Introduction

Breast cancer is the most common tumor in women and its incidence is slowly increasing worldwide, in association with many factors related to socioeconomic conditions and lifestyle, but its death rate is decreasing thanks to better prevention, improved early detection and a broad range of therapeutic strategies1. Despite this success, breast cancer heterogeneity implies that some cancer cells may be resistant to chemotherapy, radiation therapy or immune therapy, becoming able to survive through the treatment1. Resistant cancer cells lead, therefore, to treatment failure and recurrence after treatment. Breast cancer subtypes exhibit differential recurrence risks with triple negative breast cancer (TNBC) displaying the highest recurrence rate, particularly in the first 5 years, partially due to the lack of an effective targeting strategy2.

The tumor microenvironment (TME) plays a crucial role in the mechanism of resistance: the cellular components of the TME engage in dynamic and extensive cross-talk based on both cell–cell interactions and paracrine signaling between each other and with the cancer cells3. Circulating cytokines are known to play a crucial role in TME with recent data showing that CD8+ T cell-secreted cytokines lead to local and global tissue modifications4. The role of TME for tumor development and the appearance of resistant tumors is a complex problem due to the interplay between cancer cells and infiltrating T-cells and other immune components5. Furthermore, in TME, there is another functional arm of the immune system, which is pro-angiogenic5. Inflammation can also be induced in response to various anti-cancer therapies, including chemotherapy, radiotherapy, and immunotherapy5.

Recent progress in immunobiology has led the way to successful host immunity enhancement against breast cancer. In TNBC, the combination of immunotherapy based on PD-1/PD-L1 immune checkpoint inhibitors with chemotherapy has been effective both in advanced and early setting clinical trials6. Initially, a multicenter non-randomized Phase Ib trial (KEYNOTE-012) showed an overall response rate of less than 20% in advanced TNBC patients expressing PD-L17. More recently, however, clinical trials (Impassion 130, KEYNOTE-355, KEYNOTE-522) showed prolonged progression-free survival and overall survival for patients with metastatic TNBC treated with PD-1 and PD-L1 inhibitor combined with chemotherapy with respect to chemotherapy alone8,9,10. A recent meta-analysis showed that addition of immunotherapy leads to an increase in pathological complete response (pCR) both for TNBC and for hormone receptor-positive, HER2-breast cancer11. In the case of TNBC, inclusion of immunotherapy increased the rates of pCR from 46.6% to 59.9% (OR = 1.64, 95% CI, 1.19–2.25) and similar relative increases were found for HER2-breast cancers, which also included TNBC11.

From a clinical standpoint, it would be desirable to identify in advance the patients who are most likely to benefit from the addition of immunotherapy to chemotherapy in the neoadjuvant setting12. Among potential biomarkers for immunotherapy in TNBC, only the expression of the PD-L1 protein has been validated in prospective clinical trials8,13, but pembrolizumab was effective in both patients with postive and negative PD-L1 expression in KEYNOTE-52210. Several genomic signatures for TNBC have been reported over the years in the literature14,15,16, but their clinical utility in the context of immunotherapy has not yet been fully established17.

It was recently shown in metastatic breast cancer patients treated with chemotherapy that circulating cytokines identified patients outcomes according to their longitudinal changes18. Furthermore, in patients treated with neoadjuvant chemotherapy along with moderate physical activity, changes in circulating cytokines were found in a cluster of patients more likely to achieve a pCR19. These studies suggest that circulating cytokines change during treatment and might represent an interesting target to screen patients.

In the past few years, our group has developed an algorithmic strategy, named ARIADNE, that is able to address the complexity and heterogeneity of TNBC and potentially of all solid tumors20. ARIADNE identifies hybrid epithelial-mesenchymal phenotypes by mapping gene expression data into the states of a Boolean network model of the epithelial-mesenchymal pathway, accounting for gene network effects20. ARIADNE has been validated on TNBC gene expression data obtained from core biopsies and provided prognostic information on patients’ disease-free survival20. ARIADNE stratifies patients for whom immunological signatures21 were not predictive of outcome, highlighting its relevance to predicting the response to immunotherapy22. The aim of the present paper was to investigate whether ARIADNE can predict the response to immunotherapy in TNBC and other HER2-negative breast cancer patients.

Methods

Gene expression data

We considered two patients cohort corresponding to the I-SPY 2 and NCT03366844 trials. The I-SPY 2 trial study is an adaptive clinical trial platform aimed at improving the outcome of high-risk breast cancer patients by testing different drugs in the neoadjuvant setting23. For this trial, we analyzed gene expression data for HER2-breast cancer patients treated with pembrolizumab in addition to chemotherapy (n = 69) and HER2 patients treated with chemotherapy (n = 179). HER2- patients included also TNBC patients. Data were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) under accession numbers GSE19404. Throughout the manuscript, we referred to this dataset as DATA1. The NCT03366844 trial combined pembrolizumab with radiotherapy (RT) in patients with operable TNBC24. We downloaded single-cell RNAseq data obtained from TNBC patients (n = 34) from GEO under accession number GSE246613. We referred to this dataset as DATA2 throughout the manuscript.

Data normalization and computation of ARIADNE scores for bulk gene expression data

We downloaded fully normalized and batch corrected data on the gene level for GSE194040 as provided in the GEO database without further normalization (n = 987 samples in total). The ARIADNE algorithm20 is based on the binarization of gene expression data in which highly expressed genes are coded as s = 1 and lowly expressed genes as s = −1, using a Boolean representation of the EMT regulatory network20. Since the ARIADNE was originally defined with respect to TNBC gene expression data20, we considered all the TNBC samples in the dataset (n = 364) and computed the ARIADNE scores for each sample α as \({E}_{\alpha }=-{\sum }_{i,j}{J}_{ij}{s}_{i}^{(\alpha )}{s}_{j}^{(\alpha )}\), where \({s}_{i}^{(\alpha )}\) is the binarized expression of the gene i in sample α and Ji,j = ± 1 for activating and repressive interactions, respectively20. We then defined ARIADNE groups by sorting and splitting the dataset into two risk groups (low and high) based on quantile discretization, with intervals [0, 0.75] and [0.75, 1], respectively. In other words, out of n = 364 TNBC samples, 75% (263) were assigned to low and 25% (91) were assigned to high. This process defined a score threshold that separated the two risk groups. We then computed the ARIADNE score20 and assigned risk group also for the other HER2- samples (not TNBC) present in the dataset. Out of the n = 69 HER2- samples receiving pemborlizumab (29 TNBC and 40 hormone-positive), 35 were classified as high and 34 as low.

Cytokines and calculation of the cytokine score

We considered a panel of 23 cytokines (interleukins, chemokines, interferon and growth factors) playing a critical role in innate and adaptive immunity18,19: IL-1β, IL-2, IL-12, IL-15, IL-18, IL-21, CXCL-10, TNF-α, CCL-2, CCL-4, IFN-γ, IL-4, IL-5, IL-10, IL-13, CCL-22; IL-6, IL-8, IL-7, TGF-β, VEGF, IL-17, and sIL-6R. We defined the cytokine score for each sample by taking the average of the expression level of the 23 cytokine genes described above, both for DATA1 and DATA2.

Immunological genomic signatures

We considered two immunological genomic signatures that have been proposed to screen TNBC patients before immunotherapy. The first signature, known as the IGG score, was based on the expression of 14 genes15, while the second, known as IO score, was based on the expression of 27 genes14. Here, we computed the IGG score using the mean expression of the 14 genes for the TNBC samples present in the I-SPY 2 dataset (n = 85). For the IO score, we used the values reported by Dugo et al.16 for the same TNBC samples.

Analysis of single-cell data

We downloaded single-cell gene expression data from GSE246613 (NCT03366844 trial). Among cells classified as non-immune, we selected those classified as “cancer cells" (n = 50185) to compute the ARIADNE landscape and scores as discussed in ref. 20. We also considered cells classified as cancer-associated fibroblasts (CAF) to compute the cytokine score as discussed above. Among immune cells, we selected those classified as T-cells, defining CD8+ T-cells and CD4+ T-cells as those expressing the CD8A and CD4 genes, respectively. Exhausted CD8+ cells are defined as those expressing also the PDC1, HAVCR2, LAG3, and CTLA4 genes. For CD4+ T-cells, we considered those expressing the IL2RA (CD25) and FOXP3 genes.

PD-L1 pathway analysis

We downloaded the genes related to the PD-L1 pathway (hsa05235), the KEGG pathway database (https://www.genome.jp/dbget-bin/www_bget?hsa05235), and selected the genes associated with the cancer cell component of the pathway. We computed the Pearson correlation coefficients among all the pairs of genes in the cancer cell component of the pathway, separately for patients belonging to the two ARIADNE groups (low/high), and then calculated the differences between the correlation coefficients of the two groups for each pair of genes. The network representation was obtained in Python using the networkX package https://networkx.org/.

Statistics and reproducibility

Statistically significant differences in pCR fractions among different groups was established using the Fisher exact test as implemented in the python scipy.stats.fisher_exact package. Statistically significant differences in scores were established with the Student’s t test as implemented in the scipy.stats.ttest_ind package. Multivariate logistic regression for ARIADNE and cytokine score was performed by first standardizing the variables, subtracting their mean and then dividing by their standard deviations, and the fitting using the statsmodels.api.Logit Python package. Differences between correlation coefficients were established using the Fisher transformation using the Python package CorrelationStats (https://github.com/psinger/CorrelationStats). Differences in the fraction of exhausted CD8 and CD4 T-cells were established according to the Kolmogorov–Smirnov test as implemented in the Python scipy.stats.ks_2sample package.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

ARIADNE stratified HER2- patients into risk groups with differential response to immunotherapy

Women with high-risk clinical stage II or III breast cancer were enrolled in the I-SPY2 clinical trial that was designed to improve outcomes by testing new drugs in the neoadjuvant setting, including immunotherapy23. Overall, I-SPY 2 evaluated new agents combined with standard therapy through pCR as the primary endpoint23. Since oncological drugs have a high clinical failure rate due to the heterogeneity of the cancers, we have exploited the capability of ARIADNE to be able to subdivide breast cancer patients into high or low risk groups to assess how each group responds to molecular targeted therapies. In the I-SPY2 trial, 987 patients were subdivided into 10 arms as follows: 210 patients received paclitaxel for 12 weeks followed by 4 cycles of anthracycline as standard therapy and were considered as reference for all the subsequent analysis (control arm). In the experimental arms, patients received for the first 12 weeks an additional drug in addition to the standard neoadjuvant regimen. In particular, we focused on the group of HER2- 69 patients who received pembrolizumab, a monoclonal antibody against PD-113. This group was composed of 29 patients classified as TNBC and 40 who were positive for PR or ER. The precise details of the cycle of treatments were reported in ref. 23.

As illustrated in Fig. 1a, RNASeq data were obtained from the biopsy that was acquired before the treatment (DATA1, as described in the methods section), while pCR was evaluated at the end of all the treatment. We computed the ARIADNE score from the RNASeq data obtained from HER2-negative patients and subdivided patients in two risk groups (low and high risk) according to the score, as detailed in the section “Methods”. Figure 1b displays the rates of pCR for immunotherapy and chemotherapy for the two risk groups. For patients who received immunotherapy in addition to chemotherapy, the high-risk group had a pCR rate of 26% as compared with 62% for the low-risk group with an odds ratio (OR) of 4.7, with 1.68–11.32 95% confidence interval (CI) and p-value p < 0.01 according to the Fisher exact test. Conversely, for patients who received chemotherapy alone, no significant differences were detected between the high and low risk groups (OR = 1.1 with 0.48–2.52 95% CI), in agreement with previous results25. We then investigated the differences in pCR between the immunotherapy and control arms separately for the different ARIADNE risk groups (Fig. 1c). In the original study, the immunotherapy arm reached a pCR rate of 44.9% versus 17.3% for chemotherapy alone arm (OR = 3.89 2.11–7.18, 95% CI, p < 10−3). When considering only the low risk group, the difference in pCR rate was considerably higher 62.8% versus 18.0% (OR = 7.82 3.4–17.97, 95% CI, p < 10−3), while for the high risk group, the difference in pCR rate was not statistically significant. Similar effect sizes were found restricting the analysis to TNBC patients, but the results were not significant possibly due to limited sample size (Fig. S1).

Fig. 1: ARIADNE predicts the response to immunotherapy.
figure 1

a Schematic of the I-SPY 2 clinical trial and the analysis performed here. b The rates of pCR for the two ARIADNE groups (low and high) for two arms of the trial. For the immunotherapy arm, the OR is 4.7 (1.68–11.32 95% CI) with significant differences in pCR rate (p < 0.01). No significant differences are found for the chemotherapy arm. c Differences in pCR between immunotherapy and chemotherapy for all patients (OR = 3.8), for the ARIADNE low group (OR = 7.8) and for the ARIADNE high group (OR = 1.84). Statistical significance is established with the Fisher exact test. Results based on DATA1, see “Methods”.

Full size image

ARIADNE and cytokine scores are associated with the response to immunotherapy

We constructed a cytokine score by averaging the expression value of 23 cytokines, as illustrated in the Methods section. Figure 2a illustrates that both the ARIADNE score and the cytokine score were statistically associated with pCR in patients receiving immunotherapy with chemotherapy: Patients displaying pCR had a lower ARIADNE score (low risk) and a higher cytokine score (p < 0.01). When we performed the same analysis on patients receiving chemotherapy alone, we found no significant association with ARIADNE score or cytokine score (Fig. 2b). This indicated that these two scores are associated specifically with the response to immunotherapy. Figure 2c summarizes all the results, illustrating that patients with high cytokine scores and low risk according to ARIADNE are more likely to respond to immunotherapy. The expression value of individual cytokines had no association with the ARIADNE class (Fig. S3). Considering the complex network between cytokines and other immune mediators, the analysis of individual cytokines might be misleading19.

Fig. 2: ARIADNE and cytokine scores are associated with the response to immunotherapy.
figure 2

a Ariadne score and cytokine score for patients with and without pCR to immunotherapy (p < 0.01). b ARIADNE score and cytokine score for patients with and without pCR to chemotherapy (p > 0.05). c The relation between ARIADNE and cytokine score shows that the highest pCR rates are found for ARIADNE low and cytokine high. Statistical significance is established with the Student’s t test. Results based on DATA1, see methods. In boxplots, the box is bounded by the first and third quartiles with whiskers representing 1.5 times the interquartile range.

Full size image

We next investigated if the cytokine score had a predictive value for the response to immunotherapy, independently of the ARIADNE score. To facilitate the comparison between the two scores, we performed a multivariate logistic regression considering standardized versions of the ARIADNE and cytokine scores, as described in the methods section. We found that both scores were significantly associated with pCR with similar effects: ARIADNE score had an OR of 0.5 (0.3–0.9 95% CI, p < 0.05) and the cytokine score an OR 2.0 of (CI 1.1–3.7 95% CI, p < 0.05) for each standardized score point (Table S1). As a further test, we computed the log-likelihood ratio Lr between a logistic regression with the ARIADNE score alone and when including also the cytokine score (Lr = 6, p < 0.05), indicating that the cytokine score improved the predicting abilities of ARIADNE.

ARIADNE score reveals regulatory disruption within the PD-L1 pathway

Our results showed that ARIADNE score was predictive of the response to immunotherapy in breast cancer patients treated with pembrolizumab, which interferes with the PD-L1/PD-1 pathway. By design, ARIADNE focuses on the EMT pathway and recognizes hybrid epithelial–mesenchymal cell phenotypes that are more aggressive and prone to metastasis. To better clarify the interactions between the EMT and PD-L-1 pathway, we analyzed the expression of the genes present in the PD-L1 pathway for the 69 HER2- breast cancer patients discussed above. We first stratified patients into two groups according to ARIADNE score and then compared the expression value for the two groups for all the genes present in the PD-L1 pathway. No significant differences were found for most of the genes (Fig. S2). When we analyzed instead correlations between the expression of the genes in the pathway, we found that the two groups were characterized by significantly different gene-gene correlation structure as summarized in Fig. 3. This indicated that the two ARIADNE groups displayed differences at the regulatory level within the PD-L1 pathway which might explain the differential response to immunotherapy of the two groups.

Fig. 3: ARIADNE groups display different regulation in the PD-L1 pathway.
figure 3

a Differences between gene expression correlations between the ARIADNE high and the ARIADNE low groups for the genes belonging to the PD-L1 pathway. b Network representation of the differences in correlation function, where only statistically significant changes in the correlation coefficient, according to the Fisher transformation, are reported. Results based on DATA1, see methods.

Full size image

We also checked possible relations between the ARIADNE score and two gene expression immunological signatures used in the context of the response to immunotherapy for TNBC14,15,16. We analyzed the correlation among the scores, focusing on 85 TNBC patients whose gene expression data were present in the I-SPY 2 dataset. As shown in Fig. S4, the ARIADNE score was not correlated with either the IGG score or the IO score, while the two scores appeared to be strongly correlated among each other (r = 0.62, p = 10−8). This result indicates that ARIADNE is distinct from the two immunological signatures. We also considered the cross-correlation between the expression values within each of these signatures and found strong associations between the genes comprising the signature (Fig. S4).

Mapping the heterogeneity of the response to immunotherapy through single-cell RNAseq data in TNBC patients

To corroborate the results obtained from bulk RNAseq extracted from biopsies of patients treated with immunotherapy, we also analyzed RNAseq data obtained from single cells in the NCT03366844 clinical trial for n = 50 TNBC patients (DATA2 as described in the methods section)24. As illustrated in Fig. 4a, biopsies were taken at baseline (T0), after one cycle of pembrolizumab (T1), and after a second cycle of pembrolizumab given in combination with radiotherapy (T2). We computed the ARIADNE score from scRNAseq data for cells previously classified as cancer cells24 and the results confirmed the association with pCR. As in the case of bulk RNAseq, patients that responded to immunotherapy had a lower ARIADNE score (Fig. 4b).

Fig. 4: Single-cell RNAseq data confirm the relations between ARIADNE/cytokine score and pCR.
figure 4

a Schematic of the NCT03366844 trial. b ARIADNE score measured in cancer cells is significantly associated with pCR (p < 10−4). The cytokine scores measured in (c) cancer cells and (d) CAF is significantly associated with pCR. e The relation between the ARIADNE, measured in CANCER cells, and the cytokine score, measured in CAF, shows that the highest pCR rates are found for ARIADNE low and cytokine high. Each dot corresponds to scores aggregated over each individual patient. f Temporal variation of the ARIADNE score for patients with or without pCR. Temporal variations of the cytokine score for (g) CAF and (h) cancer cells. Time variations in (f, g, h) are small but statistically significant (p < 10−4). Results based on DATA2, see methods. In boxplots, the box is bounded by the first and third quartiles with whiskers representing 1.5 times the interquartile range.

Full size image

We next considered the cytokine score in cells classified as cancer cells or CAF, specialized fibroblasts that are able to organize and modify both the cellular and the non-cellular components of the TME. Our results showed that patients responding to immunotherapy had lower cytokine scores for cancer cells (Fig. 4c), but higher scores for CAF (Fig. 4d), at T0. When we plotted the average ARIADNE score from cancer cells and average cytokine scores from CAF for each patient, we found similar patterns as observed for bulk RNAseq (Figs. 4e and S5). We then analyzed the time dependence of the ARIADNE and cytokine scores. The ARIADNE score obtained from cancer cells decreased in time for patients not responding to immunotherapy, while it remained roughly constant for responders (Fig. 4f). The cytokine score decreased for CAFs (Fig. 4g) and increased for cancer cells (Fig. 4h) after immunotherapy (T1). The time-dependent variations were small but always statistically significant.

The data also allowed us to study the time dependence of T-cells during the course of the therapy. We first identified CD8+ T-cells, considering all the single cells already classified as T-cells24 and expressing the CD8A gene. We then counted the fraction of these cells that also expressed a panel of additional genes (PDCD1, HAVCR2, LAG3, CTLA4) that are characteristic of exhausted T-cells. The results shown in Fig. 5a illustrate that the patients responding to immunotherapy displayed a much larger fraction of exhausted T-cells than non-responders and that this fraction was increasing during the course of the therapy. We also estimated the amount of terminally exhausted T-cells by considering the expression of the CD28 gene and found that about half of the exhausted T-cells could be considered as terminally exhausted. We then considered regulatory CD4+ T-cells, considering cells classified as T-cells and expressing the CD4A gene and examining the fraction of those also expressing the CD25 and FOX3P genes (Fig. 5b). The fraction CD4+CD25+FOX3P+ regulatory T-cells (T-reg) was larger for patients responding to immunotherapy than in non-responders. We also checked the association between the ARIADNE groups and the fractions of exhausted CD8+ and regulatory CD4+ T-cells, but the results were not statistically significant (Fig.  S6).

Fig. 5: Time dependence of T-cells during immunotherapy displays different patterns depending on pCR.
figure 5

a Fraction of exhausted CD8+ T-cells as a function of time and pCR status. The differences between respondent and non-respondent are statistically significant at T1 and T2, but not at T0, according to the Kolmogorov–Smirnov test (**: p < 0.01, *: p < 0.05) b Fraction of CD25+/FOX3P+ CD4+ T-cells as a function of time and pCR status. The differences between respondent and non-respondent are statistically significant at T0 and T2, but not at T1, according to the Kolmogorov–Smirnov test (**: p < 0.01). Results based on DATA2, see methods. In boxplots, the box is bounded by the first and third quartiles with whiskers representing 1.5 times the interquartile range.

Full size image

Discussion

In the present paper, we showed that the ARIADNE algorithm, which is based on the identification of epithelial-mesenchymal phenotypes in gene expression data, is able to stratify HER2- patients into high and low risk groups with differential response to immunotherapy. This result is of great interest for oncologists, since the ARIADNE algorithm can help reduce treatment failure and avoid overtreatment, if it can be confirmed for the TNBC subset for which immunotherapy is FDA-approved for treatment in the neoadjuvant setting. In the I-SPY 2 trial, patients that had a low ADRIADNE score before treatment displayed much higher pCR rate after treatment with chemotherapy plus immunotherapy than those with a high score (OR 7.82). This effect, however, was not seen in patients who received chemotherapy alone. The result suggests that ARIADNE score is able to capture some specific biological features that are relevant for immunotherapy. To validate this hypothesis, we analyzed the gene–gene interactions within the PD-L1/PD-1 pathway and found significant differences between ARIADNE groups. This result suggests that there is an association between ARIADNE targeted pathways and the regulatory interactions within the PD-L1/PD-1 pathway, which is targeted by immunotherapy. Notice, however, that a number of the genes involved in the PD-L1/PD-1 pathway are also part of the PI3K/AKT and MAPK pathways, important for cancer cell proliferation, so an independent explanation for differences in gene expression correlation could also be possible.

The presence of a complex network of cytokines in the TME could also affect the response of the tumor to the immune cells, but can also interfere with the capability of immune cells to mount an immunological response against tumor cells. To investigate this issue, we constructed a cytokine score and showed that it is statistically associated with pCR in patients receiving immunotherapy and chemotherapy, but not with patients receiving chemotherapy alone. Therefore, the analysis of ARIADNE on bioptic samples before treatment, combined with the analysis of the cytokine score, could provide a useful tool for oncologists to choose the best treatment for each patient, in view of a personalized approach. Further clinical studies will be necessary to understand if the cytokine score could add relevant predictive information with respect to ARIADNE alone.

Although the tumor-immune ecosystem is highly complex and comprises a heterogeneous collection of cells, single-cell technologies have emerged as powerful tools for the dissection of the TME, with recent application to TNBC patients treated with pembrolizumab24. To corroborate the results obtained from bulk RNA-Seq extracted from biopsies of patients treated with immunotherapy, we analyzed RNA-Seq of single cells from TNBC patients participating in the NCT03366844 clinical trial24. From this data, we computed the ARIADNE and cytokine score in cancer cells and CAFs, respectively, at baseline (T0), after a cycle of pembrolizumab (T1) and after a second cycle of pembrolizumab in combination with radiotherapy (T2). The analysis of single cells confirmed the results obtained from bulk transcriptomic data. These results implied that ARIADNE is able to capture the heterogeneity of the tumor and its complexity from bioptic samples obtained before the treatment.

Single-cell data also provided information on T cells. In patients displaying pCR, we found a higher fraction of exhausted CD8+ T (eCD8T) cells at baseline (T0) and this number increased significantly after treatment (T1 and T2). It is known that CD8+ T cells are the most important antitumor effector cells in the process of immunotherapy, and their number and functional status largely determine the antitumor effect26. Due to the continuous antigen exposure and the stimulation of inflammatory factors, these cells gradually decrease their proliferation and effector functions27. CD8+ T cell exhaustion is, however, not irreversible unless they have reached the terminally exhausted stage28. These cells have an exceptionally dynamic and complex process regulated by a combination of antigen exposure, the tumor metabolic microenvironment, the immune cells, cytokines, and other factors27. Exhausted T cells can, in fact, give rise to terminally eCD8T cells, the only irreversible stage of exhaustion, which display higher toxicity29. The high numbers of eCD8T cells at baseline in patients who reach pCR are in line with the fact that those tumors are inflamed and therefore will respond better to immunotherapy30. The increase in eCD8T at T1, which is much more pronounced in patients reaching pCR, seems to confirm that in these patients immunotherapy is able to reactivate the immune response, with more CD8 T cells reaching the tumor31. The turnover of these T cells increases, and thus also the proportion of exhausted and terminally exhausted T cells. Since the advent of immunotherapy, recent evidence has emerged associating high eCD8T levels at baseline and in the early part of treatment with a good prognosis. The reason for this association is that high eCD8T indicate that the tumor is inflamed and therefore responds better to immunotherapy than immune-excluded or deserted tumors32.

In conclusion, we demonstrated that the ARIADNE algorithm applied to transcriptomic data from biopsies successfully predicts the response to immunotherapy in HER2- breast cancer patients. We also showed that our approach is distinct from other immunological gene expression signatures and could instead be combined with the cytokine score.