Integrating Tumor-Intrinsic and Immunologic Factors to Identify Immunogenic Breast Cancers from a Low-Risk Cohort: Results from the Randomized SweBCG91RT Trial Open Access

This article is featured in Highlights of This Issue, p. 1643

Postoperative radiotherapy (RT) has long been a cornerstone for breast cancer treatment and significantly reduces the risk of an ipsilateral breast tumor recurrence (IBTR; ref. 1). Despite this, predictive biomarkers are still lacking (2, 3). The immune infiltrate in the primary tumor represents a biologically relevant variable as indicated by its prognostic (4), therapeutic (5), and, for anti-HER2 and chemotherapy, treatment-predictive effect (6–8). An improved understanding of the interaction between RT and the immune system can pave the way for new RT predictive tools.

Studies on the de-escalation of RT have mainly focused on clinically low-risk groups where the favorable prognosis limits the absolute benefit (9). In addition, trials using molecular profiling to de-escalate RT are ongoing (10, 11). We have previously shown that tumors that lack immune infiltration appear to have the most significant benefit from RT, but a less favorable prognosis (11, 12). Immunologic biomarkers have not yet been investigated as treatment-predictive biomarkers in prospective RT studies. However, they may provide additional independent information allowing for an improved RT individualization of these low-risk groups. Preclinical studies indicate that RT can induce immunogenic cell death, favoring the activation of the immune system (13). Consequently, there is a need to understand if the state of the antitumoral immune response can also predict the benefit from RT in a clinical context.

The primary mechanism behind immune cell-mediated tumor growth inhibition is the development of a specific immune response that leads to T-cell–mediated tumor killing (14). Mutated genes of tumor cells give rise to foreign proteins (so-called neoantigens), which can be taken up by antigen-presenting cells (APC) and presented on MHC molecules to T cells, which are activated provided adequate costimulatory signals are present (15). Consequently, a prerequisite for an antitumoral immune response is the availability of neoantigens. Post-activation modification can occur through the upregulation of inhibitory checkpoint molecules, activation-induced cell death, and T-cell exhaustion (14, 16, 17). Tumor cells can attenuate and inhibit the immune response during several stages by, for example, downregulating MHC molecules or inducing immunosuppressive signaling (18). The interaction between tumor-intrinsic and host factors thus determines the establishment and maintenance of an effective antitumoral immune response.

The immune system's role in breast cancer appears multifaceted and partly dependent on subtype (19), making it challenging to implement as a prognostic biomarker. Microsatellite instability (MSI) and tumor mutational burden (TMB) predict immune checkpoint inhibitor effect in several cancers (20–22), perhaps due to the correlation with neoantigen load, which in turn provides better conditions for a robust and polyclonal T-cell response. In breast cancer, subtype correlates with TMB (23), which may explain why estrogen receptor (ER) status has emerged as a predictor of benefit from immune checkpoint inhibition (24). Consequently, genomically stable ER-positive tumors may lack the conditions for activating and maintaining an effective immune response, which may explain the limited benefit from checkpoint blockade treatment and the lack of a favorable prognostic impact of an immune infiltrate (25–27). However, TMB alone may not be sufficient to identify breast tumors responsive to immunotherapy, highlighting the need to understand better the interplay between tumor-intrinsic and immunologic factors for tumor progression (28).

The study aimed to investigate how tumor aggressiveness and immune infiltration interact regarding the risk of IBTR and the benefit from RT. We hypothesized that an immune infiltrate is beneficial only among highly aggressive breast tumors—a characteristic we will henceforth refer to as immune responsiveness. We hypothesize systemic and local dissemination among such patients is delayed, consistent with a greater probability of being cured from surgery alone, which could allow for RT de-escalation. We used publicly available cohorts to train an immunologic model and a model designed to capture tumor-intrinsic characteristics associated with tumor aggressiveness. We tested our hypotheses in the randomized SweBCG91RT cohort and two additional well-annotated cohorts.

Publicly available breast cancer data sets were downloaded using the R package MetaGxBreast (ref. 29; Supplementary Table S1). For the TCGA cohort, updated follow-up data were retrieved using the R package curatedTCGAData (30). The cohorts with available outcome data (n = 21) were selected as training cohorts. Pam50 subtypes were inferred using the genefu package (31). The endpoints used were chosen in the following order based on availability: (i) Distant metastasis. (ii) Any recurrence. (iii) Overall survival. Age was included as a covariate for overall survival. The follow-up time was 10 years for distant metastasis and any recurrence and 15 years for overall survival.

We chose to use the rank-based single-sample Gene-Set Enrichment Analysis (ssGSEA; ref. 32) method to minimize batch effects, as rank-based methods show robustness across different array platforms (33). We used HGNChelper (34) to harmonize the names of genes across the different training cohorts. We hypothesized that primarily biological processes, rather than individual genes, drive prognosis and treatment prediction, so we used gene sets from msigdb (35) as features. Unlike individual genes, which can be involved in several different biological pathways and expressed in various cell types, we believe a gene set is more specific for a given underlying biological process. We deemed this important because our hypothesis testing required an accurate assessment of the quantity and quality of the local immune infiltrate. Finally, because enrichment scores are calculated on the basis of the available genes in each gene set, an additional advantage is that the analysis does not need to be limited to genes profiled in all 21 training cohorts.

The immunologic model was trained within Basal and Her2 tumors (n = 2,230) as we hypothesized that less aggressive subtypes exhibit a more heterogeneous immunologic prognostic signal. This assumption was based on previous literature, which indicates that the favorable prognostic effects of an immune infiltrate may be limited to aggressive subtypes (19). SsGSEA scores (32) were obtained using the msigdb (35) and GSVA packages (36) for gene sets from the C7 category (immunologic signature gene sets) and all additional gene sets, including any of the keywords; “LYMPHOCYTE|T_CELL|PD1|PD-1|PDL1|PD-L1|LAG3|CHECKPOINT_RECEPTOR|B_CELL|PERFORIN|GRANZYME|NK_CELL|CD8|CYTOTOXIC” (n = 5,661 gene sets).

The features were analyzed in Cox regression models, and the P values were saved as two-tailed (each P value was multiplied with the sign of the respective coefficients) and converted to one-sided P values using the metap package (37). A meta-analysis of the 21 P values for each feature was then performed using the weighted sum Z (Stouffer) method (38). The square roots of the number of observations in each cohort were used as weights (39). The top 50 ranked gene sets were selected, and Spearman correlations were computed for each possible pairwise combination of the 50 gene sets within each cohort. A mean value for each pairwise correlation was generated using all 21 cohorts. To reduce variable multicollinearity, we removed the gene set with the lowest P value for all gene sets with a mean pairwise correlation >0.7 or <(−)0.7, indicating a strong association (40). This resulted in 22 remaining gene sets (Supplementary Table S2), which were scaled and centered within each cohort, and the cohorts were merged. Finally, an elastic net model was fitted using the caret package (ref. 41; Supplementary Fig. S1). The resulting model was named Immunescore (Supplementary Fig. S1).

Because the model was trained against prognosis, we hypothesized that it measures the quantity and quality of the local immune response. However, because we did not use direct measurements of the immune infiltrate when developing the model, we validated the association by comparing it with IHC assessments of stromal tumor-infiltrating lymphocytes (TIL) on whole-tissue sections scored according to Salgado and colleagues (42) as described previously (43). Furthermore, we used xCell (44) and ESTIMATE (45), deconvolutional methods developed to quantify immune infiltrates from gene expression data, as controls for the correlation with TILs.

We then wanted to define and quantify tumor-intrinsic characteristics capable of predicting the biological effect of a preexisting immune infiltrate. We will henceforth refer to tumors where an immune infiltrate is associated with an improved prognosis as being immune-responsive. Gene sets from the H (hallmark gene sets), C2 (curated gene sets), and C6 (oncogenic signature gene sets) categories were chosen as potential features (n = 6,529). To identify relevant tumor-intrinsic biological processes, we performed Cox regression analyses, including all available tumors (n = 7,008), with an interaction term between each gene set and the previously created immunologic signature, Immunescore. A meta-analysis of interaction P values for all gene sets, selecting the 50 top-ranked gene sets and removing highly correlated processes [mean ρ > 0.7 or <(−)0.7], was performed using the same method as for Immunescore. A total of 12 gene sets were included in the final model (Supplementary Table S3). We hypothesized that tumor-intrinsic biological processes associated with tumor aggressiveness and immune-responsiveness are, in fact, two sides of the same coin. We based this assumption on previous literature indicating that immunologic biomarkers convey a favorable prognosis only among aggressive tumors (19). Consequently, to simultaneously isolate the proposed prognostic and immune-responsive signal of aggressive tumor-intrinsic pathways, we trained the model among tumors with an Immunescore in the lowest third (n = 2,312) to limit the attenuation of the prognostic signal of aggressive tumor characteristics by a coexisting immune infiltrate. The final model showed a clear dominance of proliferation-related processes, which is why we called it Proliferative Index (Supplementary Fig. S1; Supplementary Table S3).

To answer our clinical question of whether the integration of tumor-intrinsic and immunologic factors allows for downgrading high-risk tumors, we created a model that merges these dimensions. This was done by scaling, centering, and calculating each training cohort's Immunescore and Proliferative Index. The cohorts were then combined, and a new model was trained where the included variables were Immunescore, Proliferative Index, and an interaction term according to the expression: (Immunescore + Proliferative Index)² (Supplementary Table S4). Using tertiles, this model, integrating Immunescore and Proliferative Index, was then used to stratify patients ages <70 with histologic grade III tumors and patients ages <60 regardless of histologic grade. We chose this subgroup as it may be recommended RT boost based on today's guidelines (46). We hypothesized that this would allow us to identify a high-risk clinical group where RT boost de-escalation is feasible.

The well-annotated publicly available Servant (n = 341; ref. 47) and Sjöström (n = 172; ref. 48) cohorts include irradiated patients followed for IBTR (Supplementary Table S5). These cohorts were used to evaluate the implications among irradiated patients of Immunescore on the risk of IBTR based on Proliferative Index of the tumor.

To understand how aggressive tumor characteristics influence the RT benefit associated with an active immune infiltrate, we used the randomized SweBCG91RT cohort (49, 50). In summary, 1,178 patients with lymph-node negative (N0) stage I or IIA breast cancer were randomly assigned between 1991 and 1997 to breast-conserving surgery with or without whole-breast RT and followed for a median time of 15.2 years (Fig. 1; Table 1). GeneChip Human Exon 1.0 ST Arrays (Thermo Fisher Scientific) were used to obtain gene expression data (GEO GSE119295). Altogether, 7% of patients received endocrine treatment, 1% received chemotherapy, and 0.4% received both endocrine therapy and chemotherapy. Analyses were performed on treatment-naïve tumor samples. Invasive carcinoma was histologically confirmed by a board-certified pathologist. Clinical variables did not differ significantly between included and excluded patients except for tumor size and histologic grade. Excluded patients had slightly smaller (median size: 11 mm vs. 12 mm) tumors that were of a lower histologic grade compared with included patients (Supplementary Table S6).

Informed oral consent was obtained from all patients, which was determined appropriate and approved by the regional ethical review board for the original study and for this study (approval nos. 2010/127 and 2015/548). The study was conducted per the declaration of Helsinki.

Time to IBTR as the first event within 10 years from the date of diagnosis was used as the primary endpoint. Other recurrences and death were considered competing risks for IBTR. Multivariable regressions in the framework of flexible parametric survival analysis (Stata macro stpm2; ref. 51) were used to estimate Royston–Parmar models and HRs and to predict the 10-year cumulative incidence of IBTR. Time was measured from the date of randomization (SweBCG91RT cohort) or the date of diagnosis (Servant and Sjöström cohorts) to an IBTR event, censoring for competing events (SweBCG91RT cohort), death, or the last date of follow-up. Analyses were performed with up to 10 years of follow-up. The continuous values of Immunescore, Proliferative Index, and Integrated score were used for interaction analyses (52). The regression models included time-dependent effects for covariates that did not fulfill the proportional hazards assumption. Interactions were tested by comparing models with and without an interaction term using the likelihood ratio test. Covariates included in the SweBCG91RT cohort were age, ER status, histologic grade, and tumor size. Immunescore and Proliferative Index variables were standardized and rescaled with a mean of 0 and a standard deviation of 1. The predicted cumulative incidence for different percentiles of Immunescore and Proliferative Index from the regression models were plotted with 95% confidence intervals. A P value <0.05 was considered statistically significant. The full models are provided in the Supplementary Table S7. Figures of cumulative incidence were created according to the method of Fine and Gray (53) and based on the multivariable Cox models of subhazards. P values for differences between compared groups (denoted PCIF) were calculated by using a weighted log-rank test as described by Geskus (using the stcrprep command in Stata; ref. 54). R version 4.1.2 (1) and Stata/MP 17.0 for Mac were used for the statistical analysis (55).

Gene expression data have been deposited in the Gene Expression Omnibus under accession number GSE119295. However, due to regulations of the ethical review board and laws related to patient privacy, all clinical information has not been made publicly available. Questions and requests for additional data can be directed to the corresponding author.

Immunescore correlated with TILs (ρ 0.42, P < 0.001; Supplementary Fig. S2; Table 2). ESTIMATE and xCell showed similar but weaker correlations with TILs of 0.33 (P < 0.001) and 0.27 (P < 0.001), respectively. Furthermore, Immunescore correlated with histologic grade (ρ 0.25; P < 0.001) and was inversely correlated with ER status (ρ −0.26; P < 0.001) and age (ρ −0.076; P = 0.038; Table 2). Immunescore and Proliferative Index correlated (ρ 0.23; P < 0.001) and were generally enriched among tumors with prognostically unfavorable characteristics (Table 2). Proliferative Index showed similar correlations as Immunescore with TILs (ρ 0.41; P < 0.001), histologic grade (ρ 0.55; P < 0.001), and ER status (ρ −0.46; P < 0.001; Table 2).

Proliferative Index was associated with an unfavorable prognosis in unadjusted analysis in the Sjöström (HR, 1.98; HR, 1.49–2.62; P < 0.001), Servant (HR, 1.46; 95% CI, 1.20–1.78; P < 0.001), and SweBCG91RT (HR, 1.36; 95% CI, 1.14–1.63; P = 0.001) cohorts (Table 3). The significance remained in the Sjöström cohort with adjustment for subtype (HR, 1.61; 95% CI, 1.01–2.56; P = 0.044), in the SweBCG91RT cohort with adjustment for RT, ER status, histologic grade, age, and tumor size (HR, 1.32; 95% CI, 1.03–1.70; P = 0.031), but not in the Servant cohort with adjustment for age and subtype (HR, 1.06; 95% CI, 0.76–1.48; P = 0.716).

The prognostic effect of Immunescore varied depending on Proliferative Index, Figs. 2; S3. An increased Immunescore was associated with a less favorable prognosis among tumors with a low Proliferative Index. The opposite was seen for tumors with a high Proliferative Index, Figs. 2; S3, and Table 3. Tumors with a high Proliferative Index and a low Immunescore had the highest risk of IBTR. An unadjusted interaction test between Immunescore and Proliferative Index to IBTR was significant in the Sjöström (HR, 0.61; 95% CI, 0.47–0.81; P < 0.001), Servant (HR, 0.78; 95% CI, 0.65–0.95; P = 0.018), and SweBCG91RT (HR, 0.78; 95% CI, 0.64–0.95; P = 0.013) cohorts. The significance remained in the Sjöström cohort with adjustment for subtype (HR, 0.63; 95% CI, 0.47–0.84; P = 0.002), in the Servant cohort with adjustment for subtype and age (HR, 0.78; 95% CI, 0.63–0.97; P = 0.029), and the SweBCG91RT cohort with adjustment for RT, ER status, histologic grade, age, and tumor size (HR, 0.75; 95% CI, 0.61–0.94; P = 0.012).

To investigate how the benefit of RT varied depending on Immunescore and Proliferative Index, we then analyzed the SweBCG91RT cohort. Patients were stratified based on RT, and a model with an interaction term between Proliferative Index and Immunescore was created for irradiated (P_interaction = 0.013) and unirradiated (P_interaction = 0.17) patients (Supplementary Table S1). The prognosis of irradiated and unirradiated patients at different quantiles of Proliferative Index and Immunescore was then compared (Fig. 2). We hypothesized that highly aggressive tumors could be stratified based on immune infiltrates regarding RT benefit. We, therefore, studied tumors with a Proliferative Index in the highest quartile. In this group, a total of 19.2% (95% CI, 11.9–30.2) of unirradiated patients and 11.1% (95% CI, 5.1–23.0) of irradiated patients with an Immunescore above the median suffered an IBTR within 10 years. This could be contrasted against tumors with a Proliferative Index in the highest quartile but an Immunescore below the median, where 27.2% (95% CI, 14.6–47.2) of unirradiated and 30.4% (95% CI, 17.7–49.2) of irradiated patients suffered an IBTR. Among tumors with a low Proliferative Index, Immunescore was not associated with a favorable prognosis (Fig. 2). In summary, among aggressive tumors, an activated immune infiltrate was associated with a greater benefit from RT (Fig. 2).

We then used a model that integrates the Immunescore and Proliferative Index, which we called the Integrated model (Supplementary Table S4), to try to downgrade what, based on clinicopathologic variables, can be regarded as high-risk tumors where intensified treatment with RT boost and chemotherapy may be recommended (46). As shown in Supplementary Table S4, the model contains an interaction term between Immunescore and Proliferative Index. The coefficient of the interaction term is negative, indicating that among tumors with a high Proliferative Index, a high Immunescore can downgrade the predicted probability of an event. Therefore, the highest predicted risk is observed among tumors with a high Proliferative Index and low Immunescore. High-risk patients from the SweBCG91RT cohort were divided into tertiles based on the Integrated model. We defined high-risk patients as those aged <70 years with grade III tumors or aged <60 years with any histologic grade (46). The distribution of subtypes in the different tertiles can be seen in Supplementary Table S11. All subtypes were represented in the different tertiles, and Luminal B was the subtype most equally distributed. The benefit from RT appeared to be greatest in the group in the lowest tertile (HR, 0.28; 95% CI, 0.09–0.85; P = 0.025), followed by the median tertile (HR, 0.42; 95% CI, 0.16–1.08; P = 0.071), and smallest among patients in the highest tertile (HR, 0.59; 95% CI, 0.31–1.13; P = 0.111).

Patients from the lowest tertile had a 10-year cumulative incidence of IBTR of 17.1% (95% CI, 10.3–27.7) without RT and 5.4% (95% CI, 2.1–13.7) with RT (Fig. 3A). Patients within the medium tertile had a 10-year cumulative incidence of IBTR of 20.3% (95% CI, 12.8–31.4) unirradiated and 9.8% (95% CI, 4.5–20.5) irradiated (Fig. 3B). The group in the highest tertile was most likely to suffer an IBTR, with a 10-year cumulative incidence of IBTR of 28.7% (95% CI, 20.3–39.7) without RT and 19.5% (95% CI, 12.2–30.3) with RT (Fig. 3C).

Finally, to investigate if the Integrated model was predictive of RT benefit, we performed an interaction analysis in the whole SweBCG91RT cohort. The interaction was significant in an unadjusted analysis (P = 0.004) and with adjustment for histologic grade, age, ER status, and tumor size (P = 0.008; Supplementary Table S10; Supplementary Fig. S4).

In this study, we investigated if an integrated assessment of immunologic biomarkers and tumor-intrinsic factors in a low-risk cohort can identify tumors with aggressive characteristics that benefit from an immune infiltrate, enabling the downgrading of RT treatment—a group that research on de-intensification of RT has not focused on so far. Our findings suggest that this may be possible, paving the way for continued individualization of postoperative RT in breast cancer.

The key mechanism for immunologic tumor rejection is activating neoantigen-specific T cells. However, only a minority of tumor mutations give rise to such T-cell responses (56). Consequently, with a higher neoantigen load, genomically unstable tumors are associated with a higher probability of being infiltrated by an activated antitumor immune response and of benefit from immunotherapy in several cancers (14, 56). The signature designed to measure tumor-intrinsic qualities, which we called Proliferative Index, included gene sets capturing ER status and histologic grade, and correlated strongly with proliferation. On the basis of previous literature, these characteristics are associated with TMB (23) and, thus, likely the number of available immunogenic neoantigens. Therefore, an association between Proliferative Index and TMB may partially explain our findings. However, TMB may be relatively constant across breast cancer subtypes (57), indicating that other explanations for the link between tumor aggressiveness and tumor immunogenicity should also be explored.

Aggressive tumor characteristics may be associated with a more immunogenic tumor microenvironment. A high proliferation rate increases replication stress (58, 59). This results in the accumulation of genomic stress and a buildup of DNA in the cytosol, activating the cGAS–STING pathway, which is central to the activation of antigen-presenting cells (60). An association between tumor proliferation and a favorable tumor microenvironment for T-cell activation, not limited to TMB, is another plausible explanation for our findings. Proliferation has previously been hypothesized to determine the significance of the immune infiltrate (61), which is in line with our findings. Continued research on tumor–host interactions can further refine and individualize predictive immunologic biomarkers and pave the way for new individualized therapies, for example, vaccine treatments (62).

We have previously shown that an active immune response in the primary tumor appears to be associated with an improved prognosis (11) and a lower need for RT. This study shows that these findings primarily apply to patients in high-risk clinical groups. Despite highly aggressive characteristics, no RT boost, and a low frequency of systemic therapy, these patients have a 10-year IBTR risk of well below 10% with standard RT treatment provided they have an active immune infiltrate. In contrast, aggressive tumors without an immune infiltrate constitute the group with the highest risk of IBTR, and these patients may also derive the least benefit from RT. Previous studies have shown that these tumors are enriched for immunosuppressive mechanisms, and treatments targeting immunosuppressive signaling pathways and radiosensitization, such as TGFβ inhibition (63, 64), may be an appropriate strategy for eradicating residual tumor cells after surgery. Given that tumor aggressiveness and immune infiltration are predictive of immunotherapy and chemotherapy benefit (6–8), we hypothesize that the prognosis in clinical high-risk groups with and without an effective immune response diverges even more in the modern setting. This further strengthens the indication for individualized RT treatment based on immunologic biomarkers among patients with aggressive tumors. We believe other approximations of immune activation and tumor proliferation may provide equally useful information for treatment individualization. Therefore, future studies should investigate if integrating biomarkers already used in clinical practice, such as TILs, immune checkpoint molecule expression, ki67, and histologic grade, allows for an improved treatment individualization of high-risk tumors.

Several prospective studies have been conducted regarding the individualization of RT in breast cancer, but these have mainly focused on de-escalation in clinically low-risk groups, where the absolute benefit of RT is limited by the favorable prognosis (65). Low Proliferative Index scores most likely reflect this group, and we found no signs of a favorable immunologic effect here. Instead, tendencies towards the opposite were seen, in line with other studies that associate immune infiltrates with poorer prognosis in clinical low-risk groups (25–27). These indicate that prerequisites for immune activation may be lacking among slowly proliferating tumors. Furthermore, immune infiltration in low-risk groups may represent dysfunctional inflammation (66) or correlate with unfavorable tumor-intrinsic characteristics (67), warranting caution regarding the omission of RT among such patients. We hypothesize that the less aggressive subgroup of tumors, as represented by a low Proliferative Index in the present study, explains our previous findings of an increased RT benefit among immune-depleted tumors (11, 12).

The methods of this study were based on assumptions from previous literature, and the purpose was to analyze how tumor-intrinsic and immunologic factors interact regarding IBTR risk and RT benefit. Our study includes a large training cohort and three independent validation cohorts. The randomized setting of SweBCG91RT makes it a suitable cohort for investigating the effect of RT based on biomarkers. However, several limitations need to be acknowledged. The SweBCG91RT cohort is underpowered to demonstrate differences in RT effect along the two axes of tumor-intrinsic factors and immune activation. Several subgroup analyses were included in this study, and no adjustments were made for multiple hypothesis testing. Nevertheless, the consistent findings in the SweBCG91RT, Sjöström, and Servant cohorts are reassuring. Patients of the SweBCG91RT cohort largely lacked systemic treatment, indicating that the risk of relapse would likely have been lower in the modern setting. At the same time, immunotherapy and chemotherapy may have further accentuated the difference in IBTR rates between highly proliferative tumors with and without immune infiltration due to the treatment-predictive effects of an immune infiltrate (6–8). In addition, high-risk patients would likely have received an RT boost had they been diagnosed today and were, therefore, also undertreated with RT. This may limit the generalizability. At the same time, we believe this provides an opportunity to study which patients do well without systemic therapy and RT boost, and who may safely be omitted from intensified RT treatment. Finally, we did not base our immunologic model on direct measurements of an activated immune response. However, despite this, the correlation of our immunologic model with stromal TILs, measured on whole-tissue sections, was stronger than that of other established deconvolutional methods for estimating immune infiltration based on bulk RNA data. This indicates that our methods allowed for an accurate assessment of the local immune response.

RT can affect the tumor microenvironment differently depending on the dose (68). Preclinical studies suggest high RT doses can turn an immunosuppressive tumor microenvironment proinflammatory (69). In addition, combining RT with immunotherapy may further favor RT-induced immune activation, and clinical trials investigating this are ongoing (70). Finally, neoadjuvant RT, which entails a high tumor burden, can be hypothesized to elicit RT-induced cell death more effectively than adjuvant RT. It remains to be determined if alternative fractionation schemes, the addition of immunotherapy, or neoadjuvant RT is more effective in inducing an antitumoral immune response compared with adjuvant RT with standard fractionation used in this study.

In summary, we have shown that implementing the immune system as a biomarker for treatment individualization of RT in breast cancer requires consideration of tumor-intrinsic characteristics. Tumor-intrinsic features of aggressivity and immune responsiveness may be two sides of the same coin. Patients with aggressive tumors derive strong protection from an activated immune infiltrate regarding IBTR and may be candidates for RT de-escalation.

The grants were used to pay for salaries.

A. Stenmark Tullberg reports other support from Prelude Dx outside the submitted work; in addition, A. Stenmark Tullberg has patents pending. S.L. Chang reports patent application no. 63/154,821 (genomic classifiers for breast cancer radiotherapy) pending, a patent for PCT/IB2019/001181 (genomic classifiers for breast cancer radiotherapy) pending, and a patent for PCT/US2019/065098 (genomic classifiers for breast cancer adjuvant systemic therapy) pending; in addition, S.L. Chang was employed by Exact Sciences during the conduct of the study (until October 2022). F.Y. Feng reports personal fees from Bluestar Genomics, Astellas, Foundation Medicine, Exact Sciences, Tempus, POINT Biopharma, Janssen, Bayer, Myovant, Roivant, Bristol Meyers Squibb, and Novartis and other support from SerImmune and Artera outside the submitted work. L.J. Pierce reports unpaid consultancy for Exact Sciences. E. Holmberg reports other support from PFS Genomics during the conduct of the study; in addition, E. Holmberg has a patent for Prelude DX pending and a patent for Exact Sciences pending. P. Karlsson reports other support from PFS Genomics during the conduct of the study; in addition, P. Karlsson has a patent for Prelude DX pending and a patent for Exact Sciences pending, and has an advisory role with AstraZeneca. No disclosures were reported by the other authors.

A. Stenmark Tullberg: Conceptualization, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. M. Sjöström: Resources, supervision, methodology, writing–review and editing. E. Niméus: Resources, writing–review and editing. F. Killander: Resources, writing–review and editing. S.L. Chang: Resources, data curation, supervision. F.Y. Feng: Resources, supervision. C.W. Speers: Resources, supervision. L.J. Pierce: Resources, supervision. A. Kovács: Supervision. D. Lundstedt: Supervision. E. Holmberg: Formal analysis, supervision, funding acquisition, validation, investigation, visualization, writing–review and editing. P. Karlsson: Resources, supervision, funding acquisition, investigation, writing–review and editing.

This work was supported by the Swedish state under the agreement between the Swedish government and the county councils; ALF-agreement Grant No. ALFGBG-965020; the Swedish Cancer Society Grant No. Can- 21 1889-S; and the King Gustav V Jubilee Clinic Foundation Grant No. 2021–351

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.