Nanoengineered Disruption of Heat Shock Protein 90 Targets Drug-Induced Resistance and Relieves Natural Killer Cell Suppression in Breast Cancer

The high mortality in breast cancer is primarily due to most late-stage patients relapsing on chemotherapy and becoming resistant to other drugs (1). This is particularly true in breast cancers that are negative for the cell surface HER2 and estrogen and progesterone receptors, known as triple-negative breast cancer (TNBC; ref. 2). Indeed, the primary treatment for TNBC includes taxanes alone or combined with anthracyclines (3). Despite some success, recurrence and resistance happen at a substantially higher rate than other breast cancer subtypes, which associate significantly with diminished likelihood of survival (4, 5). The mechanisms of resistance in TNBC are poorly defined and even emerging modalities, such as immunotherapy, in which drugs aim to reactivate immune cells to induce tumor rejection and eradication (6), have yet to markedly enhance duration of response (7–9). Elucidating the drivers and contributors of resistance and identifying modalities to target these mechanisms in TNBC is, therefore, a critical need toward achieving a sustainable cure.

Intratumor heterogeneity, cancer stem cells (CSC), and mutational evolution have long been implicated as the drivers of both intrinsic and acquired drug resistance (10). An emerging paradigm, however, is drug-induced resistance, or tolerance, which has been described as phenotypic transitions within subpopulations of cancer cells in the presence of drugs (11), which we previously showed can arise from non-CSCs via protein expressions, kinase scaffolding, and signaling activations (12). The heat shock protein 90 (Hsp90) plays a broad role in cellular signaling, including a direct effect on protein kinases, operating as an ATP-dependent dimeric molecular chaperone to form the core of large complexes with cochaperones and substrates (13). Indeed, combinations of Hsp90 inhibitors and chemotherapies have been studied (14) with the goal of targeting multiple prosurvival pathways, including STAT, ERK, Src family kinases, and PI3K families of proteins, which are augmented under external stress (15). However, targeting Hsp90 in clinical studies has been somewhat lackluster (16), suggesting novel approaches that deploy rational combinations of drugs could help to address the existing challenges.

A concerted effort to understand the biological interaction between tumor, stroma, and immune cells within the tumor immune microenvironment (TIME) will contribute to clinical treatment success (17). Not only are the activity and exhaustion status of cytolytic immune cells, such as CD8⁺ cytotoxic T cells and natural killer (NK) cells, implicated in tumor rejection (18), their spatial arrangement and locations within the tumor are critical for prognostic benefit of anticancer cytotoxics and cancer immunotherapies (19). Attempts to improve tumor surveillance via augmenting immune cell activities (20) or suppressing the “do not eat me signals” on tumor cells have been tested (21). Few studies, however, have sought to increase tumor cell surface ligands that invigorate NK-cell or T-cell surveillance, such as MHC class I polypeptide-related sequence A, B (MICA/B; ref. 22), to “unmask” tumors from immune evasion.

The goal of this study was to interrogate the TIME in drug-induced resistance and the role that chaperones contribute as druggable targets in this effect. Using in vitro coculture experiments, molecular and computational screening approaches, and cancer nanomedicines as a tool and in vivo translational models, we describe a tumor-targeted, engineered therapeutic approach that reinvigorates NK cells to combat resistance phenotypes that emerge under drug pressure.

All in vivo experiments were performed in compliance with active Institutional Animal Care and Use Committee protocol approved through Harvard Medical School (Boston, MA) and Brigham and Women's Hospital (Boston, MA), and in accordance with institutional guidelines, supervised on-site by veterinary staff. Mice with tumors were closely monitored by careful clinical examination to detect deterioration of their physical condition and sacrificed at any sign of stress. Human samples for ex vivo experiments were obtained from patients clinically diagnosed with TNBC and were collected by Mitra Biotech under institutional review board approval with written informed consent from each patient.

Radicicol was a kind gift from Dr. Leslie Gunatilaka (University of Arizona, Tucson, AZ). All chemical reagents were of analytic grade, used as supplied without further purification, and purchased from Signal-Aldrich, unless indicated. Recombinant human cytokines were reconstituted in a solution containing 0.1 mmol/L acetic acid and 0.1% BSA (PeproTech).

Human MDA-MB-231, MDA-MB-468, MDA-MB-436, MCF-7, and mammary carcinoma 4T-1 cells (ATCC) and SUM159 (BioIVT) were purchased in the last 10 years and cultured in 10% FBS in DMEM or RPMI media (Invitrogen). Cell lines were validated for absence of Mycoplasma prior to use, by the manufacturer. Cells were used within 10 passages from frozen stock vials obtained from the manufacturer. NK-92MI cells (ATCC) were cultured according to manufacturer's protocol. Primary human peripheral blood CD56⁺ NK Cells (Stem Cell Technologies, catalog #70037) were cultured using Immunocult XF T-cell Expansion Media (Stem Cell Technologies) with 10% FBS and 100 U/mL human recombinant IL2. Three-dimensional tumor spheroid NanoCulture plates were used whenever indicated (MBLI).

siRNA gene knockdown was performed on cells at a concentration of 5 × 10⁴ cell/mL. Prevalidated silencer select siRNA targeting (sense sequences) MICA (siRNAs ID#1: s8772 and ID#2: s458040; Thermo Fisher Scientific), and were transfected using Lipofectamine 2000 (Invitrogen) following the manufacturer's protocol. Scrambled siRNA was used as a control.

Cell viability was measured as described, using 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium (MTS Reagent, Promega) or water-soluble tetrazolium salts (WST Reagent, Dojindo Molecular Technologies Inc.) following the manufacturer's protocol, and absorbance was read at the recommended UV wavelength (450 nm) using BioTek Microplate Reader (BioTek Instruments Inc.). To evaluate the pharmacologic interaction of different combinations of drugs, we followed the method proposed by Chou and colleagues (23).

The Proteome Profiler Human Phospho Array (R&D Systems) was used to identify phosphorylated residues affiliated with different proteins. Following the Bradford protein analysis assay to normalize total protein content, cell lysate from the indicated cell line was applied to the phosphorylation membranes following the manufacturer's protocol. Optical densities were determined by ImageJ Software (NIH.gov) and Adobe CS5. Reference spots were used to normalize between array membranes.

Protein samples were resolved by SDS-PAGE and transferred to polyvinylidene difluoride membranes prior to incubation at 4°C with indicated primary antibodies, mTOR and pMTOR^Ser2448, pAKT^Thr308 and AKT, Phospho-p44/42 MAPK (Erk1/2)^{Thr202/Tyr204}, p44/42 MAPK (Erk1/2) pPRAS40^Thr246, pSTAT3^Tyr705, STAT3, PRAS40, and β-Actin antibodies were purchased from Cell Signaling Technology, pHck^Tyr410 (Thermo Fisher Scientific), MICA/B (R&D Systems), and heat shock factor 1 (HSF1) and HSF1^Ser326 (Abcam). Western blot images chosen as representative depictions in the figures demonstrate equivalent results taken from biological replicates (n ≥ 3).

Cells were cultured as indicated, fixed with 4% paraformaldehyde, washed twice with PBS, and blocked in 10% goat serum (volume for volume). Cells were incubated with fluorescently labeled antibodies for NK group 2 member D (NKG2D), CD158a, NKB1, CD244, and KLRG1 (BioLegend) or MICA/B (R&D Systems) overnight at 4°C and analyzed (C6 Accuri Cytometers Inc.). Data analysis was performed using FlowJo Software (Tree Star Inc.) and Accuri cFlow plus Software to obtain and confirm mean fluorescent intensity (BD Biosciences). Isotype IgG control was used to subtract for background noise.

Human TNBC was collected immediately after surgical resection (See Supplementary Table S2 for metadata). Matched-patient nonheparinized blood (5–10 mL) was also collected at the time of biopsy in BD-Vacutainer tubes (Becton Dickinson) following published protocol and established quality control criteria (24). Tissue slices were maintained in customized tumor matrix protein–coated (TMP) plates as described in prior report (25). Tissue fragments (∼300 μm–2 mm in size) were treated with the indicated drugs at the clinical max concentration (Cmax) for 72 hours as determined by published literature on each drug's pharmacokinetics profile (See Supplementary Table S3 for related drug concentrations used). DMSO was used as a vehicle control.

Media were collected from parental cell lines or drug-tolerant cancer cells (DTCC) cultured for indicated timepoints, centrifuged (“neat”), and the resulting supernatant was aliquoted and stored at −80°C. Using thawed conditioned media (25 μL “neat” media), a panel of 41 cytokines and chemokines was profiled using the MILLIPLEX MAP Human Cytokine/Chemokine Magnetic Bead Panel (MilliporeSigma) according to the manufacturer's instructions and plates were read on the Luminex200 (Luminex Corp.). Analyte measurements were reported using the MILLIPLEX Analyst Software (MilliporeSigma).

4T1 mouse mammary carcinoma cells (1 × 10⁶ cells) were injected into the mammary fat pads of 5- to 6-week-old female balb/c mice (BALB/cAnNCrl, Charles River Laboratory, strain code: 028). Docetaxel was dissolved in pure ethanol at a concentration of 50 mg/mL mixed 1:1 with polysorbate 80 (Tween 80) and brought to a final working concentration with 5% glucose in PBS. Once tumors became palpable (∼100 mm³), docetaxel, radicicol, DocRad-nanoparticle (NP), or vehicle treatments were administered intravenously on indicated days at the indicated doses. Tumor volumes were quantified using digital calipers (Starlett) by a third party unaware of treatment conditions.

Cells were permeabilized by incubation with 0.5% Triton X-100 at 4°C for 10 minutes, washed three times with 1 × PBS-T (1 × PBS + 0.05% Tween 20), and blocked using 10% BSA solution (dilution with PBS-T) at room temperature for 1 hour. The samples were stained with a primary antibodies: Hsp90 (Cell Signaling Technology), phosphorylated HSF1^Ser326 (Novus Biological), and phosphorylated Hck^Tyr410 (Thermo Fisher Scientific). For nuclear staining, the samples were counter stained with 4′,6-diamidino-2-phenylindole (DAPI, Thermo Fisher Scientific). Images were taken on a Nikon Eclipse Ti Camera (Nikon Instruments) with NIS Elements Imaging software (3.10). Confocal fluorescence imaging was performed on Zeiss LSM 800, Airyscan Confocal Laser Scanner Microscope with ZEN 2.3 software. Postprocessing of the images was completed either in ImageJ or ZEN lite software.

For murine tissue IHC, formalin-fixed, paraffin-embedded (FFPE) sections were incubated with the following primary antibodies: phosphorylated PRAS40^Thr246 (clone C77D7), STAT3^Tyr705 (D3A7) (Cell Signaling Technology), CD49b (PA5-87012, Thermo Fisher Scientific), MULT-1 (ABIN966609, Antibodies Online), and Rae-1 (PA5-93166, Invitrogen). Sections were then incubated with a horseradish peroxidase–conjugated secondary antibody (SignalStain Boost IHC Detection Reagent, Cell Signaling Technology). Chromogenic development of signal was performed using 3,3′-diaminobenzidine (DAB Peroxidase Substrate Kit, Vector Laboratories). Terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) was used following the manufacturer's protocol (FITC Kit, GenScript). For ex vivo human tumor experiments, tissue was prepared from FFPE sections in serial 4 μm sections and cut onto charged slides, which were stained with hematoxylin and eosin (H&E) for pathologic determination of tumor viability and area (determined by a clinical pathologist), cleaved caspase-3 in vitro diagnostic (IVD) antibody (catalog no., 229, BioCare), or stained with a 4-plex panel of fluorophore dyes (Opal DAPI; catalog no., FP1490), Discovery FAM (catalog no., 760-243, green), and Discovery Cy5 (catalog no., 760-238)] with corresponding primary marker antibodies [FAM-CD56 IVD antibody (clone# MRQ-420, Ventana, catalog no., 790-4596), Anti-CD3 IVD antibody (clone#2GV6, rabbit monoclonal, Ventana, catalog no., 790-4341), and anti-pan keratin (PanCK; clone AE1/AE3/PCK26, Ventana, catalog no., 760-2595)] selected for profiling NK cells (DAPI⁺PanCK⁻CD56⁺CD3⁻).

H&E stains were annotated digitally by a clinical pathologist (D. Goldman) to designate tumor tissue, nontumor tissue, and stromal areas using the HALO Digital Image Analysis Software version 2.3.1.2089.70 (Indica Labs) to establish tumor, nontumor, and stroma regions of interest (ROI). ROI groups were then trained on the basis of “ground truth” and cell populations were segmented and optimized using the DAPI stain. Once all algorithms had been fully developed, they were bulk applied to the appropriate patient, establishing a dataset identifying the absolute count and spatial distribution of DAPI⁺PanCK⁻CD56⁺CD3⁻ cells in tumor, nontumor, and stromal ROI. HALO Spatial Analysis (Indica Labs) module was used for plotting the NK cell dataset containing the requisite x and y coordinate map. Computer software settings and details are provided in the Supplementary Data. For Spearman correlation analysis, data were normalized and read into the R statistical computing package “car.” Data tables were created by caspase 3–high and caspase 3–low samples (respectively). These data were fed into a variety of visualization packages (GGPlot, GGPairs, scatterplotMatrix, and corr). For correlation analysis, the R corrplot package was utilized to visually interpret the output from the above analysis. Spearman output was visualized by creating a heatmap to express the individual correlation values that were observed.

Cholesterol (500 mg, 1.29 mmol) was dissolved in 5 mL of anhydrous pyridine. Succinic anhydride (645 mg, 6.45 mmol) and catalytic amount of DMAP were added to the reaction mixture to form a clear solution. The reaction mixture was stirred for 12 hours under argon atmosphere. Removal of pyridine was carried out under vacuum and the crude residue was diluted in 30 mL DCM, washed with 1 N HCl (30 mL) and water (30 mL), and the organic layer was separated and dried over anhydrous sodium sulfate, filtered, and concentrated in vacuo. Completion of the reaction was confirmed by TLC in 1:99 methanol:DCM solvent mixture. Radicicol (25 mg) was dissolved in 2 mL anhydrous DCM followed by addition of 1.2 mol/L equivalent of cholesterol hemisuccinate, EDC, and DMAP. The reaction mixture was stirred at room temperature for 48 hours under argon atmosphere. Upon completion of reaction as monitored by TLC, the solvent was evaporated under vacuum and the crude product was purified by column chromatography, eluting with DCM:methanol gradient, to give radicicol-cholesterol conjugate as a yellow solid. The obtained conjugate was analyzed by ¹H NMR and mass spectrometry. Further details on synthesis, characterization, and release kinetics of nanoparticles can be found in the Supplementary Data. See Supplementary Fig. S1 for reaction schemas.

Docetaxel, radicicol-cholesterol conjugate, L-α-phosphatidylcholine, and DSPE-PEG2000 at 0.01:0.09:0.6:0.3 molar ratios were dissolved in 1.0 mL DCM. Resulting clear solution was evaporated and thoroughly dried. The resulting thin film was hydrated with PBS with constant rotation at 60°C for 1 hour to get white turbid solution containing supramolecular nanoparticles (SNP). SNPs were eluted through a Sephadex column and extruded through 0.4 μm polycarbonate membrane using mini-extruder. Nanoparticles solution (10 μL) was diluted to 1 mL using deionized water and three sets of 10 measurements each were performed at 90 degree scattering angle to get the average particle size by dynamic light scattering method using Zetasizer Nano ZS90 (Malvern). The zeta potential was measured using a Zetasizer ZS90 with the nanoparticles diluted in water for measurement according to the manufacturer's protocol.

Drug-loaded nanoparticles (1 mg drug/mL, 5 mL) were suspended in PBS buffer (pH 7.4) and 4T1 cell lysate, and sealed in a dialysis tube (MWCO = 3,500 Dalton, Spectrum Lab). The dialysis tube was suspended in 1 L PBS (pH 7.4) with gentle stirring to simulate the infinite sink tank condition. A 100 μL portion of the aliquot was collected from the incubation medium at predetermined time intervals and replaced by equal volume of PBS buffer, and the released drug was quantified by high-performance liquid chromatography and plotted as cumulative drug release.

Statistical analysis was performed using Prism Software (GraphPad) and determined by ANOVA, followed by a Newman–Keuls post hoc test when values were represented between multiple groups, and, unless otherwise noted, two-tailed Student t test was used to identify statistical significance between individual groups. Two-way ANOVA was employed to track significance between groups from in vivo tumor volume experiments.

To interrogate the activity of NK cells in the presence of drug-naïve versus drug-induced resistant (i.e., tolerant) TNBC cells, we deployed an in vitro coculture model. We generated a population of DTCCs that temporarily display a hybrid epithelial–mesenchymal cell state implicated in therapy failure in multiple humanized models (12). Briefly, the parental TNBC cell line, MDA-MB-231, was exposed to docetaxel, a routine cancer chemotherapy for TNBC (26), at >20× the published IC₅₀ value (Fig. 1A). We cocultured constitutively active NK cells (NK-92MI) or CD56⁺ primary human peripheral blood NK cells with either parental cells or DTCCs to assess cytolytic activity (Fig. 1B). We determined that NK cells were incapable of killing DTCCs compared with parental cells across multiple TNBC models tested with different genetic backgrounds (Fig. 1C; Supplementary Fig. S2A). To study NK-cell activity, we used a coculture experimental design in which tumor spheroids, generated using a NanoCulture system (27), were separated from NK cells by a 0.2-μm porous membrane that restricts diffusion to secreted factors, such as growth factors, cytokines, chemokines, and lipids (Fig. 1D). Using flow cytometry, we evaluated several NK-cell inhibitory and activating biomarkers including the well-established activation marker, NKG2D, which plays a key role in NK-cell activity (18). Notably, the expression of NKG2D, and minimally expressed NKG2C, which binds and elicits immune responses on ligand receptors such as MICA/B (28), was significantly diminished on NK cells in coculture with DTCCs versus the parental cells (Fig. 1E; Supplementary Fig. S2B). In parallel, we noted MICA/B was diminished on cancer cells in response to taxanes and on DTCCs (Supplementary Fig. S2C). To elucidate a mechanism underlying the decrease of NKG2D on NK cells, we isolated cell culture supernatant from parental cells or DTCCs and interrogated the excreted cytokine milieu using multiplex Luminex cytokine analysis over the course of 24 hours (Supplementary Fig. S1F). Clusters of cytokines that affiliated with DTCCs versus parental cells emerged (Fig. 1G), yet a smaller cohort was found to overlap between two independent TNBC cell lines tested, which included regulated on activation, normal T-cell expressed and secreted (RANTES), G-CSF, and GM-CSF (Supplementary Fig. S2D). On the basis of these results, we attempted to phenocopy the DTCC microenvironment in parental cells by introducing the top induced cytokines or a cocktail of those that clustered together by Euclidean distance in the Luminex array (i.e., VEGF, G-CSF, GRO, RANTES, and IL1α). Cell viability analysis confirmed that G-CSF and GM-CSF alone and in combination with other cytokines were the only ones tested to recapitulate the DTCC TIME in a parental cell line (Fig. 1H; Supplementary Fig. S2E). Indeed, using flow cytometry, we confirmed a reduction of NKG2D expression on NK cells (Supplementary Fig. S2F), which is consistent with reports in other physiologic contexts (29, 30).

Next, we deployed systems biology and computational modeling to establish a chemical reaction network that integrated drug-induced protein kinetics with a systems biology approach. We used this strategy to infer drug effect on the DTCC phenotype using a framework that provided some mathematical certainty in which we could “toggle,” in silico, the effect of pathway perturbations. To do this, we interrogated phosphorylation status of proteins in DTCCs compared with parental cells and integrated these observations with empirical evidence from a kinetic analysis of drug-induced protein phosphorylation to establish the system of proteins and parameters that are induced by therapy, rather than “drug selected.” This approach implicated multiple protein families and pathways that are induced in the DTCCs and activated in discrete, time-dependent patterns following drug pressure in drug-naïve parental cells (Fig. 2A; Supplementary Fig. S3A; Supplementary Table S1). The systems biology and chemical reactions network that was built from these empirical observations, and from a review of the literature, identified Hsp90 as a potential “node” with the closest determined relationship between a prosurvival phenotype in DTCCs as well as modulator of NK-cell recognition of tumor cells, which functions via suppression of MICA/B through sequestration of the HSF1 (Fig. 2B; Supplementary Fig. S3B; ref. 31). We confirmed an increased expression of Hsp90 in the DTCCs compared with parental cells using fluorescence microscopy and Western blot analysis (Fig. 2C and D; Supplementary Fig. S3C and S3D) and found that docetaxel drug pressure induced the active form of HSF1 (Ser326; Supplementary Fig. S3E; ref. 32), which appeared to sequester in the perinuclear space of DTCCs versus parental cells (Fig. 2E and F). Disruption of Hsp90 using the macrocyclic antifungal antibiotic, radicicol (33), or various other small-molecule inhibitors, including ganetespib (34) and PU-H71 (35), reversed cytoplasmic sequestration of HSF1^Ser326 and activation of prosurvival proteins in DTCCs, as evidenced by confocal microscopy and Western blot analysis, respectively (Fig. 2E and F; Supplementary Fig. S4A–S4C).

The timing and sequence of anticancer drug combinations are important considerations, which influence responses to therapy (36). Indeed, simultaneous administration of taxanes in combination with targeted inhibitors was previously determined to be suboptimal (12). We tested the hypothesis that sequencing Hsp90 inhibitors and docetaxel can optimize anticancer effects while also improving immune detection by NK cells via NKG2D receptor ligand expression. We tested the anticancer effect of different dosing schedules on cancer cell viability, timing the separation of docetaxel and radicicol in discrete order (Fig. 3A). Values from cell viability analysis were used to plot the fraction affected [F(a)], which indicates the fraction of cells inhibited by treatment administered. A combination index (CI) was calculated at each F(a), CI below 1 indicates synergism and above 1 indicates antagonism (23). Schedule 2 (radicicol → docetaxel) resulted in antagonism. In contrast, schedule 1 (docetaxel → radicicol) resulted in synergism across all cell lines (Fig. 3B; Supplementary Fig. S5A). To validate these results, we simulated protein signaling and cell death by considering the reaction rates of the chemical reaction network, which were used to construct a system of ordinary differential equations to represent the protein and drug dynamics. The genetic algorithm in MATLAB was then used to explore our multi-dimensional parameter space and find a local minimum for the error between the simulation results and the in vitro data. With the given parameter fit, in silico experiments confirmed a direct correlation between docetaxel and radicicol sequencing and the effect on Hsp90 pathway induction and perturbation when drugs are administered in discrete sequence (Supplementary Fig. S5B).

We next tested how Hsp90 disruption affects NK-cell recognition and cytolysis in cells using the optimized temporal schedule (i.e., docetaxel → radicicol). Indeed, Hsp90 inhibition “primed” tumor cells, significantly increasing both the intensity of expression and percent (%) positive expressing MICA/B cells (Fig. 3C and D; Supplementary Fig. S6). Moreover, cocultures of NK cells with DTCCs that had been “primed” were significantly more sensitive to NK-cell cytolysis, as determined by cell viability analyses (Fig. 3E and F; Supplementary Fig. S7A–S7C). We confirmed that this effect was indeed tumor cell dependent by treating NK cells with Hsp90 inhibitors and determining there was no change in tumor cell cytolysis versus vehicle control (Supplementary Fig. S7D and S7E). Finally, we used siRNA gene knockdown of MICA (Supplementary Fig. S7F) and determined that NK cells lost a significant proportion of their cytolytic capacity in MICA-knockdown DTCCs, which we had “primed” by overnight inhibition of Hsp90 (Fig. 3G and H). The in silico modeling data together with in vitro evidence support a rationale for sequencing docetaxel prior to radicicol, but not inversely, as a means to improve antitumor effects while simultaneously repriming tumor cells for NK-cell cytolysis.

Cancer nanomedicines are useful tools to differentially release drug payloads in distinct, controlled, and temporal constraints (37), or for codelivery of two drugs to control spatial distribution of drugs (38). In this study, our evidence suggested that (1) drug order was important to improve the anticancer effect of the combination of docetaxel and radicicol and (2) DTCCs suppress NK cells via prolonged secretion of extracellular factors, which can be remedied by sustained inhibition of Hsp90. Given the physiologic limitations of drugs, which have shortened half lives in vivo compared with in vitro cell cultures, we hypothesized that a two-in-one drug delivery strategy (docetaxel and radicicol chimera) would achieve two goals: (i) fast release of docetaxel will eliminate the drug-sensitive population in the tumor and (ii) sustained release of radicicol will suppress survival and optimally improve tumor immunity in the residual tumor population by boosting expression of NK activity ligand receptors. In silico drug delivery comparisons predicted improved anticancer outcomes using time-delayed “chimeric” approach (Supplementary Fig. S8A and S8B). Next, we engineered a nanoparticle containing both docetaxel and radicicol (DocRad-NP), wherein radicicol was conjugated to cholesterol and held in the lipid bilayer, designed for slow release, while free-form docetaxel was encapsulated for rapid release (Fig. 4A and B). The dual payload nanoparticle was constructed using a thin film hydration followed by an extrusion approach. Dynamic light scattering confirmed the formation of a supramolecular nanostructure of 225 ± 42 diameter (Fig. 4C), where the ζ-potential and size were consistent over time at 4°C (Fig. 4D). Release kinetics showed minimal release of radicicol nanoparticles in PBS (20%) when compared with release in cell lysate (65%) over 125 hours (Fig. 4E). Notably, differential release kinetics of docetaxel and radicicol were observed such that rapid release of docetaxel was seen within 4 hours compared with radicicol (35% vs. 23%, respectively), with a clear separation in release kinetics followed by saturation of docetaxel (70%) observed at 96 hours compared with equivalent level of release of radicicol (P < 0.05), which was not achieved until 120 hours (Fig. 4E). We reasoned that DocRad-NP may serve as a suitable tool to selectively toggle the release of the chemotherapy agent (docetaxel) and the Hsp90 inhibitor (radicicol) in a time-dependent fashion to achieve optimal cell killing and sustained “priming” of any residual cancer mass.

Next, we characterized the pharmacodynamics and in vitro efficacy of DocRad-NP. DTCCs were treated with the radicicol-NP or free form (free drug), followed by an immediate wash-out after 4 hours and analyzed at 16, 36, and 48 hours later by Western blotting (Fig. 5A). The radicicol-NP–treated cells showed sustained inhibition of phosphorylated proteins up to and beyond 48 hours compared with free drug, which showed rescue of signaling disruption by 36 hours in most cases (Fig. 5B). Furthermore, cell viability analysis in drug-naïve parental cells indicated that DocRad-NP achieved 50% cell killing at concentrations below those of the single-drug–loaded nanoparticle individually or together (Fig. 5C). Next, we interrogated MICA/B expression in the context of free-drug radicicol or radicicol-NP. Transient (4 hours) treatment of the radicicol-NP resulted in a significant increase in the percent positive expression of MICA/B on DTCCs compared with the free drug (P < 0.001) and improved cytolytic capacity of NK cells in the residual tumor cell fraction (Fig. 5D–F). These data suggested an improvement of antitumor effects while simultaneous priming of the residual tumor mass for NK-cell surveillance (Fig. 5G).

Next, we used an in vivo immunocompetent orthotopic syngeneic mammary carcinoma model (4T-1 in Balb/C) and treated mice with either docetaxel, radicicol, a combination of individual compounds, or DocRad-NP at equivalent drug concentrations. In a previous report, we showed that MTD of docetaxel chemotherapy in this syngeneic model will induce the DTCC phenotypic transition, in vivo (12). First, we determined that DocRad-NP displayed superior anticancer efficacy compared with equivalently dosed individual or combination drugs via significant reduction in tumor burden over the course of treatment (P < 0.01 by two-way ANOVA; Fig. 6A). Using TUNEL, we evaluated tumor-specific killing and organ toxicity, which confirmed DocRad-NP increased cell death in the residual tumor while avoiding other organs and systemic toxicities (Fig. 6B; Supplementary Fig. S9). Using IHC on the residual tumor tissue, we tested pathway activation of Hsp90 via expression of phosphorylated proline-rich Akt substrate of 40 kDa (PRAS40) and STAT3. IHC indicated that DocRad-NP sustained pathway inhibition up to 4 days beyond the last dosing, evidenced by reduced antibody staining at day 16 compared with the free-drug combination (Fig. 6C). Blinded pathology assessment indicated overexpression of the NKG2D murine ligand receptor, murine ULBP-Like transcript 1 (MULT-1) in the DocRad-NP cohort, while another receptor, retinoic acid early inducible gene 1 (Rae-1), remained at similar expression status across cohorts (Fig. 6D). Serial sections confirmed that regions of high MULT-1 expression tended to localize with increased incidence of CD49b⁺ cells, an indication of NK cells (39), while treatment conditions with regions of low expressing MULT-1 showed minimal infiltration (Fig. 6E).

Our data demonstrate an important dynamic relationship between NK cells and drug-induced death. We sought to confirm the important role of CD56⁺ NK cells in response to multiple clinically approved drugs as they affiliate with anticancer response or resistance in human tumors. To elucidate this, we deployed a human autologous ex vivo tumor model using primary human TNBC (Supplementary Table S2). Fragments of living, fresh tumor biopsies and autologous patient-derived peripheral blood mononuclear cells were cultured on a substrate of TMPs following a previously published procedure (Fig. 7A; ref. 25). To this, we introduced clinically approved (and off-label) anticancer drugs at their respective pharmacokinetics clinical max concentration (C_max; n = 7; Supplementary Table S3).

We developed a “sequential imaging” strategy to study the TIME using 4-μm serial sections from FFPE tissue following drug treatment, ex vivo, to locate: (i) regions of tumor versus stroma (H&E), (ii) drug-induced apoptosis by cleaved caspase-3, and (iii) NK cells via multiplex IHC (mIHC; PanCK⁻CD3⁻CD56⁺), which demarcate a population poised for activity, excreting high amounts of immune-stimulating cytokines (Fig. 7B; Supplementary Fig. S10A and S10B; ref. 40). This imaging approach allowed us to identify regions of tumor that were positive for apoptosis while also allowing us to pinpoint the location of NK cells within tumor versus stroma (see Materials and Methods; Fig. 7C).

We then performed several quantitative measurements. First, drug-induced cleaved caspase-3 in the tumor region determined as a log₂ fold change increase of 0.5 calculated between vehicle and drug (Fig. 7D). To quantify the role of NK cells with high drug-induced death (i.e., caspase-3 Hi), we deployed the HALO mIHC software platform to quantify five independent parameters related to the location and density of NK cells within the tumor versus stroma, as well as the proximity to the tumor interface following drug pressure, which are features of tumor-infiltrating lymphocytes (TIL) that contribute an anticancer effect (19). We then used Spearman correlation rank order analysis to compare these with drug effect. We made two key observations: (i) we identified a higher correlation within the five cellular metrics in the caspase-3 Hi versus Lo cohorts, with a direct affiliation between NK cells within the tumor versus stroma and proximity to tumor interface in relationship to drug-induced caspase-3 (Fig. 7E) and (ii) a significant increase in NK-cell density within the tumor (P < 0.05), as well as a trend toward diminishing of the distance between NK cells and tumor in the caspase-3 Hi versus Lo cohort when comparing the vehicle treatment with drug treatment (Fig. 7F). These preliminary observations support a critical role for the dynamics of NK cells as they are linked to drug-induced cell death in human cancers.

Resolving drug resistance is penultimate to finding a sustainable cure for cancer. While conventional models of drug resistance rely on stochastic mutations conferred through Darwinian evolution, drug-induced resistance is seen as a measure of cellular “fitness,” wherein the entirety of the tumor ecosystem contributes to the effect while under drug pressure. There is a paucity of literature to support how drug-induced resistant cancer cells and other cells, such as NK cells, “cooperate” or “compete” to drive tumor growth. We focused on the role of innate NK cells in drug-induced resistance and on the basis of this information, we engineered potential therapeutic strategies and established Hsp90 as a putative “lynch pin” in the survival signaling pathway while simultaneously “putting the brakes” on the surveillance of NK cells for tumor cell clearance. To some degree, we relied on the systems biology model to establish the shortest molecular relationship among this effect. While this simplified the protein interactions involved and may overestimate the effect of Hsp90 due to its simplicity, it provided the necessary evidence that the effect of Hsp90 on Src, ERK, STAT3, and Akt is significantly changing the cell's response to docetaxel while simultaneously depressing innate immune surveillance.

The role of Hsp90 in oncology as a drug target is not new, Hsp90 has been identified as an upstream regulator of many oncogenes, making it of long interest in cancer, and small-molecule inhibitors have come and gone in the last two decades, with many failing in phase I trials due to lack of efficacy and poor bioavailability (16). More importantly, Hsp90 inhibition irreversibly downregulates expression of critical antigens and activating receptors on NK cells and T cells (41). To address this, we developed a bioengineered strategy that takes advantage of the enhanced permeability and retention to avoid cells in systemic circulation (42). Indeed, the nanoparticle formulation elicited death only in the tumor, improved therapeutic response, and presence of CD49⁺ cells proximal to MULT-1, in vivo. We chose radicicol as the Hsp90 disruptor in our nanoparticle design for the following reasons: (i) our data to this point indicated that radicicol led to the greatest reduction of surviving cells after exposure to NK92 and greatest induction of MICA/B expression compared with the other inhibitors tested, (ii) the free phenolic groups on radicicol enables a simple conjugation chemistry, and (iii) potential for rapid clinical translation, which may be inhibited by clinically deployed agents protected by intellectual property. While the strategy of using nanomedicine to augment the efficacy of Hsp90 inhibitors has been used by other groups, we show the improved benefit of a rational combination of drugs in tandem with nanotechnology to support tumor immunity and drug resistance. Indeed, based on our evidences, sustained inhibition of Hsp90 using drug reformulations in combination with other chemotherapeutic agents may be an approach to revitalize these failed clinical compounds. Future work to understand how nanotherapeutics can aid this effort is needed.

To examine the relevance of NK-cell dynamics under drug pressure in human cancers, we deployed the ex vivo tumor model. The data demonstrated that drug pressure affiliates with dynamic changes to the immune cell populations and their location within the tumor versus stroma without changing the overall number of cells. These evidences have implications in vivo, which suggest a critical role for NK cells and potentially other cytolytic lymphocytes under drug pressure. However, it should be acknowledged that ex vivo human tumor models have their limitations and more work is needed to understand dynamics of the tumor–immune interface as it relates to response versus resistance in humans. For example, controlling for the baseline level of TILs in each tumor sample and the degree to which lymphocytes (e.g., NK cells) infiltrate the tumor fragment during drug treatment prior to posttreatment analysis are not fully established here.

Finally, our findings may impact other drug modalities including immunotherapies. For example, our evidences suggest that engineered disruptors of Hsp90, cytotoxic drugs, and, potentially, immune checkpoint blockade could act in-concert to improve immune recognition, diminish drug-sensitive clones, and invigorate exhausted T cells, respectively. NK-cell therapy is another surging field in immuno-oncology (43). Not only are NK cells the focused target of interest for biological manipulation (44), they are being explored as tools for “off-the-shelf” therapy with a chimeric antigen receptor, including the same NK-92 cell line we deployed in this study (45). An unexplored challenge in this space is the development of “boosters” to augment activity of NK-cell therapy. The data in our report suggest that Hsp90-inhibiting nanovehicles induce NK-activating ligand receptors that have been previously unreported to associate. For example, while MULT-1 has previously been connected to heat shock (46) and xenobiotic stress (47), there is no evidence linking it to Hsp90 inhibition. We hypothesize other stress-related NK-activating ligands may also emerge under pressure of Hsp90 nanovehicles, in vivo. We acknowledge, however, there is more work to be done to understand how critical the NK-cell population was, in vivo, to the anticancer effects that we observed. For example, we determined a correlation between the expression of MULT-1 and CD49b (pan-NK, but also some non-NK-cell targets). Definitive evidence that anticancer effects were mediated by NK cells could improve our understanding for clinical use of Hsp90 and chemotherapy combinations.

Taken together, these findings highlight a potentially novel role for the tumor immune contexture in drug-induced tolerance. They further support a rational approach to sequence cancer therapies using advanced nanomedicines that can deliver drugs in temporal order and sustain inhibition of key signaling pathways to thwart resistance and improve tumor immunity.

M. Smalley reports personal fees from Mitra Biotech Inc. during the conduct of the study. B. Shanthappa reports grants from DOD BCRP during the conduct of the study. M. Pellowe reports other compensation from NSERC during the conduct of the study. S. Sengupta reports other compensation from Invictus Oncology and Farcast Biosciences during the conduct of the study, Farcast Biosciences, and Invictus Oncology, and personal fees and other from Akamara Therapeutics and Vyome Biosciences outside the submitted work, as well as has a patent pending for BWH 2020-687 to Brigham and Women's Hospital. A. Goldman reports personal fees from Mitra Biotech outside the submitted work, as well as has a patent for BWH 2020-687 pending to Brigham and Women's Hospital. No disclosures were reported by the other authors.

M. Smalley: Data curation, formal analysis, methodology, writing-review and editing. S.K. Natarajan: Data curation, formal analysis, methodology, writing-review and editing. J. Mondal: Data curation. D. Best: Data curation. D. Goldman: Formal analysis. B. Shanthappa: Data curation. M. Pellowe: Formal analysis, writing-review and editing. C. Dash: Data curation. T. Saha: Data curation. S. Khiste: Data curation. N. Ramadurai: Formal analysis. E.O. Eton: Data curation, formal analysis. J.L. Smalley: Data curation. A. Brown: Formal analysis. A. Thayakumar: Data curation. M. Rahman: Resources, methodology. K. Arai: Resources, methodology. M. Kohandel: Software, formal analysis, supervision. S. Sengupta: Conceptualization, resources, supervision, funding acquisition, writing-original draft. A. Goldman: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, methodology, writing-original draft, project administration, writing-review and editing.

The authors would like to thank Dr. Leslie Gunatilaka for preparing, purifying, and providing the radicicol necessary for all experiments; Saravanan Thiyagarajan, Hans Gertje, Vipin Menon, and Shruthi Subramanian for helpful discussions during experimentation and providing information for article preparation; Drs. D.C. Doval and Partha Sarathi Ghosh for providing resources during development of the mIHC panel; Dr. Ashish Kulkarni for discussions while characterizing the nanoparticles. S. Sengupta was supported by a DoD BCRP Breakthrough Award and an NIH UO1 (1U01CA214411). A. Goldman was supported by an American Cancer Society Postdoctoral Fellowship (122854-PF-12-226-01-CDD) and Breast Cancer Alliance Young Investigator Award. Medical images and illustrations used for schematics were obtained from http://www.servier.com.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.