Targetable S100-RAGE signaling mediates intrinsic trastuzumab deruxtecan resistance in metastatic breast cancer

Abstract

Purpose:

Trastuzumab deruxtecan (T-DXd) has improved outcomes in metastatic breast cancer; however, a substantial subset of patients experience early lack of clinical benefit that is not reliably predicted by routine clinicopathologic variables. In an institutional cohort of 109 T-DXd–treated metastatic lesions, HER2 expression level, ER/PR status, Ki-67 index, and metastatic site were not significantly associated with response, highlighting the need to define mechanistic determinants of intrinsic resistance.

Experimental Design:

We performed spatial transcriptomic and proteomic profiling using NanoString GeoMx Digital Spatial Profiling on pretreatment bone, brain, and soft-tissue metastases from patients with clinical benefit (response) versus early progression (resistance) on T-DXd. Tumor (PanCK⁺) and immune (CD45⁺) compartments were analyzed to link tumor architecture and region-resolved signaling states with therapeutic response. Candidate resistance pathways were functionally evaluated in HER2-positive breast cancer cell lines and in vivo metastasis models treated with T-DXd alone or combined with the RAGE inhibitor TTP488.

Results:

Spatial proteogenomics revealed recurrent upregulation of S100 family alarmins in resistant tumor regions, associated with activation of a RAGE-centered pro-survival signaling program characterized by ERK, AKT, and STAT3 phosphorylation. Pharmacologic RAGE inhibition enhanced T-DXd–induced apoptosis, restored drug sensitivity in vitro, and significantly reduced metastatic burden in lung and brain metastasis models.

Conclusions:

Spatial proteogenomics identifies a conserved S100-RAGE–driven survival state coupled to immune-excluded tumor architecture as a mechanism of intrinsic T-DXd resistance across metastatic niches. Targeting this pathway with an orally available RAGE antagonist restores T-DXd responsiveness and offers an immediately translatable strategy to overcome resistance in metastatic breast cancer.

Introduction

Antibody-drug conjugates (ADCs) have reshaped therapy for metastatic breast cancer by coupling tumor-targeting antibodies to highly potent cytotoxic payloads^1,2. Trastuzumab deruxtecan (T-DXd) is a leading example, combining a humanized anti-HER2 antibody to a topoisomerase I inhibitor through a cleavable tetrapeptide-based linker, enabling efficient intracellular release of its payload^3–5. In the phase 3 DESTINY-Breast03 trial⁴, T-DXd significantly improved progression-free survival and produced almost double the objective response rate compared with trastuzumab emtansine, establishing its role as the preferred second-line therapy for HER2-positive metastatic breast cancer. More recent results data from the phase 3 DESTINY-Breast09 trial demonstrate that T-DXd–based combinations outperform current first-line therapy in metastatic HER2 positive breast cancer, with early results suggesting that T-DXd may soon become a first-line standard for HER2-positive metastatic disease⁶. DESTINY-Breast04⁵ and DESTINY-Breast06⁷ have shown clinically significant activity of T-DXd in HER2 low and ultra-low metastatic breast cancer respectively. As T-DXd adoption expands to > 90% metastatic breast cancer patients, intrinsic resistance and early non-response have emerged as major barriers to durable benefit.

Despite these advances, treatment outcomes vary widely. HER2 abundance alone does not explain clinical variability. In the DAISY trial⁸, response rates declined progressively across HER2 expression categories, and even among HER2-IHC 3-positive tumors, approximately 30% of patients failed to respond to T-DXd, highlighting that intrinsic resistance can operate independently of HER2 abundance. Moreover, metastatic lesions exhibit pronounced spatial heterogeneity in tumor architecture, stromal organization, immune infiltration, and signaling states, features that can critically shape ADC efficacy by influencing target accessibility, internalization, payload delivery, tissue penetration, and local survival signaling^9–11.

Metastatic niches further complicate this problem. Bone, brain, and soft-tissue metastases represent distinct tumor ecosystems with divergent cellular composition and physical constraints that shape drug access, payload penetration, and tumor-microenvironment interactions^12–14. Bone metastases arise within a mineralized and cellularly diverse marrow niche that shapes tumor survival, nutrient availability, and therapeutic access. Brain metastases must traverse the blood-brain barrier and adapt to a neurovascular environment dominated by astrocytes, microglia, neurons, and tightly regulated vascular units. Soft-tissue metastases, including pericardial and chest-wall lesions develop within fibro-inflammatory stroma with prominent extracellular matrix remodeling and local immune responses. A central unresolved question is whether early failure of T-DXd reflects site-specific microenvironmental adaptations or instead arises from shared, actionable tumor states that recur across metastatic environments.

To address this question, we performed integrated spatial transcriptomic and proteomic profiling of metastatic lesions obtained immediately prior to T-DXd initiation from patients with HER2-positive breast cancer. Our cohort included bone and brain metastases, the two dominant sites of dissemination in HER2-positive disease, as well as soft-tissue metastases involving the pericardium and chest wall, which represent clinically relevant but less molecularly characterized metastatic niches. Using GeoMx Digital Spatial Profiling (DSP), we interrogated region-specific molecular programs within PanCK-positive tumor cells and CD45-positive immune compartments and compared these spatially defined features between patients who achieved clinical benefit (response) and those who exhibited early progression (resistance) on therapy. We hypothesized that intrinsic T-DXd resistance is encoded not by anatomical site per se, but by shared microenvironmental architectures and tumor-intrinsic signaling programs that recur across metastatic environments.

To test whether these resistance-associated pathways were therapeutically actionable, we validated candidate mechanisms using in vitro cellular models and in vivo metastasis assays. Through this integrated spatial-mechanistic approach, we identified a conserved S100 family-RAGE signaling axis as a central mediator of intrinsic T-DXd resistance across metastatic sites. We further demonstrate that pharmacologic RAGE inhibition restores T-DXd sensitivity by enhancing apoptotic signaling, establishing this pathway as both a predictive spatial biomarker and a rational, immediately translatable combination strategy to improve ADC efficacy in HER2-positive metastatic breast cancer.

Materials and Methods

Clinical cohort

We retrospectively analyzed 47 patients with metastatic breast cancer and 109 metastatic lesions treated with T-DXd at Emory University between Oct 2022 to Mar 2025. Clinical variables collected included age, ER, PR, HER2 status, prior systemic therapies, metastatic sites, and radiographic response assessments. Metastatic sites at treatment initiation included brain (n=9), bone (n=21), soft tissues including chest wall and pericardium (n=14), lymph nodes (n=17), and other visceral or cutaneous sites. Patients received T-DXd 5.4 mg/kg intravenously every 3 weeks, consistent with the FDA-approved dosing schedule. Patients were categorized at first restaging by RECIST v1.1 as having clinical benefit (complete response, partial response, or stable disease) versus early progression (progressive disease). In parallel, progression-free survival (PFS) was calculated from T-DXd initiation to radiographic progression or death, and patients without events were censored at last follow-up. The study was approved by the institutional IRB (# PRO00038108) and conducted in accordance with the Declaration of Helsinki. Informed consent was obtained per institutional requirements; consent for retrospective data analysis was waived when appropriate.

Spatial proteogenomics analysis

Tissue preparation and morphological staining

We retrospectively collected metastatic tissue samples from six breast cancer patients with one or more metastatic lesions. The samples included two bone metastases, two brain metastases, one pericardial metastasis, and one chest wall metastasis. Formalin-fixed, paraffin-embedded (FFPE) tissue sections (5 μm) were mounted on standard microscope slides and processed for spatial proteogenomics analysis using the NanoString GeoMx DSP (https://brukerspatialbiology.com/products/geomx-digital-spatial-profiler/geomx-protein-assays/io-proteome-atlas/). To enable tissue compartmentalization and spatial segmentation, sections were stained with a panel of three fluorescently labeled morphological markers: anti-pancytokeratin (PanCK-AF532) to identify epithelial tumor cells, anti-CD45 (CD45-AF594) to delineate immune cell populations, and Syto13 nucleic acid stain to visualize nuclei and define cellular boundaries. Staining was performed according to standard immunofluorescence protocols with overnight incubation at 4°C.

Patch-based spatial enrichment and immune abundance analysis

Spatial interactions between tumor and immune compartments were quantified using a patch-based co-occurrence framework applied to multiplex GeoMX DSP immunofluorescence images. Whole-section images were down-sampled for computational stability, and binary masks for tumor (PanCK⁺) and immune (CD45⁺) compartments were generated using fixed color-channel thresholding followed by morphological cleanup. Each image was subdivided into fixed-size patches (32×32 pixels for the primary analysis). Within each patch, tumor–immune neighborhood enrichment was calculated by measuring the observed boundary co-occurrence between adjacent tumor and immune pixels (4- or 8-connected adjacency) and comparing it to a null distribution generated by 1,000 random spatial permutations of tumor and immune masks within the patch while preserving compartment proportions. An enrichment Z-score was computed as (Observed − μ_random)/σ_random, where μ_random and σ_random represent the mean and standard deviation of permuted co-occurrence values. Positive Z-scores indicate greater-than-random tumor–immune adjacency within a patch, whereas lower Z-scores reflect reduced enrichment relative to the random expectation. In this dataset, enrichment values were predominantly positive, indicating varying degrees of spatial engagement rather than complete spatial avoidance. Patch-level Z-scores were aggregated to generate lesion-level enrichment metrics.

In parallel, immune abundance was quantified using the same patch-based framework. CD45⁺ area fraction per patch was calculated as the number of CD45⁺ pixels divided by total tissue pixels within the patch (excluding background). For tumor-restricted infiltration analysis, CD45⁺ fraction was additionally computed within PanCK⁺ tumor masks. Patch-level immune density metrics were aggregated to generate lesion-level measures for comparison between responder and non-responder samples.

Spatial proteogenomics workflow

The GeoMx spatial proteogenomics assay was performed following the manufacturer's two-day protocol. On day one, tissue sections were incubated overnight with a cocktail of DNA-barcoded RNA in-situ hybridization (ISH) probes targeting the whole transcriptome atlas (WTA). On day two, sections underwent overnight incubation with DNA-barcoded antibodies for simultaneous protein detection. Following completion of the assay protocol, slides were imaged using the GeoMx DSP instrument.

Region selection and segmentation

Regions of interest (ROIs) were manually selected based on tissue morphology and tumor architecture as visualized by the fluorescent morphological markers. Selected ROIs were computationally segmented into distinct areas of interest (AOIs) based on marker expression: tumor compartments were defined as PanCK-positive areas, immune compartments as CD45-positive areas, and nuclear regions by Syto13 staining. This segmentation approach enabled independent molecular profiling of spatially distinct tissue microenvironments within the same histological section.

Data collection and library preparation

DNA barcodes corresponding to RNA and protein targets were photocleaved and collected from each segmented AOI using sequential UV light exposure. Collected barcodes were processed using the GeoMx library preparation protocol with unique dual indexing (i7/i5) for each AOI to enable multiplexed analysis.

Data preprocessing and quality control

Spatial proteogenomic and transcriptomic data were generated on the GeoMx DSP platform and processed with the GeoMxWorkflows R package (v1.0+) following the standard NanoString pipeline. Raw DCC files were imported with corresponding PKC files and sample annotations using readNanoStringGeoMxSet. Segment-level QC required: minimum total reads ≥1,000, trimmed reads ≥80%, stitched reads ≥80%, aligned reads ≥75%, sequencing saturation ≥50%, negative control counts ≥1, NTC counts ≤9,000, nuclei count ≥20, and segment area ≥1,000 μm². Probe-level QC was performed with setBioProbeQCFlags (minimum probe ratio 0.1; maximum percentage failing Grubbs test 20%; local outlier removal enabled), and probes failing these metrics or flagged as global Grubbs outliers were excluded. Probe counts were aggregated to genes with aggregateCounts. For each segment, the limit of quantification (LOQ) was defined as the geometric mean of negative controls plus 2 SD, with a minimum LOQ of 2 counts. Segments with <5% gene detection were removed, and genes detected above LOQ in <10% of segments were excluded (negative control probes retained for normalization). The final dataset comprised 7,709 genes across 65 high-quality segments.

Spatial multi-omics data from all metastases samples were obtained from both PanCK+ tumor cell regions and CD45+ immune cell regions. Sample identifiers were standardized by combining scan names with ROI identifiers to ensure proper matching between expression data and metadata across both data modalities. Low-expression gene filtering was applied independently to brain and bone datasets using stringent criteria: genes were retained only if they showed ≥1 count in ≥3 samples per tissue type to ensure adequate gene coverage.

Normalization

To enable cross-modality comparison, RNA and protein data were normalized with a unified spatial procedure that adjusts for ROI size and cellularity. For each ROI, counts were scaled to counts per 1,000 μm² and per 10,000 nuclei usingnormalized_counts = raw_counts × (1000 / Area [μm²]) × (10,000 / Nuclei_count).For RNA, this spatial correction was applied to the GeoMx-processed counts prior to DESeq2 modeling. For protein, raw counts were corrected using the same formula; normalized values were then rounded to integers and any negative values set to zero to retain count-like properties for DESeq2.

Differential expression analysis

Differential RNA and protein expression between treatment responders and non-responders was evaluated separately for PanCK+ tumor ROIs and CD45+ immune ROIs, with independent analyses for bone and brain metastases to capture tissue-specific biology. For each tissue × cell-type stratum, DESeq2 (RRID:SCR_000154) models were fit with response status as the primary factor; area and nuclei count were included as covariates to control residual metadata effects. For protein analyses in PanCK+ bone ROIs, features with <5 total counts across all samples were removed before modeling. Non-responders were set as the reference group for fold-change estimation. P values were adjusted by Benjamini–Hochberg FDR; significance was defined as FDR < 0.05 with an additional biological threshold of |log2FC| > log2(1.5).

Targeted S100 family RNA and protein expression comparisons

Expression of S100-family genes/proteins was extracted from all ROIs and compared between response groups using two-sided Welch’s t-tests. Distributions were visualized with violin plots overlaid with jittered ROI points and embedded box plots.

Principal component analysis

Variance-stabilizing transformation (VST) was applied to spatially normalized counts using vst (DESeq2) with blind = FALSE to preserve group structure. PCA was computed with plotPCA; samples were colored by response (responder vs non-responder). PC1/PC2 coordinates were exported and visualized with ggplot2 (RRID:SCR_014601) to assess clustering and potential confounders.

Pathway enrichment

Significantly dysregulated proteins were mapped to gene symbols by manual curation and automated lookup. Enrichment was performed with clusterProfiler (RRID:SCR_016884) (KEGG, organism “hsa”)(KEGG RRID:SCR_012773), GO Biological Process, and ReactomePA. Over-representation was tested by hypergeometric tests with FDR correction (q < 0.05 for KEGG/Reactome; q < 0.20 for GO to accommodate pathway redundancy).

Cell lines and cell culture

A panel of human breast cancer cell lines was obtained from the (American Type Culture Collection) ATCC. Cells were maintained in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin at 37°C with 5% CO₂. All cells used in this study were routinely tested for mycoplasma contamination and used within six passages.

Cell viability assay

BT474 (RRID: CVCL_0179) and SKBR3 (RRID: CVCL_0033) cells were seeded in 96-well plates (5,000–10,000 cells/well) and treated with T-DXd (0.01–10 μg/mL) alone or in combination with TTP488 (0.5 μM) for 72 hours. Cell viability was measured using CellTiter-Glo Luminescent Cell Viability Assay (Promega) and calculated as percentage relative to vehicle control. IC50 values were determined using four parameter non-linear regression method (https://www.aatbio.com/tools/ec50-calculator). CRISPR-based RAGE-depletion in SKBR3 cells were conducted using RAGE CRISPR Plasmids (Sant Cruz Biotechnology, sc-400284-KO-2). Data represent mean ± SEM from three independent experiments performed in triplicate.

Immunostaining assay

Cells were seeded on glass coverslips and treated with vehicle control, TTP488 (0.5 μM), T-DXd (4 nM), or the combination for 72 hours. Following treatment, cells were fixed in 4% paraformaldehyde and permeabilized with 0.2% Triton X-100. Immunofluorescence staining was performed using antibodies against cleaved Caspase-3 (Casp3)(RRID: AB_725947) and BCL-2 (RRID: AB_2715467), followed by Alexa Fluor 488–conjugated secondary antibodies. Nuclei were counterstained with DAPI. Images were acquired under identical exposure conditions using a FV3000 confocal microscope with a 20× and 63× objective.

Quantification was performed using a custom Python pipeline (OpenCV-based, python-bx RRID: SCR_024202). The green (Casp3⁺ or BCL-2⁺) and blue (DAPI) channels were extracted, thresholded (intensity cutoff = 60 a.u.), and binarized. Connected components in each channel were counted as discrete objects. The ratio of green to blue components was defined as the Casp3⁺ or BCL-2⁺ ratio, representing relative apoptotic or anti-apoptotic cell prevalence per field. Data from ten 20x field of view and 3 independent experiments per condition were pooled. Violin plots show group distributions with median and 95% confidence intervals; statistical comparisons were performed using Welch one-way ANOVA test plus Games-Howell.

In vivo studies

All animal studies were approved by the Institutional Animal Care and Use Committee guidelines of Houston Methodist Research Institute (# IS00007758). Female nude mice were housed under specific pathogen-free conditions with ad libitum access to food and water. On day 0, SKBR3 cells (1.75×10⁵ in 100ul PBS) were injected into the left cardiac ventricle or tail vine of immunodeficient mice. Treatments began 3 days post-injection and continued for 4 weeks. Mice were randomized into four groups (n=10 per group): vehicle control, T-DXd alone, TTP488 alone, and T-DXd plus TPP-488. T-DXd was administered intravenously at 10 mg/kg every three days. TTP488 was given orally at 3 mg/kg once daily. At 24 hours after the last dose, mice were euthanized and lung, liver, kidney, and brain tissues were harvested, fixed in formalin, and paraffin-embedded. Hematoxylin and eosin (H&E) staining was performed, and metastatic burden was assessed in a blinded fashion using three parameters: (i) number of metastatic foci per section, (ii) cumulative metastatic area fraction across all organs, and (iii) mean lesion size in lung and brain were collected for tumor burden assessment.

Statistical analysis

Clinical predictors of T-DXd response were evaluated using univariable and multivariable logistic regression models. Group comparisons for categorical variables were performed using chi-square or Fisher’s exact tests, as appropriate. Continuous variables were compared using two-sided Student’s t-tests or Wilcoxon rank-sum tests based on distributional assumptions. Statistical significance was defined as p < 0.05. All clinical analyses were performed in R version 4.2.1 using packages survival (3.3–1) and rms (6.3–0). Unless otherwise indicated, all tests were two-sided, and multiple hypothesis testing was corrected using the Benjamini–Hochberg false discovery rate (FDR).

For in vitro assays, pairwise comparisons between treatment conditions (T-DXd vs T-DXd + TTP488 combination) were performed using Welch one-way ANOVA followed by Games-Howell’s post hoc test.

For in vivo metastasis experiments, differences in metastatic nodule counts, metastatic area, and protein expression (e.g., cleaved caspase-3) were analyzed using one-way ANOVA with Tukey’s correction or Kruskal-Wallis test with Dunn’s correction when data were non-normally distributed. Data are presented as mean ± SEM unless otherwise specified.

Tissue-specific molecular analyses (bone vs brain metastases) were conducted separately for spatial transcriptomic and proteomic datasets, followed by integrative qualitative comparisons to identify shared versus site-restricted mechanisms of T-DXd response.

Data availability statement

The raw GeoMX data in the study has been deposited at GEO database with session number GSE327983. Additional raw data supporting the findings of this study are available from the corresponding author upon reasonable request.

Results

Clinical variables do not reliably predict response to T-DXd

In our curated cohort of 47 metastatic breast cancer patients contributing 109 metastatic lesions obtained prior to T-DXd treatment (Supplementary Table 1), we evaluated clinical and pathological predictors of treatment outcome using univariable and multivariable logistic regression analyses (Table 1). Patients were categorized according to clinical benefit, defined as complete response (CR), partial response (PR), or stable disease (SD), versus progressive disease (PD) at first restaging based on RECIST v1.1 criteria. Among the 47 patients, 31 achieved clinical benefit, whereas 16 experienced progressive disease. Because multiple lesions were sampled from individual patients, multivariable analyses were performed using mixed-effects logistic regression models with patient ID specified as a random intercept to account for within-patient clustering.

Table 1.

Univariable and multivariable mixed-effects logistic regression analysis of clinical predictors of T-DXd response.

Characteristics	Total(N)	Univariate analysis		Multivariate analysis
		Odds Ratio (95% CI)	P value	Odds Ratio (95% CI)	P value
Specimen site	109	0.856 (0.707 – 1.037)	0.113	0.859 (0.660 – 1.116)	0.255
ER intensity	105
3	64	Reference		Reference
2	18	0.714 (0.241 – 2.114)	0.543	0.297 (0.065 – 1.370)	0.120
0	19	1.705 (0.502 – 5.791)	0.393	1.175 (0.192 – 7.198)	0.862
1	4	1.364 (0.133 – 13.933)	0.794	0.257 (0.015 – 4.382)	0.348
ER percentile	107	0.997 (0.986 – 1.008)	0.559
PR intensity	102
0	51	Reference		Reference
2	14	1.042 (0.282 – 3.848)	0.951	1.061 (0.195 – 5.763)	0.945
3	29	0.682 (0.261 – 1.784)	0.435	0.614 (0.145 – 2.608)	0.509
1	8	2.917 (0.330 – 25.805)	0.336	3.135 (0.293 – 33.555)	0.345
PR percentile	104	0.997 (0.986 – 1.008)	0.574
HER2	99
3	15	Reference		Reference
2	59	1.300 (0.405 – 4.168)	0.659	0.841 (0.159 – 4.458)	0.839
1	21	2.133 (0.505 – 9.010)	0.303	1.198 (0.162 – 8.846)	0.860
0	4	10434240.5266 (0.000 – Inf)	0.989	11524191.4202 (0.000 – Inf)	0.993
Prior TDM1 Exposure	102
Y	16	Reference		Reference
N	86	4.534 (1.492 – 13.776)	0.008	6.665 (1.502 – 29.580)	0.013

Note: Clinical benefit was defined as PR, or SD at first restaging per RECIST v1.1 criteria. Analyses were performed using mixed-effects logistic regression models with patient ID specified as a random intercept to account for non-independence of multiple lesions obtained from the same patient.

Metastatic sites represented among the 109 lesions included brain (n=9), bone (n=21), soft tissue (n=14), lymph node (n=17), and other visceral or cutaneous locations (lung, liver, pleura, and skin). No statistically significant difference in T-DXd clinical benefit was observed between patients with brain metastases and those with extracranial disease (p>0.05), supporting preserved T-DXd activity in the central nervous system (CNS)¹⁵.

HER2 expression was classified per ASCO-CAP 2023 criteria. Among the 109 metastatic lesions analyzed, 81 were classified as HER2-low (IHC 1+, 0, or IHC 2+ with FISH−) and 18 as HER2-high (IHC 3+ or IHC 2+ with FISH+) based on ASCO-CAP 2023 criteria, whereas HER2 status was not available for 10 lesions. Consistent with published evidence demonstrating T-DXd efficacy in HER2-low tumors^5,16, there was no statistically significant association between HER2 status and response to T-DXd (p >0.05) in mixed-effects analyses accounting for lesion clustering within patients. ER status, PR status, and Ki-67 index were likewise non-predictive (p>0.05).

In contrast, prior exposure to trastuzumab emtansine (T-DM1) was significantly associated with reduced T-DXd response in both univariable (OR=4.534, p=0.008) and multivariable analyses (OR=6.665, p=0.013), suggesting that T-DM1 treatment may desensitize tumors to benefit from T-DXd, consistent with potential cross-resistance between HER2-targeted antibody-drug conjugates¹⁷. Overall, these findings underscore the limited predictive value of conventional clinicopathologic variables and motivate investigation of biologically grounded mechanisms underlying intrinsic or early resistance to T-DXd.

Spatial proteogenomic profiling reveals distinct microenvironmental features in bone and brain metastases associated with response to T-DXd

To investigate molecular and architectural determinants of T-DXd response across metastatic sites, we performed spatial multi-omics profiling on pretreatment bone and brain metastases from patients with HER2-positive breast cancer (Figure 1A–B, Supplementary Table 2). Using GeoMx DSP, we analyzed four metastatic samples for spatial proteogenomics: two bone metastasis (one responder, one resistant, 30 and 26 ROIs, respectively) and two brain metastasis (one responder and one resistant, 22 and 24 ROIs, respectively). ROIs were segmented into tumor-enriched (PanCK+) and immune-enriched (CD45+) compartments to enable compartment-specific analyses (Figure 1C).

Figure 1. Spatial proteogenomic profiling of T-DXd treatment response in bone and brain metastases.

A. Schematic overview of the study workflow. Bone or brain metastatic lesions were collected from four patients immediately prior to T-DXd treatment. Spatial transcriptomic and proteomic profiling was performed using NanoString GeoMx DSP, and patients were classified as responders or resistant following clinical assessment.

B. Whole-section immunofluorescence images from representative responders and resistant cases in bone (left) and brain (right) metastases. Scale bars = 1 mm.

C. High-resolution immunofluorescence images of tumor–immune regions of interest (ROIs) before and after ROI sampling. PanCK+ tumor compartments are outlined in cyan; CD45+ immune compartments are outlined in orange. Scale bars = 50 μm.

D. Additional representative high-magnification fields illustrating microarchitectural differences across conditions, including compact tumor clusters with immune exclusion in resistant lesions and interdigitated tumor–immune interfaces in responders. Scale bars = 50 μm.

E. Quantification of tumor–immune interface length per 32 × 32-pixel patch in bone metastases. Responders exhibit significantly longer tumor–immune contact zones than resistant lesions (p = 5.5 × 10⁻¹²).

F. Neighborhood enrichment scores of tumor–immune spatial interactions across bone and brain metastases. Responders show consistently higher enrichment values relative to resistant lesions, indicating increased spatial mixing.

G. Quantification of tumor–immune interface length per patch in brain metastases. Responders again display markedly greater interface lengths compared with resistant lesions (p = 1.1 × 10⁻¹⁰). Violin plots show distribution and mean ± SEM.

To extend and contextualize these findings, we additionally performed whole-transcriptome DSP on two independent metastatic lesions, including a responding pericardial metastasis (22 ROIs) and a resistant chest wall metastasis (18 ROIs). Although limited in number, these pretreatment biopsies span anatomically distinct metastatic niches that are rarely accessible and were analyzed at high spatial resolution across hundreds of ROIs, enabling comparative analysis of tumor architecture and signaling states associated with therapeutic response.

Whole-section immunofluorescence imaging revealed marked architectural differences between bone and brain metastases and between responding and resistant lesions (Figure 1D). In bone metastases, responding lesions showed fragmented PanCK⁺ tumor islands extensively interdigitated with CD45⁺ immune cells, whereas resistant lesions exhibited consolidated tumor masses with immune cells restricted to peritumoral regions. These qualitative patterns were supported by quantitative spatial analyses that responding lesions exhibited higher tumor-immune spatial engagement (interface length)(Figure 1E) and immune abundance within tumor regions (CD45 area fraction within PanCK+ tumor) (Figure 1F), whereas non-responding lesions displayed higher tumor-tumor neighborhood enrichment scores, consistent with a tightly clustered, immune-exclusion phenotype (Figure 1G).

Brain metastases displayed even sharper contrasts, with responders showing a more open tumor-immune interface and resistant lesions forming dense, cohesive epithelial structures with minimal immune penetration (Figure 1E–F). Among all samples, resistant brain metastases exhibited the strongest immune-excluded phenotype, whereas responding lesions across both anatomical sites showed greater tumor-immune intermixing, underscoring the potential role of microenvironmental organization in shaping T-DXd sensitivity.

To further characterize immune composition, we performed CIBERSORT-based deconvolution analysis to estimate immune cell-type enrichment from the GeoMx data. These results have now been included in Supplementary Figure 1A–B, where we compared immune cell-type–associated signatures between responder and resistant groups. Consistent with our spatial analyses, these data suggest that differences in response are more strongly associated with tumor-immune spatial organization than large shifts in overall immune cell composition.

Conserved transcriptional programs distinguish T-DXd response and resistance

To define molecular determinants of T-DXd efficacy, we next profiled region-resolved transcriptional programs within PanCK+ tumor and CD45+ immune compartments. Principal component analysis of tumor-enriched ROIs revealed clear separation between responding and resistant lesions (Figure 2A), with bone and brain metastases intermingling within each response group, indicating that response-associated tumor programs are partly conserved across anatomical sites. In contrast, immune-enriched ROIs showed weaker stratification by treatment response (Figure 2B), suggesting site-dependent and more heterogeneous immune states.

Figure 2. Molecular characterization of T-DXd response and resistance across bone and brain metastases.

A–B. Principal component analysis (PCA) of spatial transcriptomic profiles from PanCK+ tumor regions (left) and CD45+ immune regions (right). Points represent ROIs from bone (circles) and brain (squares). Samples cluster distinctly by treatment outcome, with responders (blue) and resistant lesions (red) separating along major principal components.

C–D. Volcano plots showing differentially expressed genes (DEGs) between responders (blue) and resistant lesions (red) within PanCK+ tumor compartments from bone (C) and brain (D) metastases. Genes passing FDR-adjusted significance thresholds are highlighted; selected top response-associated and resistance-associated genes are annotated.

E. Venn diagram illustrating the overlap of T-DXd–associated transcriptional programs between bone and brain tumor cells. Of 314 DEGs in bone and 200 DEGs in brain, 22 genes were shared between both metastatic sites.

F. Log2 fold-change values (Resistance vs Response) for the 10 shared T-DXd resistance-associated genes, demonstrating consistent directionality across bone and brain tumor compartments.

G–I. Violin plots showing RNA expression levels of S100A7, S100A8/A9, and S100P across bone, brain, and soft-tissue tumor cells. Resistant lesions exhibit significantly higher expression of S100 family members compared with responders. *p < 0.05, *p < 0.01 (two-sided Wilcoxon test).

Differential gene expression analysis identified both shared and site-specific markers of T-DXd sensitivity and resistance. Responding bone and brain metastases demonstrated enrichment of metabolic and detoxification pathways (for example GST family members, FABP7, HMGCS2 and LGALS3BP), whereas resistant tumors across both sites upregulated inflammatory and cytoskeletal regulators, prominently including S100A7 and S100A9 (Figure 2C–D). Venn analysis confirmed a core set of response-associated genes (n = 22) and resistance-associated genes (n = 10) shared across bone and brain lesions (Figure 2E–F), supporting the presence of a conserved transcriptional architecture underlying T-DXd response.

Given their consistent enrichment in resistant tumors across both metastatic sites, we examined S100A7 and S100A9 expression in greater detail. Violin plots showed significantly elevated expression of both genes in resistant tumors in bone, brain and soft-tissue metastases (Figure 2G–I). Their recurrent upregulation, together with known roles in inflammation, immune remodeling and therapeutic resistance^18,19, highlights the S100 family as a potential mediator of de-novo T-DXd resistance.

To evaluate the broader clinical relevance of S100 signaling, we analyzed its association with progression-free survival in an independent chemotherapy-treated breast cancer cohort (n = 1,936)(https://kmplot.com/analysis/). High S100 biomarker expression (mean expression of S100A7 and A9) was significantly associated with shorter PFS (HR = 1.15–1.58; log-rank p = 0.0002; Supplementary Fig. 2A). Furthermore, multivariate analysis stratified by HER2 status in this cohort demonstrated no significant difference between HER2-positive and HER2-negative tumors (p = 0.1039; HR = 0.97–1.43). These findings suggest that S100 signaling is unlikely to be specifically associated with the HER2 antibody component. In the spatial omics cohort, a strong inverse correlation between S100A7 and A9 score and PFS was observed (Spearman R = −0.77), although this did not reach statistical significance due to limited sample size (p = 0.103; Supplementary Fig. 2B–C).

Pathway analysis reveals distinct biological processes underlying T-DXd response in tumor and immune compartments

To uncover mechanisms driving differential T-DXd responses, we performed pathway enrichment analysis on differentially expressed genes from tumor and immune regions. In PanCK+ tumor cells, resistant lesions showed strong enrichment of cell cycle and DNA repair pathways, as well as extracellular matrix and adhesion programs (Figure 3A). In contrast, responding tumors were enriched in active cell proliferation, and metabolic processes, including peptide elongation and lipid metabolism. These patterns may suggest that resistant tumors activate DNA repair and structural remodeling programs that may blunt T-DXd cytotoxicity, whereas responding tumors maintain high metabolic and translational activity that may increase susceptibility to the drug. Complete pathway enrichment outputs, including pathway name, adjusted p values, and biological category are provided in Supplementary Tables S3.

Figure 3. Pathway-level determinants of T-DXd response and resistance across tumor and immune compartments in bone and brain metastases.

A. Dot plot of significantly enriched biological pathways in PanCK+ tumor cells comparing responders and resistant lesions. Dot size reflects pathway impact scores, and color intensity corresponds to statistical significance (–log10 p value). Distinct pathway programs characterize Response versus Resistance groups, including proliferative and metabolic pathways in responders and DNA repair or extracellular matrix pathways in resistant tumors.

B–C. Volcano plots showing differentially expressed genes (DEGs) in CD45+ immune compartments from bone (B) and brain (C) metastases. Responders (green) and resistant lesions (orange) exhibit distinct immune transcriptional signatures. Total genes analyzed and the number of significant DEGs (FDR-adjusted) are indicated; representative genes enriched in each group are labeled.

D. Pathway enrichment analysis of CD45+ immune regions displaying programs associated with Resistance (left) and Response (right) to T-DXd. Pathway impact is represented by bar length, and statistical significance by color saturation (–log10 p). Resistance-associated immune programs include ER stress, protein misfolding, immunoglobulin production, and ECM remodeling, whereas response-associated programs include T-cell activation, cytokine signaling, antigen presentation, and metabolic fitness pathways.

The immune microenvironment exhibited pronounced tissue specificity. Bone and brain CD45+ regions displayed largely distinct transcriptional profiles, with many more differentially expressed genes detected in brain metastases (Figure 3B–C). Despite this heterogeneity, clear functional trends emerged. Immune compartments from resistant samples were enriched in pathways related to ER stress, protein misfolding, chromatin remodeling, immunoglobulin synthesis, and fibrosis (Figure 3D). Conversely, responding immune regions showed enrichment for T cell activation, cytokine signaling, interferon response, antigen presentation, cytoskeletal organization, and oxidative metabolism. These findings indicate that T-DXd response is associated with an immune microenvironment that is transcriptionally active and metabolically competent, while resistance corresponds to a stressed, remodeled, and potentially dysfunctional immune state.

Proteomic profiling identifies HER2 signaling and S100 family proteins as key determinants of T-DXd response across metastatic sites

To complement the transcriptomic analyses, we performed spatial proteomic profiling of PanCK+ tumor regions from bone and brain metastases. Resistant bone metastases showed increased expression of S100P and CD36, whereas responding tumors demonstrated higher Hsp27, MX1, and AIF. In brain metastases, resistant lesions displayed elevated HER2, S100A8/A9, Fibronectin, and Tuberin, indicating a partially overlapping yet tissue-specific resistance profile (Figure 4A–B).

Figure 4. Proteomic determinants of T-DXd response and resistance in PanCK⁺ tumor cells from bone and brain metastases.

A–B. Heatmaps showing significantly altered protein expression profiles in PanCK⁺ tumor regions from bone (A) and brain (B) metastases. Each column represents an individual ROI, and each row corresponds to a quantified protein (z-scored log₂ expression). Samples are annotated by clinical outcome (Response = blue; Resistance = red). Distinct proteomic programs characterize responders and resistant lesions within each metastatic site.

C–D. Protein–protein correlation analyses with linear regression and 95 percent confidence intervals.

C. In bone metastases, phosphorylated HER2 (pY877) shows a significant positive correlation with S100P expression; points colored by Response (blue) and Resistance (red). Pearson’s r and p values shown.D. In brain metastases, S100A8/A9 expression strongly correlates with total HER2 levels, with corresponding Pearson’s r and p displayed.

E–G. Violin plots comparing expression of key proteins in bone PanCK⁺ tumor regions:

E. HER2, F. Topoisomerase I, and G. S100P. Dots represent individual ROIs; central lines denote medians.

H–J. Violin plots comparing expression of the same markers in brain PanCK⁺ tumor regions:H. HER2, I. Topoisomerase I, and J. S100A8/A9. Dots represent individual ROIs; medians shown.

K. Protein-expression heterogeneity matrix summarizing the coefficient of variation (CV percent) for indicated proteins across four groups: Bone_Responder, Bone_Non-responder, Brain_Responder, and Brain_Non-responder. Warmer colors indicate greater intra-group heterogeneity. Non-responder groups show consistently higher heterogeneity, reflecting molecular diversity associated with resistance.

Given the prominence of HER2 and S100 family proteins in both datasets, we examined their expression relationships. In bone metastases, phosphorylated HER2 (phospho-Y877) correlated positively with S100P (Pearson's r = 0.604, p < 0.001), and both were higher in resistant tumors (Figure 4C). In brain metastases, S100A8/A9 strongly correlated with total HER2 (Pearson's r = 0.723, p <0.001), again enriched in resistant lesions (Figure 4D). These findings suggest that S100 signaling may intersect with HER2 pathway activity to promote T-DXd resistance.

Quantitative comparisons revealed clear site-specific resistance mechanisms. In bone metastases, HER2 protein levels did not differ between groups, whereas Topoisomerase I, the DXd payload target, was significantly reduced in resistant tumors, alongside elevated S100P (Figure 4E–G). In contrast, brain-resistant tumors demonstrated markedly higher HER2 expression without differences in Topoisomerase I, accompanied by increased S100A8/A9 (Figure 4H–J). The paradoxical elevation of HER2 in resistant brain lesions may suggest altered HER2 functional states or impaired ADC internalization rather than insufficient target abundance.

To evaluate intra-tumoral protein heterogeneity, we calculated the coefficient of variation (CV) for key markers including HER2, S100A8/A9 and S100P across responder and resistant groups (Figure 4K). HER2 showed consistently high heterogeneity (CV: 9.8–18%), with phospho-HER2 displaying slightly lower variability (7.8–15.2%). S100A8/A9 and S100P exhibited moderate to high heterogeneity (CV ranges: 5.9–14.4% and 5.7–16.7%, respectively), with the greatest variability observed in resistant bone samples. Topoisomerase I showed the widest CV range (5.5–18.8%), particularly elevated in resistant brain samples, while Topoisomerase IIα remained more uniform across groups. Overall, resistant lesions demonstrated greater heterogeneity across multiple proteins, suggesting that resistant tumors may harbor more diverse molecular states. This increased variability may reflect coexisting resistance programs and clonal selection pressures, reinforcing the link between S100 family signaling, altered HER2 activity, and reduced Topoisomerase I levels in T-DXd resistance.

RAGE inhibition improves sensitivity to T-DXd through caspase-3 activation

S100 family proteins are calcium-binding alarmins that signal primarily through the RAGE receptor to activate NF-κB and MAPK survival pathways. Although GeoMX spatial proteomics showed no significant differences in total RAGE levels between responders and resistant lesions (Figure 5A), multiple downstream RAGE effectors were markedly upregulated in PanCK⁺ tumor regions of resistant lesions, including pAKT1(T450) (95% CI 0.17–0.64, p=0.0014), pERK1/2(T202/Y204; T185/Y187) (95% CI 17.2–130.7, p=0.009), and pSTAT3(Y705) (95% CI 3.87–29.94, p=0.0067) (Figure 5B). These results may suggest that S100-RAGE signaling amplifies prosurvival ERK, AKT, and STAT3 activity in resistant tumors, despite unchanged RAGE abundance.

Figure 5. TTP488 overcomes T-DXd resistance by activating caspase-3 and suppressing BCL-2.

A. Violin plots showing RAGE protein expression in PanCK⁺ tumor ROIs (left) and CD45⁺ immune ROIs (right) comparing T-DXd Response (blue) versus Non-response (red) groups. Statistical comparisons shown above plots.

B. Violin plots of downstream RAGE-pathway signaling proteins measured by DSP: NF-κB p65, AKT1 (phospho T450), ERK1 (phospho T202/Y204) + ERK2 (phospho T185/Y187), and STAT3 (phospho Y705). Non-responders show significantly elevated activation of AKT, ERK, and STAT3. P-values indicated above each pairwise comparison.

C. Western blot of S100 family protein expression across HER2+ breast cancer cell lines (e.g., CRL-2330, BT-474, MB231-HER2, CRL-2338, SKBR3), confirming differential endogenous S100 levels. β-Actin serves as loading control.

D. Western blot showing HER2 expression across the same panel of cell lines.

E. Western blot showing RAGE and p-p65 expression of RAGE knockout, non-target control and wild-type SKBR3 cells.

F. SKBR3 wild-type (WT) and RAGE-knockout (RAGE KO) cells were treated with T-DXd (4 nM) for 72 hours with or without TTP488 (0.5 μM), as indicated. Cell viability was normalized to the WT vehicle control. Data are presented as mean ± SEM with individual values shown. Statistical significance was assessed by Welch one-way ANOVA test followed by Games-Howell post hoc test. * p<0.05; *** p<0.001; ns non-significant.

G. Representative immunofluorescence images of cleaved caspase-3 and BCL-2 in SKBR3 cells treated with control, TTP488, T-DXd, or combination therapy. Nuclei counterstained with DAPI (blue). Scale bar = 100 μm.

H-I. Quantification of cleaved caspase-3-positive and BCL-2-positive cell ratios in SKBR3 cells across treatment groups. Combination therapy induces the highest apoptotic signal. Data shown as mean ± SEM. ***p < 0.001.

J. Western blot of cleaved caspase-3 and γH2AX (DNA damage marker) in SKBR3 cells, confirming enhanced apoptosis and DNA damage with combination therapy. β-Actin is loading control.

K. Densitometric quantification of cleaved caspase-3 expression from panel J, normalized to control. Combination therapy significantly increases apoptosis relative to monotherapies. *** p < 0.001.

We next screened HER2⁺ breast cancer cell lines for endogenous S100 expression (Figure 5C–D). BT474 cells exhibited robust S100 expression, whereas SKBR3 showed minimal levels. Co-treatment with the RAGE inhibitor TTP488 (0.5 μM, non-toxic) significantly sensitized both lines to T-DXd. In SKBR3 cells, the T-DXd IC₅₀ decreased from 13 nM to 0.27 nM, and in BT474 cells from 28 nM to 0.04 nM. We also tested combination on 4T1.2-HER2 and MB231-HER2 cells. Combined treatment with T-DXd (4 nM) and TTP-488 (0.5 μM) resulted in significantly lower viability than T-DXd alone in both models (Supplementary Figure 3).

To directly evaluate the functional role of RAGE signaling in mediating resistance to T-DXd, we generated CRISPR-based RAGE-depleted SKBR3 cells, which showed markedly reduced RAGE protein expression (Figure 5E). RAGE depletion significantly sensitized SKBR3 cells to T-DXd, with an estimated 11.7-fold decrease in IC₅₀. This was accompanied by attenuation of canonical downstream signaling, including reduced phosphorylation of p65, consistent with suppression of NF-κB pathway activity (Figure 5E). Notably, pharmacologic inhibition with the RAGE antagonist TTP488 did not further sensitize RAGE-deficient cells, supporting that the effects of TTP488 are largely mediated through on-target inhibition of RAGE (Figure 5F). Collectively, these data provide genetic evidence that RAGE signaling is a functional mediator of T-DXd resistance.

To determine whether RAGE inhibition enhances T-DXd-induced cell death, we examined apoptotic signaling downstream of topoisomerase I–mediated DNA damage. Combination treatment markedly increased cleaved caspase-3 expression relative to T-DXd alone or TTP488 alone (Figure 5G–H, J–K), indicating that RAGE blockade augments T-DXd-triggered apoptotic execution rather than independently inducing apoptosis. Consistently, BCL2 protein levels were significantly reduced only in the combination group (Figure 5G, I), supporting a cooperative mechanism whereby TTP488 potentiates T-DXd cytotoxicity by suppressing anti-apoptotic signaling and enhancing caspase-3 activation.

To evaluate whether RAGE inhibition enhances T-DXd efficacy in vivo, we established SKBR3-derived lung and brain metastasis models and treated mice with vehicle, TTP488, T-DXd, or the combination. Across both metastatic sites, combination therapy produced the most pronounced reduction in metastatic burden. In the lung model (Figure 6A–B), the number of metastatic foci per section was significantly reduced in the combination group compared with control (p = 0.0092) and T-DXd alone (p = 0.01), whereas TTP488 alone showed only a mild, non-significant reduction (p = 0.09). Mean metastasis size (Figure 6C) followed the same trend: combination therapy significantly decreased lesion size relative to T-DXd monotherapy (p = 0.0013), while TTP488 alone showed no effect (p = 1), and T-DXd alone produced only an intermediate reduction (p = 0.06 vs control). Total metastatic tumor burden, quantified as cumulative tumor area fraction per lung (Figure 6D), further demonstrated the superior efficacy of the combination treatment. The combination group exhibited significantly lower tumor area compared with control (p = 0.0061) and a trend toward reduction relative to T-DXd alone (p = 0.11). TTP488 monotherapy again showed no significant effect (p = 0.79). These data indicate that RAGE inhibition not only reduces lesion number but also restricts overall metastatic expansion.

Figure 6. TTP488 enhances T-DXd efficacy and promotes caspase-3 activation in metastatic mouse models.

A. Representative H&E-stained lung (left) and brain (right) sections from SKBR3 metastasis–bearing mice following 4-week treatment with control, TTP488, T-DXd, or the combination. Metastatic lesions are outlined in red. Scale bar = 2 mm.

B. Number of metastatic foci per section showing a significant reduction with combination therapy compared to control and monotherapy groups (p = 0.0092 vs control; p = 0.01 vs T-DXd).

C. Mean metastasis size per lesion, showing a strong reduction in the combination group (p = 0.0013 vs T-DXd; trend vs control p = 0.06).

D. Cumulative tumor area fraction per lung demonstrating marked suppression with combination treatment (p = 0.0061 vs control; p = 0.11 vs T-DXd).

E. Western blot of cleaved caspase-3 from lung lysates across treatment groups.

F. Quantification of cleaved caspase-3 levels normalized to control, showing significantly elevated apoptosis in the combination group compared with T-DXd monotherapy (***p < 0.001, unpaired t-test). Data are mean ± SEM (n = 3 mice per group).

To determine whether enhanced apoptosis contributed to this improved response, we examined cleaved caspase-3 expression in lung tissues. Western blot analysis (Figure 6E) and quantification (Figure 6F) showed that combination-treated mice had a marked increase in cleaved caspase-3, approximately a 2.3-fold increase relative to vehicle control (p < 0.001), whereas T-DXd or TTP488 alone induced only modest or minimal changes.

In our studies, the selected T-DXd dose of 10 mg/kg administered twice weekly was considered tolerable because it did not produce evidence of excessive systemic toxicity based on serial body weight monitoring (Supplementary Figure 4). Across the treatment period, mice receiving T-DXd alone showed only a mild early suppression of weight gain relative to controls, but did not demonstrate progressive weight loss or marked deterioration over time. Importantly, body weights remained close to baseline throughout the study, without the degree of decline typically considered indicative of significant treatment-related toxicity. In the combination arm, mice receiving T-DXd + TTP-488 maintained body weight trajectories comparable to, or slightly better than, those in the T-DXd-only group, further supporting that this dosing regimen was not associated with excess systemic intolerance.

Together, these quantitative analyses demonstrate that RAGE inhibition significantly enhances T-DXd efficacy across multiple dimensions of metastatic burden by reducing the number of metastatic lesions, shrinking their size, decreasing total tumor burden, and amplifying apoptosis. These data establish TTP488 as a compelling therapeutic strategy to overcome T-DXd resistance in HER2-positive metastatic breast cancer.

Discussion

In this study, we integrated spatial transcriptomic and proteomic profiling with functional in vitro and in vivo analyses to identify a spatially conserved, mechanistically actionable resistance program associated with intrinsic T-DXd resistance in metastatic HER2-positive breast cancer. Across bone, brain, and soft-tissue metastases, resistant lesions consistently showed elevated expression of S100-family alarmins, including S100A7, S100A9, and S100P. These proteins converged on activation of a RAGE-associated signaling program, accompanied by increased phosphorylation of ERK, AKT, and STAT3. Importantly, this program was observed despite preserved or even elevated HER2 expression, indicating that resistance arises from altered cellular state rather than target loss. While our data do not fully delineate the downstream wiring of this pathway, the reproducible enrichment of S100-RAGE signatures across anatomically distinct metastases strongly supports its role as a survival-prone tumor state that persists despite high HER2 expression.

By integrating spatial architecture with signaling state, our findings extend beyond correlative biomarker discovery and instead implicate a defined biological program that can be functionally perturbed. Spatial imaging further highlighted microarchitectural features linked to intrinsic resistance. Resistant lesions demonstrated compact epithelial architecture with restricted tumor-immune interfaces, whereas responding lesions exhibited more open and interdigitated boundaries. These observations support the concept that ADC efficacy is shaped not only by target expression and intracellular signaling, but also by local tissue organization, which may act as a physical and signaling scaffold that constrains immune access and payload penetration²⁰. Although spatial analysis alone cannot establish causality, the concordance between immune exclusion, S100-RAGE activation, and therapeutic failure provides a compelling structural context for understanding heterogeneous responses to T-DXd.

These findings may help address a clinical challenge: conventional clinicopathologic variables, including HER2 level, ER/PR expression, proliferative index, or metastatic site, fail to predict benefit from T-DXd²¹. In our real-world cohort of 109 metastatic lesions from 47 patients, none of these clinicopathologic variables were significantly associated with response, consistent with observations from DAISY and other prospective studies. The identification of a spatially conserved resistance program thus offers a biologically grounded explanation for why HER2 expression alone is insufficient as a predictive biomarker and highlight the importance of tumor-state- and microenvironment-defined resistance mechanisms.

A key strength of this work is that the identified resistance pathway is immediately druggable. Pharmacologic RAGE inhibition with TTP488, an orally available agent with established Phase 2/3 human safety data^22,23, powerfully resensitizes T-DXd–resistant tumors. In vitro, TTP488 reduced the IC₅₀ of T-DXd by nearly two orders of magnitude in both SKBR3 and BT474 cells and enhanced apoptotic execution via caspase-3 activation and BCL2 suppression. In both lung and brain metastatic mouse models, combination therapy significantly reduced metastatic foci, lesion size, and cumulative tumor burden, surpassing the efficacy of T-DXd monotherapy and demonstrating that the S100-RAGE axis is not merely a biomarker of resistance but a functional driver that can be therapeutically exploited. These findings establish a mechanistically grounded and readily translatable therapeutic strategy to overcome intrinsic T-DXd resistance.

Beyond these translational implications, this study demonstrates the unique power of spatial proteogenomics to resolve intratumoral heterogeneity that is invisible to bulk sequencing approaches. The distinct tumor architectures, immune exclusion patterns, and regionally confined signaling programs identified here by spatial profiling illustrate why bulk sequencing approaches have been insufficient for predicting ADC response. Our results support a precision oncology framework in which spatially resolved biology is used to identify microenvironment-driven drug resistance, guide patient selection, and inform rational combination therapies^20,24,25. This framework is particularly relevant for ADCs, whose efficacy depends on both target engagement, internalization, payload delivery, and tissue penetration, features that cannot be inferred from HER2 expression alone²⁰.

We acknowledge several limitations. This spatial proteogenomic cohort size was modest, reflecting the difficulty of obtaining high-quality pre-treatment biopsies from metastatic sites such as bone and brain. However, the consistency of the S100-RAGE signal across independent patients, metastatic locations, and molecular modalities, together with functional rescue in vitro and in vivo mitigates concerns regarding stochastic findings. In addition, our in vivo experiments were performed on immunodeficient models, which do not capture the full contribution of RAGE signaling in myeloid or adaptive immune populations. We attempted to establish an orthotopic 4T1.2-HER2 model to evaluate treatment in a more clinically relevant, spontaneous metastasis setting. However, in our hands, tumors implanted in the mammary fat pad underwent spontaneous regression in immunocompetent mice, even without treatment (Supplementary Figure 5). By 14 days after inoculation, complete tumor regression was observed in 12 of 20 mice. Given the expression of human HER2 in this model, we speculate that this reflects immune-mediated rejection of the exogenous antigen, thereby preventing stable tumor establishment. As a result, we were unable to perform delayed treatment initiation studies in this system. Finally, the absence of longitudinal biopsies precluded analysis of resistance evolution under therapeutic pressure, an important direction for future investigation.

In summary, our study identifies the S100-RAGE pathway as a central, spatially conserved, and targetable mediator of intrinsic T-DXd resistance across metastatic sites, establishes spatial tumor-immune architecture as a functional determinant of ADC efficacy, and provides a clinically translatable strategy with an already available therapeutic agent for therapeutic rescue. By integrating clinical response data with spatial transcriptomics, spatial proteomics, mechanistic signaling assays, and in vivo therapeutic validation, we establish a coherent biological model in which S100-driven RAGE activation and microarchitectural immune exclusion jointly define early T-DXd failure. This integrated spatial–mechanistic–therapeutic framework enables biomarker-driven patient stratification and provides a translational path for targeting the S100-RAGE axis to overcome intrinsic T-DXd resistance in HER2-positive metastatic breast cancer.

Significance Statement.

Intrinsic resistance to trastuzumab deruxtecan (T-DXd) limits treatment benefit for a substantial subset of patients with metastatic breast cancer, yet conventional clinical and pathological variables do not reliably predict response. In our institutional cohort of 109 metastatic lesions, HER2 status, hormone-receptor expression, Ki-67 index, and metastatic site showed no significant association with treatment outcome, underscoring the need for biologically grounded predictors of intrinsic resistance. Through spatial proteogenomic profiling of clinical metastases, we identify a conserved S100-RAGE survival program that characterizes resistant tumors across anatomic sites. Pharmacologic RAGE inhibition restores T-DXd–induced apoptotic signaling and suppresses metastatic burden in vivo, establishing a clinically actionable combination strategy and supporting development of spatial biomarkers for patient stratification.