Introduction

Breast tumors with a germline BRCA1/2 mutation have a defective homologous recombination DNA double-strand break repair pathway.[1] Poly(ADP-ribose) polymerase (PARP) inhibitors are FDA-approved to treat advanced BRCA1/2-related breast cancers.[2] PARP inhibitors work by stalling and collapsing replication forks thereby stopping DNA repair and eventual cell death.[3,4] Nevertheless, about 40% to 70% of patients fail to respond or will develop resistance to PARP inhibitors over time.[5–7] The high rate of resistance has driven efforts to develop combination therapies that enhance the effectiveness of PARP inhibitors, as well as to identify biomarkers to predict which patients are most likely to benefit.

MicroRNAs (miRNAs) are small (∼20 nucleotides), non-coding RNAs that have emerged as useful biomarkers for the detection of early cancers, including breast cancer.[8–12] As such, miRNAs may provide informative molecular readouts of PARP inhibitor response and resistance in breast tumors. However, miRNA measurements are often performed in bulk tissue, which obscures intratumoral heterogeneity and removes spatial information.[13] We previously showed that integrated analysis of expression and spatial localization of seven miRNAs in endpoint murine breast tumors following a full treatment course discriminated PARP inhibitor-sensitive from-resistant tumors.[14] These findings suggest that spatially resolved miRNA measurements can capture treatment-associated tumor states that are not readily accessible from bulk assays.

One strategy to overcome PARP inhibitor resistance involves co-targeting PARP and the PI3-kinase (PI3K) pathway. PI3K inhibitors can suppress the growth of tumors with PIK3CA mutations,[15] and are also effective in tumors with wild-type PIK3CA.[16] In addition to their role in tumor proliferation, PI3K enzymes contribute to the DNA damage response: PIK3CB is required for double-strand break sensing by regulating NBN recruitment to damaged DNA.[17] Building on this mechanistic rationale, we and others have demonstrated that combining a PARP inhibitor with a PI3K inhibitor produces significant preclinical and clinical activity in BRCA1/2-related cancers.[18–21] We have also previously published that the effectiveness of PARP inhibition depends on the tumor immune microenvironment (TME).[22–24] Olaparib increased infiltration and activation of CD8⁺ T-cells [23] and tumor-associated macrophages.[22] Therefore, another potential combination therapy could be combining PARP with an immune stimulant such as polyinosinic:polycytidylic acid (Poly(I:C)). Poly(I:C) activates dendritic cells by binding to various surface receptors such as toll-like receptor 3 which in turn activates CD8⁺ T-cells.[25] These mechanistically distinct treatment conditions provide a biologically relevant setting in which to evaluate spatial molecular readouts of treatment response and resistance.

In this work, we used our previously described hydrogel-based, nanoliter well array in situ miRNA technology [14,26] to assess tumor spatial miRNA profiles in FFPE sections after 10 days of treatment. To interpret this high-dimensional spatial readout, we developed an analytical framework integrating latent Dirichlet allocation (LDA) and principal component analysis (PCA) to identify early, spatially resolved miRNA profiles associated with treatment response. We further developed a specialized, priming-based LDA analysis to improve differentiation between treatment regimens among PARP inhibitor-sensitive tumors. Finally, we evaluated whether incorporating intratumoral immune composition, by computing Structural Similarity Index Measure (SSIM)[27] maps between immune cell infiltration and miRNA topics, could enhance interpretation of the LDA-derived spatial topics by relating them to underlying cellular architecture and highlighting regions where immune patterns and miRNA distributions co-localize. This integrative spatial miRNA assay and analysis provides a framework for linking treatment-associated miRNA patterns to tumor and immune architecture and may inform prognostic or predictive biomarker development in BRCA1/2-related breast cancers.

Results

Combination therapies produced promising therapeutic effects in PARP inhibitor-resistant tumors

We used the K14-Cre Brca1^f/f Trp53^f/f mouse tumor model to investigate the effects of combining a PARP inhibitor, olaparib, with either a PI3K inhibitor (i.e., alpelisib) or Poly(I:C) (Fig 1A). This BRCA1-related triple-negative breast cancer (TNBC) model is known to be sensitive to PARP inhibition [28], as well as in combination with alpelisib.[18] Our current results using orthotopic xenotransplanted tumor fragments from this model were consistent with prior findings: olaparib significantly lengthened the median time taken to reach study endpoint of 20 mm compared to control (75-78 days versus 19-23 days, p<0.001; Fig 1B.i). Combination therapies further lengthened the time to reach 20 mm when compared to olaparib alone (olaparib+alpelisib 115 days versus 75 days, p<0.001, Fig 1B.i; olaparib+Poly(I:C) 120 days versus 78 days, p<0.001, Fig 1B.ii and S1 Fig). This study demonstrated for the first time that Poly(I:C) enhanced PARP inhibition, with an effect size similar to alpelisib in olaparib-sensitive tumors.

Fig 1.

We derived olaparib-resistant tumors from initially sensitive tumors grown orthotopically under constant olaparib treatment.[29] Median survival was shorter in mice with resistant tumors compared to those with sensitive tumors (resistant 23-24 days versus sensitive 75-58 days; Fig 1B). Olaparib+alpelisib but not olaparib+Poly(I:C) improved the outcome in resistant tumors when compared to olaparib alone (olaparib+alpelisib 39 days versus 23 days, p=0.018, Fig 1B.iii; olaparib+Poly(I:C) 30 days versus 24 days, p=0.567, Fig 1B.iv and S1 Fig). This means that alpelisib but not Poly(I:C) was able to at least partially overcome PARP-inhibitor resistance.

Spatial miRNA analysis

Given that the combination therapies improved treatment outcomes, we investigated the extent to which different treatments altered the spatial miRNA profiles of the tumors. We performed our previously described hydrogel-based, nanoliter-well in situ miRNA assay [14,26,30–32] on tumors (n=21) harvested 10 days after treatment when the TME remodeling and transcriptional reprogramming are likely to be complete but not yet confounded by extensive necrosis (Fig 2A).

Fig 2.

Our in situ miRNA assay simultaneously quantified seven miRNAs commonly studied in breast cancer (Fig 2B, S2 Fig).[33–37] MiR-21-5p, let-7a-5p, and miR-19a-3p were highly abundant in these K14-Cre Brca1^f/f Trp53^f/f tumors (S3 Fig). Cursory univariate analyses determined that both sensitive and resistant tumors treated with olaparib+Poly(I:C) had lower let-7a expression compared to olaparib alone (both adj.p=0.01; S3 Fig). The expression of the other six miRNAs did not differ between treatment groups or sensitive/resistant tumors in this bulk analysis (S3 Fig).

Since analysis of individual miRNA expression levels was insufficient to inform tumor phenotype or capture tumor-level miRNA organization, we hypothesized that a spatially informed assessment of the tumors may reveal underlying trends, as evidenced by our prior work.[14] We applied a “naïve” implementation of an LDA-based framework, whereby we trained an LDA model on all the 10-day treated tumors (over 115,248 data points), where each point corresponded to a miRNA amount per unit of space in a tumor section. This case-study-wide LDA model learned underlying features in the form of “topics” that were themselves composed of miRNA probabilities, which could be visualized as word clouds (Fig 2D.i).

The power of the LDA topics is in the method of application; the associated word probabilities can be leveraged as weights that can be applied sequentially to each spatial unit on any given sample (explained in S4 Fig) such that each pixel is effectively associated with the topic that corresponds to the highest weighted value. This can be visualized colorimetrically by associating topics with colors that can be assigned spatially, allowing for further examination (Fig 2D.ii, S4 Fig). A more holistic analysis can be applied by summarizing the associated topic percentages within all tumor sections (Fig 2D.iii) which not only allows easier viewing of topic trends across all samples but can also be used to stratify the data via PCA (Fig 2D.iv). The resultant PCA showed that the tumors mostly stratified along the first principal component (PC1, 96% variance explained; Fig 2D.iv). Olaparib-sensitive tumors occupied the positive PC1 region, with significantly higher PC1 scores than resistant tumors (median PC1 0.42 vs −0.083; Mann-Whitney U test, p <0.001; Fig 2D.iv).

Topic 2, dominated by miR-21, was found to contribute positively to PC1, while Topic 1, dominated by let-7a, contributed negatively (Fig 2D.iv). This means that miR-21 was associated with olaparib sensitivity while let-7a was associated with resistance. Notably, the tumor-specific survival trends in Fig 1 were foreshadowed by spatial miRNA patterns observed in day 10 tumors (Fig 2D.ii, S5 Fig).

Within olaparib-resistant tumors, the case-study-wide LDA model stratified tumors that received olaparib alone and those receiving combination therapies (Fig 2D.iv). Combination treatments altered the miRNA profiles of resistant tumors, shifting them toward the profiles of sensitive tumors, specifically by increasing the proportion of the miR-21-dominated topic within the tumors (Fig 2D.iii). Within the olaparib-sensitive tumors, there was little stratification between olaparib-only and combination treatment regimens (Fig 2D.iv), despite the survival study demonstrating an approximately two-fold increase in survival with combination therapies compared with olaparib alone (Fig 1B). Together, these data indicate that the miRNA profiles distinguishing olaparib-sensitive from-resistant tumors reflect underlying functionally ‘sensitive’ and ‘resistant’ cell state rather than a simple pharmacodynamic response to treatment.

Refined spatial miRNA analysis of PARP inhibitor-sensitive tumors

Since the case-study-wide spatial LDA analysis (Fig 2C) did not provide pharmacodynamic information specific for mono- or combination therapy in olaparib-sensitive tumors as shown in Fig 1B, we performed a specialized analysis to improve the stratification of olaparib-sensitive tumors and investigate whether early spatial miRNA profiles could differentiate tumors that received combination therapy versus olaparib-only treatment (Fig 3A). Specifically, we examined whether this early stratification could reflect the greater efficacy of the combination regimens observed in Fig 1.

Fig 3.

The core of this specialized analysis lies in isolating the sensitive tumor samples and examining how their LDA-derived miRNA probabilities differ from those learned from all the tumors. First, we trained an independent LDA model using only olaparib-sensitive tumors to obtain miRNA probabilities (S2 Table), which were subsequently used to prime the spatial miRNA input for olaparib-sensitive tumors (Fig 3A). The highest miRNA probabilities from the learned topics (S2 Table) were used mathematically as weights that were reapplied to the raw miRNA input values, thus biasing the spatial miRNA data with respect to the learned miRNA probabilities (Fig 3A). The primed data was input into the original case-study-wide LDA model (Fig 2C), and similar downstream visualization techniques and PCA were performed.

The specialized analysis stratified olaparib-sensitive tumors almost exclusively along PC1, with the let-7a-dominated topic 1 now having a greater allocation in tumors treated with olaparib alone, and the miR-21-dominated topic 2 having a greater allocation in tumors that received the combination therapies (Fig 3B). This pattern aligns with the improved survival among sensitive tumors receiving combination therapy (Fig 1B) and mirrors the trend seen in olaparib-resistant tumors in which combination treatment similarly shifted the spatial miRNA profile toward the miR-21-dominant topic (Fig 2D.iii).

While successful, this stratification still relied on our a priori knowledge that the tumors were sensitive and cannot be accurately described as fully prognostic, even though the final topic assignments were allocated by applying the same case-study-wide LDA model as in the naïve analysis to the newly primed miRNA data. What can be gleaned from this specialized analysis, however, is what it reveals about the nature of the spatial miRNA expression that was impacted by the priming and contributed to the successful stratification of the samples by treatment type. In other words, the value of this analysis is not so much that it provides a standalone prognostic classifier as that it reveals which spatial miRNA patterns distinguished olaparib-only from combination therapy within sensitive tumors.

Given that LDA models produce a set of miRNA probabilities that are referred to as “topics,” examining the change between the miRNA probabilities that constitute these topics across both the case-study-wide and the sensitive-only analyses can help explain the change in topic allocation. The highest-probability miRNAs across both LDA models are let-7a and miR-21 (S2 Table), which also correspond to the miRNAs that dominate the main topics responsible for stratifying the tumors. Our priming technique utilizes these probabilities as weights applied to the inputs, which results in what effectively amounts to a contrasting effect that is applied spatially across the miRNA data.

In the specialized LDA model, the weights for let-7a and miR-21 are 0.92 and 0.88, respectively (S2 Table), while the remaining miRNAs have lower weights. It is those miRNAs with the highest weights that are impacted relatively less than the rest when the LDA model seeks to allocate topics using this newly primed data, such that topics dominated by the reduced miRNAs will have relatively lower values than the topics dominated by heavily weighted miRNAs and can be overtaken by new topic allocations. With respect to our two dominant topics, we notice that, specifically in the olaparib-only treated samples, some of the formerly allocated topic 2 tumor regions, which were miR-21-dominated, are now being assigned to topic 1, which was let-7a-dominated, as can be seen in the spatial annotation maps in Fig 3B and S6 Fig. Given that the learned let-7a weighting was only ∼4% greater than that of miR-21, this means that, for the olaparib-only treated samples, the base amounts of miR-21 and let-7a across the parts of the tumor region that switched topics were already close in value, as opposed to those regions that maintained their original topic assignment. Thus, the ability to stratify the samples by treatment type was dependent primarily on examining the tumor regions where the miR-21 and let-7a amounts were most similar, which are the same regions that changed topic under priming.

Spatial associations between miRNAs and tumor cellular composition

Although spatial miRNA profiling stratified tumors by olaparib sensitivity and improved our understanding of how relative miRNA abundance related to treatment type, limitations in the current analytical approach remain. The spatial information gleaned from the miRNA data and represented by spatially allocated miRNA-informed topics (Fig 2D.ii) was effectively reduced to a summary statistic and further processed via PCA in order to answer a higher-level question: to what extent was a tumor likely to be sensitive or resistant to treatment. By stratifying the tumor samples along a PCA with known miRNA topic contributions (Fig 2D.iv), we developed a miRNA classifier that distinguished between sensitive and resistant states. To further characterize these states, we considered that the effectiveness of olaparib depends on the TME [22–24] and that Poly(I:C) is an immune stimulant[25]. We therefore related miRNA expression patterns to the immune composition of the TME.

We performed an automated H&E image analysis to classify cells as tumor or stromal. We then quantified tumor-infiltrating immune cells by staining for CD45 on the next serial tumor section cut immediately after the section used for in situ miRNA analysis. We extended our case-study-wide LDA analysis pipeline to include an SSIM assessment (Fig 4A)[27]. We used the top two miRNA topics dominated by let-7a and miR-21 and the normalized ratio of the spatial distributions of immune cells to tumor cells as input to generate SSIM maps, whereby the structural similarity of each spatial unit within the tumor is assessed by examining its local neighborhood and then calculating the local SSIM between the miRNA topics and cell ratio maps for that neighborhood (Fig 4A).

Fig 4.

To simplify the multiple per-miRNA-topic SSIM maps into a more interpretable form, we trained a 3-topic LDA model whereby each topic is informed by the miRNA-cell SSIM data. In this model, topic 1 is dominated by the let-7a-immune SSIM, and topic 2 is dominated by the miR-21-immune SSIM (Fig 4B.i). This allowed us to determine whether the immune cell infiltration in any given part of the tissue was more similar to the spatial distribution of one miRNA topic as opposed to the other. The 3-topic LDA model was applied to the tumor SSIM maps and the tumors’ topic percentages were calculated (Fig 4B.ii). A new variation of the spatial topic map was developed, where we added an intensity scaling to the topic maps to indicate not only which miRNA spatial distributions were more similar to the immune cell infiltration, but also the strength of that similarity via incorporating the per-pixel SSIM scores into the topic map (Fig 4B.iii, iv).

Most tumors, regardless of whether they were olaparib-sensitive or -resistant, had a greater spatial similarity between the immune cell infiltration and the let-7a miRNA topic than that of miR-21 (Fig 4B.ii). The exception to this observation was found among sensitive tumors that received the combination olaparib+Poly(I:C) treatment, where immune cell infiltration was more spatially similar to the miR-21-dominated topic. Given that Poly(I:C) modulates the immune response in tumors [38], increased immune cell activity was expected in tumors receiving olaparib+Poly(I:C). Further examination of the topic maps revealed a notable pattern in olaparib-sensitive tumors treated with the olaparib + Poly(I:C) combination: miR-21-immune regions appeared non-uniform, with brighter (higher-similarity) signal concentrated near boundaries with let-7a-immune regions and dimmer signal toward the center (e.g. SOP3 in Fig 4B.iii). Thus, incorporating SSIM into the LDA miRNA analysis pipeline enabled the assessment of spatial similarity between miRNAs and the underlying biology. Given the complexity of the spatial analysis, these maps also highlight specific regions that warrant examination in the future.

Discussion

This study investigated therapeutic strategies combining PARP inhibitors with PI3K inhibitors or the immune-stimulant Poly(I:C) to overcome drug resistance in BRCA1-related breast cancers. Both combinations delayed tumor progression, especially in olaparib-sensitive tumors, but that benefit was modest in tumors that had acquired resistance. To evaluate these treatments, we applied our spatial miRNA in situ profiling technology to FFPE tumors collected after 10 days of therapy and developed novel computational frameworks to analyze those early-stage tumors. Our LDA-based analyses demonstrated that early, spatially resolved miRNA profiles distinguished olaparib-sensitive from-resistant tumors, and revealed how combination therapies reshaped resistant tumors toward a functionally sensitive state. Two spatial miRNA topics, dominated by miR-21 and let-7a, stratified the tumors as olaparib-sensitive or -resistant. By integrating SSIM into the analysis pipeline, we bridged the gap between high-dimensional molecular signatures and the physical tumor cellular composition, enabling interpretation of the spatial biology and immune-regulatory landscapes underlying drug response. Our miRNA-based spatial analysis shows potential as a useful tool for predicting therapy response and biomarker discovery, and may provide a clinical window to adjust therapeutic strategies before tumor progression occurs.

While both PI3K inhibition and Poly(I:C) enhanced olaparib efficacy in sensitive tumors in vivo, only PI3K inhibition at least partially reversed acquired PARP inhibitor resistance. Our SSIM analysis of tumors after 10 days of treatment showed that, in resistant tumors, there was little difference in the spatial similarity between immune infiltration and the let-7a miRNA topic when comparing olaparib+Poly(I:C) with olaparib alone, indicating a failure of Poly(I:C) in modulating the gross composition of the TME. Direct inhibition of the PI3K/AKT pathway overcame cell-intrinsic survival mechanisms in resistant tumors. These findings highlight that spatial miRNA profiling can potentially identify tumors that are sensitive to PARP inhibitors and may benefit from additional immunomodulatory therapies, whereas tumors likely resistant to PARP inhibitors may benefit from the addition of other targeted agents, such as PI3K inhibitors.

We developed a case-study-wide LDA analysis framework that distinguished olaparib-sensitive and -resistant tumors. Among resistant tumors, our initial analysis was able to further distinguish olaparib-only treated tumors and those that had received the combination therapies. The lack of treatment-group separation among olaparib-sensitive tumors prompted us to implement a priming-based, specialized analysis to successfully resolve therapy-associated spatial miRNA profiles. This specialized approach highlighted how targeted priming of spatial miRNA data can reveal subtle stratification patterns that were missed by the case-study-wide LDA analysis.

To understand how priming influenced topic composition, we compared the miRNA probability distributions between the case-study-wide LDA topics and the sensitive-only models. Treatment-specific differences within olaparib-sensitive tumors were encoded in relatively subtle, locally balanced patterns between miR-21 and let-7a, rather than in large global shifts. In both models, the highest-probability miRNAs were let-7a and miR-21 (S2 Table), which was consistent with their dominance in the main topics responsible for stratifying the tumors. In the sensitive-only LDA model, the weights for let-7a and miR-21 were 0.92 and 0.88, respectively, while the other miRNAs had lower weights. As a result, those lower-weight miRNAs were reduced more strongly when priming was applied, and topics dominated by the reduced miRNAs lost relative influence during topic assignment compared to topics dominated by the two heavily weighted miRNAs. In olaparib-only treated samples, some of the formerly allocated topic 2 tumor regions (miR-21-dominated) were re-assigned to topic 1 (let-7a-dominated) (Fig 3B.i vs S5 Fig and S6 Fig). Given that the learned let-7a weighting was only ∼4% greater than that of miR-21, this indicated that, in the regions that switched topics, the base amounts of miR-21 and let-7a were already close in value, as opposed to regions that maintained their original topic assignment. This explained why priming was able to expose treatment-associated separation within the sensitive tumors even when global shifts appeared modest.

Finally, by integrating SSIM with miRNA topic maps and immune cell infiltration, we not only established a proof-of-concept framework for relating miRNA-defined spatial patterns to cellular architecture, but also found that, whereas most tumors showed greater spatial similarity between immune infiltration and the let-7a-dominated topic, olaparib+Poly(I:C)-treated sensitive tumors instead showed stronger immune-miR-21 spatial similarity. The SSIM-LDA maps showed that immune localization was differentially aligned with distinct miRNA patterns across treatment groups and highlighted specific micro regions where those patterns intersect, providing targets for more detailed pathological and mechanistic follow-up in the future.

The study design should also be considered when interpreting these findings. We focused on a single BRCA1-related TNBC mouse model and profiled tumors after 10 days of treatment to capture early spatial miRNA differences before endpoint-associated tissue changes became dominant. Additional treatment stages and BRCA1/2-related tumor models would help test whether the same spatial topics persist beyond this setting. The let-7a- and miR-21-dominated topics were associated with olaparib resistance and sensitivity, respectively, but these associations do not by themselves explain the biology underlying those spatial patterns. Future studies incorporating additional spatial biomarkers and histological analysis of the miRNA-defined regions could help clarify the cellular and molecular features associated with these topics.

Looking forward, these results suggest that spatial miRNA profiling could potentially be leveraged to assess whether a tumor is likely to be responsive to PARP inhibition. Applying our analytical framework to larger preclinical studies and, ultimately, to clinical BRCA1/2-mutant breast cancer cohorts would enable the evaluation of the robustness and generalizability of our identified spatial miRNA topics and stratifications, and could facilitate their translation from prognostic separations into formal predictive classification models. In parallel, SSIM-based spatial correlations could be extended beyond miRNA-only data to incorporate additional spatially resolved modalities, such as IHC markers and mRNA, and to explore alternative topic models or more complex spatial similarity measures within the same analytical framework, further enriching the biological and translational insight that can be extracted. In the context of PARP inhibitor-based therapy, integrating spatial miRNA readouts with survival outcomes and immune measurements may ultimately help refine patient selection and guide combination therapy choice.

Methods

Animal experiments and analysis

All animal experiments were performed in accordance with protocols approved by the Institutional Animal Care and Use Committee at Beth Israel Deaconess Medical Center (protocol number 052-2020-23). Breast tumors generated in K14-Cre Brca1^f/fTrp53^f/f female mice were dissected into small fragments and transplanted into the mammary fat pad of female FVB/NJ recipient mice aged ≥6 weeks. Olaparib-resistant tumors were obtained from mice that initially bore olaparib-sensitive tumors and were treated continuously with olaparib until resistance was acquired.[29] For treatment efficacy studies, mice were randomized into treatment groups once tumors reached ∼5 mm in diameter. Treatments continued until tumors reached 20 mm in their longest dimension, which was defined as the experimental endpoint; mice were euthanized at this point. For the 10-day studies, tumors were randomized into treatment groups once they reached 8 mm in size. After 10 days of treatment, tumors were harvested, fixed in formalin, and processed into FFPE sections. A vehicle-only control group was included for the 10-day olaparib-sensitive tumor cohort. A matched 10-day vehicle-only resistant tumor cohort was not included because, when untreated, olaparib-sensitive and olaparib-resistant tumors followed a similar endpoint trajectory, with tumors reaching endpoint within approximately 14-17 days.

Olaparib was reconstituted in DMSO, diluted in PBS immediately prior to intraperitoneal injection, and administered daily at 50 mg/kg. Alpelisib was administered orally at 24 mg/kg daily. Poly(I:C) was delivered intratumorally at 1 mg/kg, three times per week. Tumor dimensions were measured every 2-3 days using electronic calipers, and tumor volumes were calculated using the ellipsoid formula: Volume=L×W²/2, where L is the longest dimension and W is the shortest dimension. The log-rank test compared the survival curves between the study arms. Statistical analyses were performed using GraphPad Prism. Growth curves and Kaplan-Meier curves of survival were plotted. Survival between cohorts was compared using log-rank tests.

H&E and immunohistochemistry (IHC) staining, and automated image analyses

Tumors were harvested and fixed in Rapid-Fixx (Fisher Scientific, Hampton, NH), then submitted to the BIDMC histology core facility (RRID:SCR_009669) for paraffin embedding, sectioning, H&E staining, and performing CD45 IHC using their standard protocol. CD45 IHC was performed using a rat anti-mouse CD45 monoclonal antibody clone 30-F11 at 1:250 dilution (#14-0451-82, ThermoFisher Scientific, Waltham, MA) and a goat anti-rat IgG horseradish peroxidase polymer secondary antibody at 1:1 dilution (#ab214882, Abcam Inc, Waltham, MA). QuPath software was used to segment and classify cells as tumor or stromal on H&E images, and quantify CD45+ cells on IHC images.[39,40]

Spatially resolved miRNA assay

A spatially resolved miRNA assay is applied to tumor samples treated for 10 days to attain the spatial miRNA data. Assay device fabrication and the assay process were performed using methods as described in prior work.[14] The assay device consists of a nanoliter well array with hydrogel posts embedded within each well. The hydrogels were functionalized with DNA probes that had complementary sequences to the miRNAs in our panel (S1 Table). The relative positioning of the hydrogel posts within each well was designed such that each post in that position would target the same miRNA sequence. The miRNA assay with Rolling Circle Amplification (RCA) was applied to each sample as previously described, and the bound miRNA/DNA complexes were fluorescently labeled which allowed for quantitation.[14]

Spatial miRNA data pre-processing

The spatial miRNA data acquired from the assay required pre-processing prior to any subsequent analysis (LDA, SSIM, etc.). Utilizing microscope images of the samples during the assay (S7 Fig), the portion of the array that was overlapped by tissue, as opposed to the surrounding area outside the sample, was manually outlined and used as a tissue mask so that only miRNA values within the tissue region were retained. The bulk miRNA expression per tumor was compared between the groups using two-way ANOVA using GraphPad Prism. Šídák’s multiple comparisons test was used to compare between two groups, and comparisons that achieved adjusted p<0.05 were reported.

For LDA model training, the masked spatial miRNA data were processed further in order to ensure optimal performance. First, pixels outside the tissue mask were assigned a value of 0, and an alternate dataset was created in which these masked-out pixels were instead assigned a small positive constant (ε), to provide a safe variant of the data for any logarithmic analyses downstream in the pipeline. Training the LDA model in MATLAB (MathWorks, Natick, MA) required restructuring the miRNA data so that the preserved 2D spatial information was linearized by row for all samples and concatenated into a single array for LDA model training. Given this systematic linearization of the data, we were able to transform the results back into the original spatial configuration for visualization and subsequent analysis.

Case-study-wide spatial LDA implementation

Given the efficacy of the LDA training hyperparameters in prior work [14], an initial 4-topic LDA model was trained on the entirety of the miRNA data derived from all the 10-day treated tumors. Collapsed Gibbs Sampling was used as the inference algorithm, as it has been shown to yield more accurate topic discovery despite taking longer to train (training time is negligible, approximately ∼21 minutes).[41] As opposed to utilizing word frequencies when training the LDA model, we used scaled word counts. The quantified miRNA values per well were on the order of 3-4 attomoles at most and LDA requires integer word counts, so we scaled the values by multiplying every miRNA count by 10⁴ to preserve as much of the data as possible prior to conversion. Once the LDA model was trained, we used the topic word probabilities to visualize the contribution of each miRNA to each topic via word clouds.

Applying the trained LDA model on all samples spatially yielded the colored topic maps (S5 Fig). The relative abundance of each color-defined topic was then calculated per tumor section to obtain the topic-tumor percentages. Ultimately the LDA output was used to stratify the tumor samples by using the topic-tumor percentages as observations in a case-study-wide PCA, whereby the contribution of the topics (PCA variables) could be noted for each principal component to enable trend interpretation.

PCA statistical methods

To assess separation between treatment response groups (resistant versus sensitive), we compared the scores along the first principal component (PC1) for treated tumors that were ultimately classified as olaparib-sensitive versus olaparib-resistant (n=9 per group). PC1 scores between these two groups were compared using a two-sided Mann-Whitney U test. A p-value<0.05 was considered statistically significant.

Priming-based specialized LDA analysis for treatment-sensitive tumors

In addition to the case-study-wide LDA workflow, a data priming technique was applied to the sensitive tumors in order to bias the model to stratify them along treatment type, a use that is conceptually consistent with other priming methods in computer vision and other machine learning applications [42–45]. We initially examined all the miRNA probabilities associated with all the LDA topics, then for each miRNA we noted the highest probabilities found across all the topics (S2 Table). The miRNA probabilities were then stored in a vector and treated as a form of weights whose values were multiplied to all the spatial miRNA input values for our sensitive tumors. The primed input was then passed through the case-study-wide LDA model and the LDA topics were reassigned and used to stratify olaparib-sensitive samples by treatment type.

Tumor and immune cell data processing

Tumor and stromal cell count data were obtained from QuPath and by noting the position and orientation of the tumor sections as they were superimposed on our spatial miRNA quantification device while being assayed (S7 Fig). Using MATLAB, we quantified the amount of tumor, stroma, and immune cells that occupy the same position as the miRNAs. We defined immune cell infiltration as the normalized ratio of immune cells to tumor cells. First, we defined a tissue mask such that all locations containing immune or tumor cells are included; we then log-transformed the ratio at all locations (x,y) while adding a small positive constant ε and where r(x,y) represents the new ratio (Equation 1):

We then remapped the ratios from zero to one via a logistic transform such that values below 0.5 correspond to relatively immune-deficient regions and values above 0.5 correspond to relatively immune-enriched regions (Equation 2):

Pixels outside the tissue mask were excluded from downstream normalization. Next, we re-normalized the ratio on a per tumor basis such that the maximum ratio was set to one and the lowest to zero. Lastly, we clipped the values at the median (50^th percentile) within the tissue mask and renormalized the clipped values to [0,1]. The reasoning behind the renormalization is to make the final ratio compatible for use with SSIM. Since our second data type (miRNA topic maps) being input to the SSIM pipeline are effectively binarized masks and the spatial similarity between these masks and a ratio where most values are close to zero (given the large difference in immune cell vs tumor cell count), we needed to highlight the significance of the immune cell infiltration via remapping and signal normalization.

SSIM map formulation

SSIM maps were created in order to encapsulate the structural similarity between our spatial immune cell infiltration ratios and the case-study-wide LDA-derived miRNA topic maps. In essence, SSIM compares two images within a local neighborhood around each pixel by evaluating agreement in local mean intensity, local contrast (variance), and local structure (covariance). This yields a so-called structural similarity index that ranges from −1 to 1, where 1 indicates high local structural similarity, 0 indicates no detectable structural correlation, and −1 indicates strong local anti-correlation (i.e., the structures being compared are inverted relative to one another). We used the built-in SSIM function in MATLAB to formulate the maps where an approximately ∼7×7 pixel neighborhood was chosen given the total 28×28 tumor map size.

LDA-SSIM implementation

An LDA model was trained on the SSIM maps between the immune cell infiltration ratio maps and the top two miRNA topic maps (let-7a- and miR-21-dominated) in the same manner described for the case-study-wide LDA implementation. The 3 LDA topics learned consisted of probabilities of the two SSIM map types (let-7a-immune SSIM and miR-21-immune SSIM) and were applied spatially and statistically summarized into topic percentages within each tumor and represented via color. The spatial topic assignment was done similarly to the miRNA LDA model but with the added complexity of adding a color depth element to the maps such that the level of structural similarity can be gleaned in addition to which SSIM was most similar for any given region in the tumor.

Data availability

All data and code are publicly available and can be accessed via Zenodo repository: https://doi.org/10.5281/zenodo.18499891.

Acknowledgements

We gratefully acknowledge funding from the National Institutes of Health (NIH) R01CA235740 to PSD and FJS, R35CA232105 to FJS, R01CA226776 to GMW, R21CA267088 to YJH and GMW. GMW is also supported by the Breast Cancer Research Foundation 24-177. This work was supported in part by the Koch Institute Support (core) Grant P30-CA14051 from the National Cancer Institute. We thank the Koch Institute’s Robert A. Swanson (1969) Biotechnology Center for technical support, specifically the microscopy core. This work was in part supported by a graduate fellowship to ONM from the Ludwig Center at MIT’s Koch Institute for Integrative Cancer Research. This work was in part supported by the Takeda pharmaceutical graduate fellowship to ONM.

Additional files

Supporting Information

Additional information

Funding

HHS | National Institutes of Health (NIH) (R01CA235740)

• Frank J Slack

• Patrick Doyle

HHS | National Institutes of Health (NIH) (R35CA232105)

• Frank J Slack

HHS | National Institutes of Health (NIH) (R01CA226776)

• Gerburg Wulf

HHS | National Institutes of Health (NIH) (R21CA267088)

• Gerburg Wulf

• Jan Heng

Breast Cancer Research Foundation (BCRF) (24-177)

• Gerburg Wulf

Koch Institute for Integrative Cancer Research, Massachusetts Institute of Technology (KI) (P30-CA14051)

• Patrick Doyle