hfill{{{\rm{local}}}}\; {{{\rm{interaction}}}}\; {{{\rm{that}}}}\; {{{\rm{is}}}}\; {{{\rm{ignored}}}}\; {{{\rm{in}}}}\, {A}_{l}\end{array},\right.$$
(1)
The parameter \({\beta }_{i,j}^{l}\) was learned by a Multilayer Perceptron based on the embeddings \({H}_{i}^{l}\) and \({H}_{j}^{l}\). Thus, we learned the factor embeddings while using these embeddings and transitions to deduce the strength of implicit interactions among factors. Finally, we used some downstream tasks to train BFReg-NN and obtain the weights in BFReg-NN. For example, we could train it to classify the cell types based on gene expression if we know the labels of cells. By analyzing the weights, we obtained the important genes and how these genes led to the cell type.We compared our proposed BFReg-NN model with various types of methods on three different tasks. These methods include: (1) traditional methods (MLP, LSTM, Random Forest, XGBoost, Transformer), (2) two GNN methods (GCN and GAT), and (3) two classical medical models (DCell and P-NET)67,68. For both the GNN and medical models, we provided gene expression data and graph topology as inputs. The graph topology included gene regulatory networks, protein-protein interaction networks, and Gene Ontology (GO) term hierarchy, etc. Therefore, we ensured a fair evaluation of its performance. The detailed experimental procedure can be found at this link (https://arxiv.org/pdf/2304.04982.pdf).Utilization of BFReg-NN to predict the drug effectsRecall the four layers are (1) relations among drug targets, (2) links between drug targets and pyroptosis genes, (3) relations among pyroptosis genes (4) the survival layer. In the first and third layers, we use GNN-based equations to obtain the overall situation among drug targets or pyroptosis genes. In the second layer, we use a DNN-based equation to transform the overall situation reflected from drug targets into the individual situation of each pyroptosis gene. The final layer is a CoxPH model to transform the overall situation reflected from the pyroptosis gene into patient survival time. Here we limited the neural network architecture of BFReg-NN by drug targets and pyroptosis genes. We defined \({A}_{1}\) as the relationship between drug targets, \({A}_{2}\) as the relationship between pyroptosis genes, and \({M}_{1}\) as the association between drug targets and pyroptosis genes. All the relationships were extracted by PPI networks (https://string-db.org/). We set the combined score of an edge as larger than 0.6 in \({A}_{1}\), \({A}_{2}\) and \({M}_{1}\). To enrich the association in \({M}_{1}\), factors were also connected if they are 2-hop neighbors. The downstream task was a survival task to predict recurrence-free survival (RFS) of TNBC patients. For all possible drug combinations, we trained the corresponding BFReg-NNs by different architectures decided by different input drug targets. If the drug targets can accurately predict the survival of TNBC patients, it indicates that these two drugs have highly relevant targets for TNBC patient survival. The c-index was used as an indicator to evaluate the fitness of the model69. It quantifies the level of concordance between predicted and observed survival times, with 1 indicating perfect concordance, and 0.5 suggesting no better than random chance. The c-index of the final selected drug pairs is larger than 0.9, which indicates the high association of drug pairs and the survival of TNBC patients. We also implemented Integrated Gradient to analyze BFReg-NN after it was well-trained. We obtained the important score of each drug target through different pyroptosis genes to influence TNBC RFS.Preparation and characterization of nanococrystalsPreparation of platelet membrane-derived vesiclesPlatelets were separated from the entire blood of female Balb/C mice of 7-week-old (20 ± 1 g) by centrifugation to separate erythrocytes and leukocytes and suspended in PBS with protease inhibitor tablets35. The purified platelet-rich plasma was subsequently placed in ice-cold PBS containing EDTA and prostaglandin E1 (PGE1, Sigma Aldrich, USA) to prevent platelet activation and then centrifuged at 800 × g for 20 min at room temperature to collect platelet precipitates. Platelet membranes were obtained via a repeat freeze-thaw cycle. Briefly, platelet suspensions were frozen in liquid nitrogen, thawed at 37°C, and centrifuged at 4000 × g for 5 min. After washing three times, it was resuspended in water and set aside.Preparation of MG and MG@PMMG nanococrystals were prepared by a one-step self-assembly method. The dimethyl sulfoxide (DMSO) solution containing 5 mg/mL of MIT and the ethanol solution containing 3 mg/mL of GA were mixed in equal volumes, and 100 µL of the mixture was added dropwise to 2 mL of ddH2O. Then stirred at room temperature for 5 h, and MG gradually appeared. After centrifugation (15,000 × g), the precipitate was collected and repeatedly rinsed in water three times to remove DMSO and unassembled drug. After the last centrifugation, MG was resuspended with 1 mL of sterile PBS and stored at 4°C for subsequent use. To prepare MG@PM, 100 μL of MG suspension (3.5 mg/mL) was mixed with 150 µL of platelet membrane suspension (1 mg/mL of membrane proteins) and then successively squeezed through 1 µm, 400 nm, and 200 nm polycarbonate porous membranes (Avanti Polar Lipids Inc, GE Healthcare, USA). ICG-labeled MG or MG@PM were obtained by the same procedure as MG and MG@PM, except that 1 mg/mL of ICG was added to the ethanol solution during the self-assembly process of MG.Nanococrystal characterizationThe morphology of MG and MG@PM was characterized by using transmission electron microscopy (TEM) (TEM-1400 Plus electron microscope, Leica, Germany). The diameter, polydispersity index, and zeta potential of nanococrystals were measured by a dynamic light scattering detector (Zetasizer, Malvern, UK). To examine the stability of MG and MG@PM in PBS, the Z-average diameter was measured for seven consecutive days, while the steady state of MG@PM in different media was observed and photographed. The loading capability (LC) and encapsulation efficiency (EE) of MIT and GA in MG@PM were measured by the UV-vis spectrophotometer (PerkinElmer Lambda 750). The characteristic absorption peaks of MIT and GA were 610 nm and 370 nm, respectively.EE (%) = (Amount of drugs in nanococrystals)/(Total amount of drugs input) × 100%LC (%) = (Amount of drugs in nanococrystals)/(Nanococrystals weight) × 100%We performed drug release of MG@PM with a dialysis method. Briefly, 2 mL of MG@PM suspension (containing 1 mg of MIT and 1.5 mg of GA) was put in dialysis bags (ThermoFisher, MWCO = 3 kDa) and incubated within 50 mL of release medium (0.01 M PBS, pH = 5.5 or 7.4) at 37 °C at 100 rpm for 24 h. The release profiles of MIT and GA from MG@PM was tested.Membrane protein characterizationThe membrane proteins of MG@PM nanococrystals were characterized by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). MG@PM nanococrystals were lysed with the RIPA lysate, and protein concentrations were quantified, with platelet membranes and MG serving as controls. The presence of specific protein markers was identified by Western blot analysis. In brief, the resulting gels were shifted onto polyvinylidene difluoride (PVDF) membranes and hatched with antibodies (anti-CD47, anti-CD41, and anti-P-selectin) overnight at 4 °C. The membranes were then incubated with horseradish peroxidase-labeled goat anti-rabbit lgG H&L (HRP) antibody (G-21234, InvitrogenTM, USA) at room temperature for 1 h. GAPDH antibody (MA5-15738, InvitrogenTM, USA) was used as a control. The resulting bands were tested with an ECL developer (Beyotime, China) and quantified using ImageJ software.Cellular uptakeTo determine the cellular uptake of MG and MG@PM, 4T1 cells were seeded into confocal dishes (2×104 cells per dish) and cultured for 24 hours. Then ICG-labeled MG and MG@PM (50 ng/mL of ICG) were added to cells and incubated for 1 h and 4 h. To acquire fluorescence signals, cells were fixed with 4% paraformaldehyde before nuclei staining with Hoechst 33258 for 8 min. Photos were subsequently taken with a confocal laser scanning microscope (CLSM) (Leica, Germany).Cytotoxicity assayA total of 5 × 103 4T1 cells per well were seeded overnight in 96-well plates, which were subsequently subjected to various treatments for 24 h. Afterward, the CCK-8 reagent was applied to the cell and hatched for 1.5 h. Then the absorbance at 490 nm of each well was determined using the microplate reader (Multiskan MK3, Thermos, USA).Besides, the live/dead cell staining experiment was carried out using CLSM. Briefly, 4T1 cells were seeded into confocal dishes (2 × 104 cells/dish) and incubated for 24 h. Then cells were cocultured with various formulations for 24 h, staining with Calcein-AM and PI according to the protocol, followed by imaging with the CLSM.Synergistic effect evaluation4T1 cells were treated with compound drugs of MIT and GA at different mass ratios, and cell viability was examined as described above. The combination index (CI) of various combinations was examined using the Chou-Talalay method, and the results of the CompuSyn software were classified as synergistic (CI < 1), additive (CI = 1), and antagonistic (CI > 1)70. Fraction affected (Fa) between 0.2 and 0.8 was deemed to be valid.Flow cytometry4T1 cells with different treatments (Control, MIT, GA, MG, MG@PM) were harvested, stained with Annexin V-FITC/PI, and subjected to flow cytometry for quantitative analysis of cell pyroptosis (CytoFLEX S, Beckman, USA). For quantitative analysis of ROS production in tumor cells, 4T1 cells were treated with different formulations for 6 h and stained with DCFH-DA (10 µM) for 30 min, after which they were washed with PBS several times and collected for flow cytometry.4T1 cells with different treatments (Control, MIT, GA, MG, MG@PM) were fixed, permeabilized, and blocked. The samples were hatched with anti-GSDME-N antibody overnight at 4 °C, then stained with anti-FITC-IgG (H + L) antibody at room temperature. After that, the cells were collected and washed with PBS several times and analyzed with flow cytometry.Pyroptosis assayTo observe cell morphological changes, 4T1 cells were seeded into 6-well plates, after which they were cultured for 24 h in different preparations respectively. The cell morphology was visualized under a phase contrast microscope (Olympus, Japan). Cellular LDH and ATP levels were tested using a firefly luciferase-based ATP assay kit (Beyotime, China) and LDH cytotoxicity assay kit (Beyotime, China) according to both manufacturers’ instructions. The luminescent value and absorbance were measured using a microplate reader (Multiskan MK3, Thermos, USA).Western blotting analysisThe cell or tissue samples were lysed with the RIPA lysate, and protein concentrations were quantified. The presence of specific protein markers was identified by Western blot analysis. The resulting gels were shifted onto PVDF membranes and hatched with antibodies overnight at 4 °C. Primary antibody against GSDMB (ab215729, 1:1000), GSDMD (ab209845, 1:1000), and GSDME (ab215191, 1:1000) was obtained from Abcam. Caspase-3 (#9662, 1:1000) and GAPDH antibody (#2118, 1:1000) were purchased from Cell Signaling Technology. The resulting bands were incubated with goat anti-rabbit lgG H&L (HRP) antibody and tested with an ECL developer (Beyotime, China) and quantified using ImageJ software.siRNA-mediated knockdown4T1 cells were seeded into 6-well plates (1 × 105/well) for siRNA-mediated knockdown. After 24 h in culture, 3 μL of siCasp3 or siGsdme (RiboBio, Guangzhou, China) was transfected with Lipo3000 according to the manufacturer’s instructions. After 72 h, transfected 4T1 cells were treated with PBS, MIT, GA, MG, or MG@PM for subsequent experiments.Consumption of intracellular GSHCells after different treatments were collected and subjected to two freeze-thaw circles to release intracellular content, and the supernatant was removed after centrifugation and processed according to the kit instructions. Intracellular GSH concentrations were measured by the GSH assay kit (Keygen Biotech, China).TMT-based proteomics analysisProteomics features of 4T1 cells treated with MG@PM were analyzed by liquid chromatography-mass spectrometry (RIGOL L-3000, RIGOL TECHNOLOGIES, Beijing, China)35. The 4T1 cells without any treatment were taken as control. Protein samples were electrophoresed by SDS-PAGE to examine the concordance of each group. Protein samples (100 µg) were reduced, alkylated, and digested overnight with trypsin at 37°C. Then, each sample solution was tagged with TMT labeling reagent. After mixing and labeling, the samples were dissolved with 100 µL of mobile phase A (water containing 10 mM ammonium formate), centrifuged at 14000 × g for 20 min, and the supernatant was extracted and graded using a high-performance liquid chromatography phase. The solubilized powder was dissolved in 10 µL of liquid A, centrifuged at 14000 × g for 20 min at 4°C, and 1 µg of the supernatant was injected into the sampler for liquid mass detection using an Orbitrap Fusion Lumos mass spectrometer. The NSI source was operated in positive mode. Other ionization source parameters are listed as follows, spray voltage: 2000 V, capillary temperature: 320 °C. The scan ranges of full scan mode were set at m/z 407–1500. The raw data of mass spectrometry detection was analyzed using the Mus musculus UniProt database71. The parameters for peptide identification using Proteome Discoverer 2.4 software (Thermo Fisher Scientific) were set as follows: Carbamidomethyl(C) as static modification; The was M Oxidation (15.995 Da), TMT-6plex (K, N-terminal), Acetyl (Protein N-terminal) as dynamic modification；Precursor ion mass tolerance was ± 15 ppm; Fragment ion mass tolerance was ± 0.02 Da; Max missed cleavages was 2. Peptide- and protein-level false discovery rates (FDRs) were filtered to 1%. Statistical analysis of the identification and quantization results was done using Perseus 1.6.7.0 software. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE72 partner repository with the dataset identifier PXD053939.Protein activity assay4T1 cells were seeded in 12-well plates at a density of 1 × 105 cells per well and cultured overnight. Afterward, cells were treated with different formulations for 12 h. After that, cells were collected, lysed in an ice bath, and spun down. The concentration of Txn, Mcl1, Top2a, and Erbb2 in the supernatant was analyzed by using the ELISA assay kits (Jianglai Biology, China) according to the manufacturer’s instructions. Then optical density (OD) value at 450 nm of each well was determined using the microplate reader.NADH and NAD+ measurement4T1 cells were cultured in 12-well plates (5 × 104 cells per well). After being treated with PBS, MIT, GA, MG, or MG@PM for 12 h, the cells were washed with PBS three times. NADH and NAD+ levels were assessed using the NAD+/NADH assay kit (WST-8) (Beyotime, China) according to the manufacturer’s instructions. Then optical density (OD) value at 450 nm of each sample was determined using the microplate reader.Mitochondrial Membrane Potential Monitor4T1 cells were seeded into 12-well plates (1 × 105/well) for 24 h at 37 °C and incubated with different formulations. Finally, cells were dyed with JC-1 dyestuff for 30 min and the fluorescence signal was obtained by flow cytometry. The change of the JC-1 indicator from red to green can facilely detect the decrease of cell membrane potential.Pharmacokinetics and biodistribution studyTo confirm that the platelet membrane modification could endow the nanococrystal with prolonged blood circulation, six female Balb/C mice of 7-week-old (20 ± 1 g) were intravenously injected with IMG and IMG@PM (1 mg/kg of ICG) separately. At the indicated time points (0.017, 0.25, 0.5, 1, 2, 4, 8, 12, and 24 hours after injection), blood samples (50 μL) were taken from the fundic venous plexus and placed into heparin-preprepared polyethylene tubes. Plasma was collected by centrifugation and then subjected to fluorescence analysis of ICG on a microplate reader (Ex = 785 nm, Em = 810 nm). The major pharmacokinetic parameters were analyzed with Drug and Statistics (DAS) software.The orthotopic breast tumor model was established by inoculating 3 × 106 4T1 cells into one mammary fat pad of female Balb/C mice of 7-week-old (20 ± 1 g). To identify the tumor-targeting efficiency of nanococrystals, IMG and IMG@PM (1 mg/kg of ICG) were intravenously injected in tumor-bearing mice, respectively. Fluorescence imaging of mice was performed with the IVIS Spectrum imaging system (PerkinElmer, USA) at pre-designed time points (1, 2, 4, 8, 12, and 24 h) post-injection. After 24 h, tumors and major organs (heart, liver, spleen, lung, and kidneys) were isolated from the mice and subjected to ex vivo fluorescence imaging using the IVIS Spectrum imaging system.LC-MS/MS analysisSix SD rats were randomly divided into 2 groups (n = 3 rats) and i.v. injected with MG and MG@PM (3 mg/kg of MIT, 4.5 mg/kg of GA), respectively. Blood samples (200 μL) were collected at 0.017, 0.25, 0.5, 1, 2, 4, 8, 12, and 24 h after injection and then centrifuged at 5000 × g for 10 min immediately. The supernatant plasma samples were stored at −20 °C until analysis. To detect the concentrations of MIT and GA in plasma, 150 μL of methanol was added to 50 μL of plasma samples in order to precipitate proteins. Then, the samples were vortexed for 1 min and centrifuged at 14,000 × g for 10 min. The supernatant was subsequently subjected to liquid chromatography–tandem mass spectrometry (LC-MS/MS, SCIEX Triple Quad™ 5500, USA) for analysis. Chromatographic separation was carried out on an Acquity UPLC column (2.1 × 100 mm, 1.8 µm). The mobile phase A was ultrapure water containing 2 mM FA, and the mobile phase B was ACN. Quantification of the ions was achieved by the multiple reaction monitoring (MRM) mode, in positive mode for mitoxantrone (monitoring the transition of the m/z 445 precursor ion to the m/z 88) and gambogic acid (monitoring the transition of the m/z 629.4 precursor ion to the m/z 545). The major pharmacokinetic parameters were analyzed with Drug and Statistics (DAS) software.Drug distribution in various organ tissues was examined in the orthotopic breast tumor model. Tumors and major organs (heart, liver, spleen, lungs, and kidneys) were isolated from the Balb/C mice after tail vein injection of MG and MG@PM (3 mg/kg of MIT, 4.5 mg/kg of GA) for 24 h. The tissues were homogenized by a homogenizer in PBS and precipitated with methanol. After centrifugation at 12,000 × g for 10 min, the supernatant was taken and then the levels of MIT and GA in the respective tissues were detected by LC-MS/MS following the above method.Anti-tumor efficacyOn day 10 after tumor implantation, murine orthotopic tumor models were assigned randomly to 5 groups, and then PBS, MIT, GA, MG, and MG@PM (3 mg/kg of MIT, 4.5 mg/kg of GA) were injected intravenously once two days for 3 times. Tumor volume was measured with calipers every two days during the period. After 15 days, tumors were isolated for photographing and weighing. The stripped tumors were soaked in 4% paraformaldehyde for 48 h, and paraffin sections were prepared to study the destruction of tumor cells induced by different treatments. Moreover, the sections were evaluated through terminal deoxynucleotidyl transferase-mediated dUTP nick-end labeling (TUNEL) and hematoxylin-eosin staining (H&E) and observed by an Inverted fluorescence Microscope (Nikon, Japan).In vivo immunostimulation experimentFor in vivo immunostimulation studies, the lymph nodes, spleens, and tumor tissues of murine orthotopic tumor-bearing models were collected on day 7, cut into pieces, and incubated in DMEM containing collagenase type I (Sigma, USA), collagenase type IV (Biosharp, China), hyaluronidase (Sigma, USA), and of DNase (Biosharp, Germany) at 37°C for 40 min. Then the sample was passed through 200 mesh nylon strainers to prepare single-cell suspensions and analyzed by flow cytometry. For the analysis of T cells, the single-cell suspensions were stained with anti-CD3-PerCP, anti-CD4-FITC, and anti-CD8-APC antibodies. For the analysis of regulatory T cells (Treg), the single-cell suspensions were stained with anti-CD25-APC, anti-CD4-FITC, and anti-Foxp3-PE antibodies. To assess the maturity of DCs, the single-cell suspensions were dyed with anti-CD11c-FITC, anti-CD86-APC, and anti-CD80-PE antibodies and analyzed by flow cytometry. In addition, sera from different groups of mice were collected on days 6, 7, and 8, respectively. The secreted cytokines in the serum, including TNF-α, IFN-γ, and GZMB were determined by ELISA kits (Absin, Shanghai).In vivo antimetastatic studiesTen days after the orthotopic breast tumor model was established, mice received different treatments on days 0, 2, 4, 6, and 8: 1) PBS, 2) Anti-PD-1 (200 µg/kg), 3) Abraxane (20 mg/kg) + Anti-PD-1 (200 µg/kg), 4) MG@PM (3 mg/kg of MIT, 4.5 mg/kg of GA). 1 × 106 4T1-Luc cells were injected intravenously on day 9 to construct a lung metastasis model. Bioluminescence imaging of mice was recorded by the IVIS Spectrum imaging system every five days. Each mouse received an intraperitoneal injection of D-luciferin sodium solution (150 mg/kg) 20 min before bioluminescence imaging. Mice were executed on day 29. Pulmonary tissues were captured, colored with Bouin’s solution (Servicebio, China), and imaged to calculate lung metastases. Lung tissue sections were also stained with H&E and observed with Olympus Slide View VS200 (Tokyo, Japan). As for survival analysis, the lifespan of the murine tumor-bearing model adopted with the same administration schedule was recorded during the treatment period.Statistics and reproducibilityNo statistical method was used to predetermine sample sizes. For in vitro studies, treatment groups were randomly assigned and were not changed when treatment was given on the culture day indicated. These experiments were completed in replicates and independent experiments. For animal studies, mice were randomly assigned to treatment groups after tumor inoculation. The starting tumor burden in the treatment and control groups was similar before treatment. The investigators were blinded to allocation during experiments and outcome assessment. Overall survival was estimated by Kaplan–Meier methods and compared with log-rank tests. Cox proportional hazard models were used for multivariate survival analysis. Multivariate logistic regression models were used to assess binary outcomes of response to treatment. Pearson’s correlation coefficient was used to assess linear correlations between variables. Two-group was compared using unpaired student’s t-test, and multiple-group analysis was performed using one-way ANOVA followed by multiple comparisons test. In vitro experiments were carried out at least triplicate independent experiments as indicated. All in vivo studies included a minimum of five mice per group for drug therapy studies. All statistical analyses were performed using GraphPad Prism 9 software.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer
