Recent advances in medical treatments have dramatically improved the outcome of triple negative breast cancers. As illustration, after a median follow-up of 36 months, only 12% of the patients included in the adjuvant bevacizumab-containing therapy in triple-negative breast cancer (BEATRICE) trial had presented a metastatic relapse. This data emphasizes the need to develop predictors of outcome in a patient with triple negative breast cancer (TNBC) who have received optimal adjuvant therapy, in order to identify those who are eligible to adjuvant trials, and need new investigational drugs.

It has been previously shown that the presence of tumor infiltration by lymphocytes after neoadjuvant chemotherapy is associated with an excellent outcome. In this study that included 304 patients, the presence of TILs > 60% after neoadjuvant chemotherapy was observed in 10% of the patients and was associated with a 91% overall survival rate at 5 years.

Interestingly, 85% of the samples with post-chemotherapy TIL+ were TIL- at baseline before chemotherapy (Dieci MV, Criscitiello C, Goubar A, et al. (2014) Prognostic value of tumorinfiltrating lymphocytes on residual disease after primary chemotherapy for triplenegative breast cancer: a retrospective multicenter study. Ann Oncol Off J Eur Soc Med Oncol ESMO 25:611-618. doi: 10.1093/annonc/mdt556).

1. Purpose

The study purpose is to develop a genomic predictor of TIL after chemotherapy and to test its prognostic value in TNBC.

The strategy consists in developing a genomic predictor of TIL after neoadjuvant chemotherapy using only information obtained before the start of the neoadjuvant treatment (biopsies), and then to test whether this predictor could identify a subset of TNBC patients who do not have a systemic relapse. One of the aims is to develop a genomic predictor of TIL after neoadjuvant chemotherapy in TNBC using only information before the start of chemotherapy.

In order to address this question, we will quantify post-chemotherapy TIL in series of TNBC treated with neoadjuvant chemotherapy and for which a genomic profile has already been generated. TIL will be assessed in post-chemotherapy samples from MDACC neoadjuvant series and TOP (Trial of Principle) trial.

The histopathologic evaluation of the percentage of intratumoral (It) and stromal (Str) TILs will be performed on Hematoxilyn and eosin-stained (HES) slides from surgical specimens and will be done according to criteria previously described and published by Denkert and colleagues. For each case, all the slides containing residual invasive breast disease will be evaluated.

The goal will be to collect information on post-chemotherapy TIL in a large series patients with TNBC treated with neoadjuvant chemotherapy that did not achieved pCR after surgery. There is a lot of discussion on the most appropriate cut-off and in the absence of a reliable gold standard; we modeled the continuous level of stromal TILS in the post chemotherapy sample as a function of gene expression. This model is more powerful than logistic models and will allow us to predict which patients would have stromal TILS superior to currently discussed cutoffs (40%, 50% or 60%>). A RT-PCR based assay will then be developed on FFPE samples matched to their frozen counterparts.

The predictive value of the RT-PCR based assay for TIL-infiltration will be then validated on FFPE samples from IEO and GBG neoadjuvant studies.

Another aim is to validate the prognostic value of the genomic predictor in TNBC treated with neoadjuvant chemotherapy

Once the genomic predictor has been generated, we will test its prognostic value in patients with TNBC treated with adjuvant chemotherapy. Several series of samples will be used. First, the ACIS validation dataset will be used where both outcome and gene expression arrays are available. Second, we will perform gene expression profilings in the IBCSG study 22 and PACS08 in order to test the prognostic value of TIL-predictor in >300 TNBC treated with adjuvant therapy.

The primary analysis was performed on TNBC patients (ER-/HER2-). Description of all the studies included in the present analysis is shown in Table 31. Tumors were identified as ER-/HER2 - based on ER assessment by IHC and HER2 assessment by IHC and fluorescent in situ hybridization, as originally reported. When unavailable, ER and HER2 status was assigned according to ESR1 and ERBB2 gene expression.

2. Invention

The present invention relates to the use of the value of the expression of at least one gene selected from the group comprising: GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, for the estimation of prognosis of distant relapse-free survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT).

In a particular embodiment, the present invention relates to said use of the value of the expression of the four genes: GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, for the estimation of prognosis of distant relapse-free survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT).

In a particular embodiment, the present invention relates to said use of the value of the expression of the four genes : GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, wherein a low value of the expression of the genes SULTlEl and HLF, and a high value of the expression of the genes GBP1 and CXCL13, measured in a biopsy taken from a patient tumor before neoadjuvant chemotherapy corresponds to an high stromal tumor-infiltrating lymphocytes (Str-TIL) after neoadjuvant chemotherapy, corresponding to a good distant relapse free-survival or overall survival of said patient.

In a particular embodiment, the present invention relates to said use of the value of the expression of the four genes : GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, wherein a high value of the expression of the genes SULTlEl and HLF, and a low value of the expression of the genes GBP1 and CXCL13, measured in a biopsy taken from a patient tumor before neoadjuvant chemotherapy corresponds to an low stromal tumor-infiltrating lymphocytes (Str-TIL) after neoadjuvant chemotherapy, corresponding to a short distant relapse free-survival or overall survival of said patient. In a particular embodiment, the present invention relates to said use of the value of the expression of the four genes: GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene for determining a genomic predictor of formula:

Genomic predictor = 0.288 * GBP1 expression + 0.392 * CXCL13 expression -1.027 * HLF expression -1.726 * SULT1E1 expression,

and wherein the expression of the four genes corresponds respectively to the value of the mR A of each one,

for the estimation of prognosis of distant relapse-free survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT).

In a particular embodiment, the present invention relates to said use of the value of the expression of the four genes: GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene wherein when the genomic predictor for a patient is more than or equal to 0.51 , the patient has a good prognosis corresponding to a good distant relapse free-survival or overall survival of said patient.

In a particular embodiment, the present invention relates to sais use of the value of the expression of the four genes: GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene wherein when the genomic predictor for a patient is strictly less than 0.51, the patient has a poor prognosis corresponding to a short distant relapse free-survival or overall survival of said patient.

The present invention also relates to an in vitro prognostic method of the distant relapse-free survival or overall survival in a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT) comprising the determination of the value of the expression of at least one gene selected from the group comprising : GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse-free survival or overall survival in a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), comprising the determination of the value of the expression of the four following genes : GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene. In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT) wherein said gene expression is determined from mRNA or proteins, in particular from mRNA. In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein said gene expression is determined by a method allowing to measure mRNA quantity such as micro array, PCR or RT-PCR. In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein said gene expression is determined by an Affymetrix gene array.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein said value of the expression of the four following genes GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, is determined in a sample from a biopsy taken from a patient tumor before neoadjuvant chemotherapy. In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT) wherein the four gene corresponding toGBPl gene, HLF gene, CXCL13 gene and SULTlEl gene, are respectively represented by the nucleotide sequences SEQ ID NO: 1, SEQ ID NO: 2, SEQ ID NO: 3, and SEQ ID NO: 4.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein said value of the expression of the four genes : GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, corresponds to a low value of the expression of the genes SULTlEl and HLF, and a high value of the expression of the genes GBP1 and CXCL13, measured in a biopsy taken from a patient tumor before neoadjuvant chemotherapy corresponds to an high stromal tumor- infiltrating lymphocytes (Str-TIL) after neoadjuvant chemotherapy, corresponding to a good distant relapse free-survival or overall survival of said patient.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein said value of the expression of the four genes : GBP1 gene, HLF gene, CXCL13 gene and SULTlEl gene, corresponds to a high value of the expression of the genes SULTlEl and HLF, and a low value of the expression of the genes GBP1 and CXCL13, measured in a biopsy taken from a patient tumor before neoadjuvant chemotherapy corresponds to an low stromal tumor- infiltrating lymphocytes (Str-TIL) after neoadjuvant chemotherapy, corresponding to a short distant relapse free-survival or overall survival of said patient.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), comprising the determination of a genomic predictor according to formula:

Genomic predictor = 0.288 * GBP1 expression + 0.392 * CXCL13 expression -1.027 * HLF expression -1.726 * SULTlEl expression, for the estimation of prognosis of distant relapse-free survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT).

In a particular embodiment, the present invention relates to an in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein when the genomic predictor for a patient is strictly less than 0.51, the patient has a poor prognosis.

In a particular embodiment, the present invention relates to said in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT), wherein when the genomic predictor for a patient is more than or equal to 0.51 , the patient has a good prognosis. The present invention also relates to a kit for the in vitro prognostic method of the distant relapse survival or overall survival of a patient with triple negative breast cancer (TNBC) having received a neoadjuvant chemotherapy (NACT) according to claim , comprising :

- 4 pairs of primers corresponding to the 4 genes GBP1, HLF, CXCL13 and SULT1E1,

- at least one pair of primers corresponding to a housekeeping gene selected from the group comprising 18S rRNA, ACTB, HPRTl, HSPCB, PPIA, PUMl, RPS13, SDHA and TBP ,

- a reverse transcriptase,

- oligonucleotides,

- a polymerase

- and suitable buffer solutions.

The present invention also relates to an use of the value of the expression of the four genes : GBP1 gene, HLF gene, CXCL13 gene and SULT1E1 gene measured in a biopsy taken before a neoadjuvant chemotherapy (NACT), for predicting the level of stromal tumor-infiltrating lymphocytes (Str-TIL) in a patient with triple negative breast cancer (TNBC) after a NACT.

3. Training phase 3.1 Materiel 3.1.1 Description of the training population

The participants' flow chart of the training dataset is shown in Figures la and lb.

The baseline characteristics of the 99 eligible patients (ER-/HER2-) in the training dataset are presented in Table 1. The baseline characteristics of patients included in the training dataset are shown in Table 32 (n = 113). Table 1: Baseline characteristics of eligible patients in the training dataset

3.1.2 Genomic data

The complete genomic data are publically available on the Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) in the series GSE16446 for TOP samples; in the series GSE25066 and GSE20271 for MDACC samples. We performed data processing on the 113 patients with stromal TIL data available (99 patients were TNBC and 14 were HER2+, see Figures la and lb, GEO accessions of the 14 HER2+ patients are shown in Table 32). 3.1.2.1 Quality checks before normalization.

For quality checks before normalization, we used boxplots and plots of the density estimates of the raw probe level data comparing all arrays. Plots are shown in Figure 2 and Figure 3.

3.1.2.2 Separate data normalization using fRMA

We applied frozen robust multiarray analysis (fRMA) preprocessing algorithm to normalize the two datasets separately. This method is implemented in the R package 'frma'. For quality checks after fRMA, we used boxplots and plots of the density estimates of the normalized data comparing all arrays. Plots are shown in Figure 4 and Figure 5.

3.1.2.3 Cross-platform normalization

We merged the two datasets using Cross-platform normalization (XPN) methods for batch effect removal [3]. This method is implemented in the R package 'inSilicoMerging'. For quality checks after cross-platform normalization, we used boxplots and plots of the density estimates of the normalized data comparing all arrays. Plots are shown in Figure 6 and Figure 7.

3.1.2.4 Unspecified filtering

Unspecified filtering consists in including only the 10 000 most variable genes (standard deviation) for further analysis. It was performed once and for all, using gene expressions from 113 samples: the 10 000 genes selected will be used for all the further analysis.

3.2 Methods and results

3.2.1 Difference in stromal TIL after chemotherapy between MDACC samples and TOP samples TILs were quantified on RD after NACT in H&E slides from surgical samples from MDACC neoadjuvant series and TOP trial (training set). All mononuclear cells (i.e., lymphocytes and plasma cells) in the stromal compartment within the borders of the invasive tumor were evaluated and reported as a percentage (TILs score). TILs outside of the tumor border, around DCIS and normal breast tissue, as well as in areas of necrosis, if any, were not included in the scoring. TILs were assessed as a continuous measure (score). For each surgical specimen, all the slides containing invasive RD have been evaluated. The reproducibility of this method has been described 12. H&E slides from TOP samples have been sent to IEO, where they have been independently read for TIL-infiltration by two investigators (CC and GP). MDACC H&E slides have been read on-site by two investigators (CC and BS).

Difference in stromal TIL after chemotherapy between MDACC samples and TOP samples was assessed on the 113 patients in the training dataset. Stromal TIL significantly deviates from normality (Shapiro-Wilk normality test p-value = 9.771e-l l). There is a statistically significant difference in stromal TIL between MDACC samples and TOP samples (Wilcoxon rank sum test with continuity correction p-value = 0.005027). Summary statistics of stromal TIL in TOP samples, MDACC samples and overall are given in Table 2. Histograms of stromal TIL in TOP samples, MDACC samples and overall are shown in Figure 8.

Table 2: Summary statistics of stromal TIL in TOP samples, MDACC samples and overall

3.2.2 Box-Cox transformation

The Box-Cox transformation is a useful data transformation technique used to stabilize variance and make the data more normal distribution- like. Box-Cox transformation applies only to positive variables, so we applied it on (Stromal Til +1).

The univariate generalized linear model on which the Box-Cox transformation was applied included one at a time of the 10 000 most varying genes (see 3.1.2.4), center (Bordet vs. MDACC) and HER2 status (- vs. +). The model was applied on data from the 113 patients of the training dataset. The multivariate generalized linear model on which the Box-Cox transformation was applied on 113 patients from the training dataset and included one at a time of the 10 000 most varying genes and center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +), grade (1-2 vs. 3) and HER2 status (- vs. +). The model was applied on data from the 1 13 patients of the training dataset.

The Box-Cox transformation formula is given below:

Summary statistics of a values derived from 10 000 Box-Cox transformations are given in Table 3. We chose to set a at the median value for all the genes (10 000) in the multivariate analysis; consequently a = 0.2000 for all the following models.

Table 3: Summary statistics of a values derived from 10 000 Box-Cox transformations

3.2.3 Procedure 1: Univariate selection with adjustment

Procedure 1 steps:

1. To fit a general linear model to model the continuous level of stromal TIL in the post chemotherapy samples using complete cases. Stromal TIL is transformed using Box- Cox transformation.

2. To correct for multiple comparisons using False Discovery Rate (FDR) method

[Benjamini Y, Hochberg Y (1995) Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J R Stat Soc Ser B Methodol 57:289- 300.] (Bonferroni p- values are reported for information purposes only).

3. To report genes that achieved the selection criterion of a corrected p-value <0.05.

3.2.3.1 Univariate analysis 3.2.3.1.1 Triple negative patients

There were 99 patients identified as triple negative. We fitted a general linear model to model the continuous level of stromal TIL in the post chemotherapy sample as a function of gene expression while controlling for the effect of a potential confounder that is the center (Bordet vs. MDACC). Summary of the 79 genes achieving selection criterion (corrected p-value <0.05) are shown in Table 33.

3.2.3.1.2 All patients stratified on HER2 status

There were 113 patients used to build the model. We fitted a general linear model to model the continuous level of stromal TIL in the post chemotherapy sample as a function of gene expression while controlling for the effect of potential confounders that are center (Bordet vs. MDACC), and HER2 status (- vs. +). Summary of the 114 genes achieving selection criterion (corrected p-value <0.05) are shown in Table 34. 3.2.3.2 Multivariate analysis

3.2.3.2.1 Triple negative patients

There were 99 patients identified as triple negative. We fitted a general linear model to model the continuous level of stromal TIL in the post chemotherapy sample as a function of gene expression while controlling for the effect of potential confounders that are center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3). Summary of the 41 genes achieving selection criterion (corrected p-value <0.05) are shown in Table 35.

3.2.3.2.2 All patients stratified on HER2 status

There were 113 patients used to build the model. We fitted a general linear model to model the continuous level of stromal TIL in the post chemotherapy sample as a function of gene expression while controlling for the effect of a potential confounder that are center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3) and HER2 status (- vs. +). Summary of the 60 genes achieving selection criterion (corrected p-value <0.05) are shown in Table 36.

3.2.4 Procedure 2: Model selection using penalization

The purpose of the shrinkage is to prevent overfit arising due to either collinearity of the covariates or high-dimensionality. We chose to apply LI absolute value ("lasso") penalty as described by Tibshirani et al. [Tibshirani R (1996) Regression Shrinkage and Selection via the Lasso. J R Stat Soc Ser B Methodol 58:267-288] [Tibshirani R, others (1997) The lasso method for variable selection in the Cox model. Stat Med 16:385-395.].

Appling an LI penalty tends to results in many regression coefficients shrunk to zero and few other regression coefficients with comparatively little shrinkage hence this method allows selection of the most significant genes.

The amount of shrinkage is determined by the tuning parameter λ. A value of zero means no shrinkage, in this case, the method is identical to maximum likelihood estimation. A value of infinity means infinite shrinkage, in this case, all regression coefficients are set to zero. It is important to note that shrinkage methods are generally not invariant to the relative scaling of the covariates. We standardized the covariates before fitting the model. This standardization makes sure that each covariate is affected more or less equally by the penalization. Note that the regression coefficients reported here have been scaled back and correspond to the original scale of the covariates.

We included only the 10 000 most variable genes (standard deviation) in this analysis (see 3.1.2.4).

The appropriate generalized linear model for the response variable stromal TIL is linear. We penalized all the gene expressions covariates. Additional clinical covariates included are center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3). Those variables were not penalized. The penalization procedure was performed on 98 patients among the 99 eligible patients in the training dataset (one missing grade).

Stromal TIL is transformed using Box-Cox transformation.

3.2.4.1 The choice of tuning parameter λ

Model selection using penalization

The purpose of the shrinkage is to prevent overfit arising due to either collinearity of the covariates or high-dimensionality. We chose to apply LI absolute value ("lasso") penalty as described by Tibshirani et al. Appling an LI penalty tends to results in many regression coefficients shrunk to zero and few other regression coefficients with comparatively little shrinkage hence this method allows selection of the most significant genes. The amount of shrinkage is determined by the tuning parameter λ. A value of zero means no shrinkage, in this case, the method is identical to maximum likelihood estimation. A value of infinity means infinite shrinkage; in this case, all regression coefficients are set to zero (Figure 44). It is important to note that shrinkage methods are generally not invariant to the relative scaling of the covariates. We standardized the covariates before fitting the model. This standardization makes sure that each covariate is affected more or less equally by the penalization. Note that the regression coefficients reported have been scaled back and correspond to the original scale of the covariates. We included only the 10 000 most variable genes (standard deviation) in this analysis. Stromal TILs was transformed using Box-Cox transformation. We penalized all the gene expressions covariates. Additional clinicopathologic covariates included are series (TOP vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3). Those variables were not penalized. The penalization procedure was performed on 98 patients among the 99 eligible patients in the training dataset (one missing grade).

Cross-validation was used to assess the predictive ability of the model described above with different values of the tuning parameter. 10-fold cross-validation was chosen to determine the optimal value of the tuning parameter λ. The allocation of the subjects to the folds is random. When using LI optimization, the cross validated likelihood as a function of λ very often has several maxima hence it is important to cover a wide range of values (see Figure 9). The optimal value of λ was found equal to 91.5 (see Figure 10).

3.2.4.2 Genes selection

Penalization was performed with the optimal value of the tuning parameter λ. The clinical covariates: center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3) were included in the model but they were not penalized. The 4 selected genes are shown in Table 4.

Table 4: Genes selected using penalization

3.2.5 Genomic predictor of post-chemo TIL

3.2.5.1 Building the genomic predictor

After model selection and in order to determine the coefficients of the 4 selected genes in the construction of the genomic predictor, we applied a generalized linear model for the response variable stromal TIL on the 4 selected genes and the clinical covariates center (Bordet vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3). The genomic predictor is the linear combination of the genes expressions weighted by the regression coefficients shown in Table 5.

Stromal TIL is transformed using Box-Cox transformation.

Table 5: Genes associated with stromal TIL after chemotherapy

3.2.5.2 Description of the genomic predictor

The genomic predictor significantly deviates from normality (Shapiro-Wilk normality test pvalue = 1.518e-06). There was no statistically significant difference in the genomic predictor between MDACC samples and TOP samples (Wilcoxon rank sum test with continuity correction p-value = 0.888). Summary statistics of the genomic predictor for the 99 TNBC patients in the training dataset are given in Table 6. Histograms of the genomic predictor are shown in Figure 11. Table 6: Summary statistics of the genomic predictor in TOP samples, MDACC samples and overall

To facilitate interpretation of the values of the genomic predictor, we used a transformation to make the genomic predictor lie approximately between 0 (low value) and 1 (high value). The transformation has no effect on the prognostic value of the genomic predictor and is shown in the formula below, where i is the patient's index, Qo.os is the 5% quantile of the genomic predictor in the training samples (99 patients, Qo.os = -11.35669) and Q0.95 is 95% quantile of the genomic predictor in the training samples (99 patients, Q0.95 = -6.511546): Summary statistics of the transformed genomic predictor in the training dataset are given in Table 7. Histograms of the transformed genomic predictor are shown in Figure 12.

Table 7: Summary statistics of the transformed genomic predictor in TOP samples, MDACC samples and Overall

We used the transformed value of the genomic predictor within the rest of the training phase, referring to it as Genomic predictor. 3.2.5.3 Assessing the prognostic value of the genomic predictor on survival

The median follow-up (years) in the training dataset was computed using inverse Kaplan- Meier method applied on distant relapse-free survival (Table 8). There is a statistically significant difference in follow-up between the two cohorts (Logrank p-value = 1.68e-13).

Table 8:

3.2.5.3.1 Distant relapse- free survival

We assessed the prognostic value of the predictor on distant relapse-free survival (DRFS). In the training dataset, 94 patients had available data. We observed 43 events. Results of the Cox model are shown in Table 9. The Cox model is stratified on center.

Table 9: Multivariate cox model - Distant relapse-free survival

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between distant relapse-free survival and the genomic predictor. There was no significant non-linear effect (p = 0.2874). Log-relative hazard profiles are shown in Figure 13. 3.2.5.3.2 Overall survival

We assessed the prognostic value of the predictor on overall survival. In the training dataset, 94 patients had available data. We observed 41 events. Results of the Cox model are shown in Table 10. The Cox model is stratified on center.

Table 10: Multivariate cox model - Overall survival

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between overall survival and the genomic predictor. There was no significant nonlinear effect (p = 0.3057). Log-relative hazard profiles are shown in Figure 14.

3.2.5.4 Building risk groups

3.2.5.4.1 Cut-offs

We build risk groups based on:

1. Tertiles (33.33%, 66.66%>), referred to hereafter as TER

2. Median (50%), referred to hereafter as MED

3. Quantiles (27%, 73%) [Cox DR (1957) Note on Grouping. J Am Stat Assoc 52:543-

547. doi: 10.2307/2281704], referred to hereafter as COX

The cut-offs defined above are frozen for all the study.

3.2.5.4.2 Distant relapse-free survival

Kaplan-Meier distant relapse-free survival curves of the three risk groups according to the different cut-offs are shown in Figure 15, Figure 16 and Figure 17.

3.2.5.4.3 Overall survival

Kaplan-Meier overall survival curves of the three risk groups according to the different cutoffs are shown in Figure 18, Figure 19 and Figure 20.

3.2.5.5 Testing for correlations

3.2.5.5.1 Gene - Gene correlation

We performed pairwise correlation between the different genes included in the predictor using Spearman correlation. The correlation was assessed on 99 patients. Correlation coefficients values and 95% confidence intervals obtained using 1000 bootstrap repetitions are given in Table 11. Heat map shown in Figure 21 reflects hierarchic clustering of pairwise correlation between the 4 genes. The cells are colored according to Spearman's correlation coefficient values with red indicating positive correlations and green indicating negative correlations.

Table 11: Correlation coefficients and p-values of Spearman correlation

3.2.5.5.2 Correlation between the genomic predictor and validated gene modules (Immunel and Immune2)

Among 99 patients in the training dataset, only 85 had all genes expression to generate the genomic predictor and available immunel and immune2 gene modules expressions [9]. We performed pairwise correlation using Spearman correlation. Correlation coefficients values and 95% confidence intervals obtained using 1000 bootstrap repetitions are given in Table 12.

Table 12: Correlation between the genomic predictor and gene modules

3.2.5.5.3 Change in stromal TIL after chemotherapy as compared to before chemotherapy

From TOP samples, 36 patients had a GEO accession and available value of stromal TIL before chemotherapy (34 from the training dataset + 2 from the validation dataset). 29 of the 34 patients in the training dataset had both information about stromal TIL before chemotherapy and stromal TIL after chemotherapy. Spearman correlation coefficient value between stromal TIL before chemotherapy and stromal TIL after chemotherapy was 0.17 (p- value = 0.384). There is a significant absolute increase in stromal TIL after chemotherapy as compared to before chemotherapy (18.28, [CI95% 6.21 - 30.34], paired t-test p- value = 0.004). Individual profiles (Grey lines) and the mean profile (Dark grey line) are shown in Figure 22.

3.2.5.5.4 Correlation between the genomic predictor and stromal TIL before chemotherapy

From TOP samples, 22 had a GEO accession and available value of stromal TIL before chemotherapy. Spearman correlation coefficient value between stromal TIL before chemotherapy and the genomic predictor was 0.41 [-0.06 - 0.77]. 95% confidence intervals were obtained using 1000 bootstrap repetitions.

3.2.6 Prognostic value of stromal TIL after chemotherapy on survival

The Cox models are stratified on center. For illustrative purposes only, we show Kaplan- Meier survival curves, considering a cut-off value of 50%> for stromal TIL. 3.2.6.1 Distant relapse-free survival 3.2.6.1.1 Univariate analysis

In the training dataset, 95 patients had available data. We observed 44 events. (Table 13). Table 13 :

3.2.6.1.2 Multivariate analysis

In the training dataset, 94 patients had available data. We observed 43 events. Results of the Cox model are shown in Table 14.

Table 14: Multivariate Cox model - Stromal TIL on distant relapse-free survival

HR 95%IC P

Age 1.01 0.98 - 1.04 0.664 cT 0.312

TO- 1-2 1

T3-4 1.39 0.74 - 2.61

cN 0.816

NO 1

N+ 1.09 0.54 - 2.17

Grade 0.816

1-2 1

3 1.09 0.52 - 2.32

Stromal TIL after chemotherapy 0.98 0.96 - 1.00 0.043

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between distant relapse-free survival and the stromal TIL after chemotherapy. There was no significant non-linear effect (p = 0.501). Log-relative hazard profiles are shown in Figure 23.

3.2.6.2 Overall survival

3.2.6.2.1 Univariate analysis

In the training dataset, 95 patients had available data. We observed 42 events. (Table 15). Table 15:

HR 95%IC P

Stromal TIL after chemotherapy 0.98 0.96 - 1.00 0.027 3.2.6.2.2 Multivariate analysis

In the training dataset, 94 patients had available data. We observed 41 events. Results of the Cox model are shown in Table 16.

Table 16: Multivariate Cox model - Stromal TIL on overall survival

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between overall survival and stromal TIL after chemotherapy. There was no significant non-linear effect (p = 0.594). Log-relative hazard profiles are shown in Figure 25.

4 Validation phase 4.1 Materiel

4.1.1 Description of the validation population

The participants' flow chart of the validation dataset is shown in Figure 27.

In the validation dataset, 373 patients were TNBC (ER-, HER-). Among them, 185 had available survival data. The baseline characteristics of the patients in the validation dataset are presented Table 17.

Table 17: Baseline characteristics of patients in the validation dataset

4.1.2 Genomic data

The complete genomic data are available at the Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/). We applied frozen robust multiarray analysis (fRMA) [McCall MN, Bolstad BM, Irizarry RA (2010) Frozen robust multiarray analysis (fRMA). Biostat Oxf Engl 11 :242-253. doi: 10.1093/biostatistics/kxp059 ] preprocessing algorithm to normalize data separately on each series.

4.2 Methods and results

4.2.1 Description of the genomic predictor

The genomic Predictor significantly deviates from normality (Shapiro-Wilk normality test pvalue = 1.444e-08). There is a statistically significant difference in the genomic Predictor between the five cohorts' samples (Kruskal-Wallis rank sum test p-value < 2.2e-16). Summary statistics of the genomic predictor in the validation dataset are given in Table 18. Histograms of the genomic predictor are shown in Figure 28. TOP samples are different Affymetrix platform from all other samples (see Table 37).

Table 18: Summary statistics of the genomic predictor in the validation dataset

We performed the same transformation on the genomic predictor of the validation dataset as in the training dataset (see 3.2.5.2) using the 5% quantile of the genomic predictor in the training samples (99 patients, Qo.os = -11.35669) and the 95% quantile of the genomic predictor in the training samples (99 patients, Q0.95 = -6.511546). Summary statistics of the transformed genomic predictor in the training dataset are given in Table 19. Histograms of the transformed genomic predictor are shown in Figure 29. Table 19: Summary statistics of the transformed genomic predictor in the validation dataset

We used the transformed value of the genomic predictor within the rest of the validation phase, referring to it as Genomic predictor.

4.2.2 Validation of the prognostic value of the genomic predictor on distant relapse-free survival

The median follow-up (years) in the validation dataset was computed using inverse Kaplan- Meier method applied on distant relapse-free survival. There is no statistically significant difference in follow-up between the five cohorts (Logrank p-value = 0.556). (Table 20).

Table 20:

In the validation dataset, data were available only on distant relapse-free survival. 185 patients had available data. We observed 57 events. The Cox model is stratified on center.

4.2.2.1 Patients with no pCR (RD)

4.2.2.1.1 Univariate analysis

115 patients were not in pCR. We observed 49 events among them. (Table 21). Table 21:

4.2.2.1.2 Multivariate analysis

98 patients were not in pCR and had complete data. We observed 39 events among them. Results of the Cox model are shown in Table 22.

Table 22: Multivariate Cox model - Genomic Predictor on distant relapse-free survival - Validation dataset - Prognostic value of the four-gene signature on survival in a multivariate Cox model

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between distant relapse-free survival and the genomic predictor in the validation dataset for patients achieving pCR. There was no significant non-linear effect (p = 0.5240). Log-relative hazard profiles are shown in Figure 30.

4.2.2.2 All patients (pCR and RD)

4.2.2.2.1 Univariate analysis

185 patients had available data. We observed 57 events among them.(Table 23). Table 23:

4.2.2.2.2 Multivariate analysis

160 patients had complete data. We observed 45 events among them. Results of the Cox model are shown in Table 24.

Table 24: Multivariate Cox model - Genomic Predictor on distant relapse-free survival - Validation Dataset - Prognostic value of the four-gene signature on survival in a multivariate Cox model

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between distant relapse-free survival and the genomic predictor in the validation dataset. There was no significant non-linear effect (p = 0.4504). Log-relative hazard profiles are shown in Figure 31.

4.2.3 Validation of risk groups

We used cut-off points assessed on the training dataset for building risk groups in the validation dataset (TER, MED, COX).

4.2.3.1 Patients with no pCR (RD)

Kaplan-Meier distant relapse-free survival curves of the three risk groups according to the different cut-offs and for patients that did not achieved pCR are shown in Figure 32, Figure 33 and Figure 34.

4.2.3.2 All patients (pCR and RD)

Kaplan-Meier distant relapse-free survival curves of the three risk groups according to the different cut-offs and for all patients are shown in Figure 35, Figure 36 and Figure 37.

4.2.4 Testing for correlation 4.2.4.1 Gene - Gene correlation

We performed pairwise correlation between the different genes included in the predictor using Spearman correlation. The correlation was assessed on 185 patients. Correlation coefficients values and 95% confidence intervals obtained using 1000 bootstrap repetitions are given in Table 25. Heat map shown in Figure 38 reflects hierarchic clustering of pairwise correlation between the 4 genes. The cells are colored according to Spearman's correlation coefficient values with red indicating positive correlations and green indicating negative correlations.

Table 25: Correlation coefficients and p-values of Spearman correlation - Validation dataset

4.2.4.2 Correlation between our predictor and validated gene modules (Immunel and Immune2)

All patients (n = 185) have expressions of the genomic predictor and available immunel and immune2 gene modules expressions. We performed pairwise correlation using Spearman correlation. Correlation coefficients values and 95% confidence intervals obtained using 1000 bootstrap repetitions are given in Table 26.

Table 26: Correlation between the genomic predictor and gene modules - Validation dataset

4.2.6 Validation of the prognostic value in the training and in the validation set at diagnosis

4.2.6.1 Study population

Study flowchart for the training set is described in Figure lb. Overall, 99 patients with ER- /HER2- BC were selected to generate the signature. Patients' characteristics in the training set are given in Table 1. Flowchart for the validation set is described in supplementary material. Overall, 185 patients with ER-/HER2- BC were selected to validate the prognostic value of the signature on DRFS. Patients' characteristics in the validation set are given in Table 17.

4.2.6.2 Prognostic value of the four-gene signature in the training set

The prognostic value of the four-gene signature was assessed in 94 patients from the training set, for whom survival data were available. All patients had RD after NACT. Median (Ql - Q3) follow-up was 7.6 years (3.7 - 8.8). In a multivariate analysis (Table 42), the four-gene signature was significantly associated with better DRFS (HR for a one-unit increase in the value of the 4-gene signature: 0.28, 95%CI: 0.13-0.63, ρ=0·002). Kaplan-Meier DRFS curves of the risk groups (low four-gene signature vs. high four-gene signature) constructed using the median value of the 4-gene signature (median=0.51) are shown in Figure 16. There was no evidence of a non- linear association between the 4-gene signature and DRFS. The 4-gene signature added significant prognostic information to the clinicopathological characteristics at diagnosis, as shown by the likelihood ratio test (ρ=0·004). The discrimination was also improved; at five years, the C-index increased from 0.617 to 0.673 (Table 42). Similar results were obtained for OS (HR for a one-unit increase in the value of the 4-gene signature: 0.35, 95%CI: 0.16-0.75, ρ=0·007; likelihood ratio test, ρ=0·012; the C-index increased from 0.631 to 0.668).

4.2.6.3 Prognostic value of the four-gene signature in the validation set

In the validation set, 68 (37%) patients achieved pCR and 115 (63%>) relapsed (2 missing information on pCR). The prognostic value of the four-gene signature was assessed in 162 patients (23 missing information on grade). Median (Ql - Q3) follow-up was 3.2 years (2.3 - 4.5). In a multivariate analysis (Table 42), the four-gene signature was significantly associated with better DRFS (HR for a one-unit increase in the value of the 4-gene signature: 0.29, 95%>CI: 0.13-0.67, p=0.004). Kaplan-Meier DRFS curves of the risk groups constructed using the same cutoff (0.51) as in the training set are shown in Figure 36. There was no strong evidence of a non- linear association between the 4-gene signature and DRFS. The 4-gene signature added prognostic information to the clinicopathologic model at diagnosis as shown by the likelihood ratio test (p=0.008). Discrimination was also improved; at five years, the C- index increased from 0.686 to 0.700 in the validation set.

Results of the conditional logistic model showed no statistically significant association between the four-gene signature and the probability to achieve pCR in the validation set (OR for a one-unit increase in the four-gene signature: 0.96, 95% CI: 0.30-3.08, ρ=0·947, detailed results are provided in the supplementary material.

Table 42: Prognostic value of the four-gene signature on survival in a multivariate Cox model

DRFS, distant relapse-free survival; OS, overall survival; cT, clinical tumor size; cN, clinical nodal status; HR, Hazard ratio; CI, confidence interval; P, p-value

5. Distribution of the genomic predictor: training vs. validation

Samples included 99 patients from the training dataset and 185 patients from the validation dataset. There was a statistically significant difference in the genomic predictor between the training dataset and the validation dataset (Wilcoxon rank sum test with continuity correction p-value = 0.001349). Summary statistics of the genomic predictor are given in Table 27. Histograms of the genomic predictor are shown in Figure 39.

Table 27: Summary statistics of the genomic predictor - Training vs. validation

Summary statistics of the standardized genomic predictor are given in Table 28. Histograms of the genomic predictor are shown in Figure 40.

Table 28: Summary statistics of the transformed genomic predictor - Training vs. validation

6. Evaluating the added value of the genomic predictor to a clinical model

We used Uno's C-statistic to quantify the capacity of the prediction models in discriminating among subjects with different event times [10]. We considered two truncation times: 3 years and 5 years. The resulting Cs tell how well the given prediction models work in predicting events that occur in the time range from 0 to 3 years and 0 to 5 years, respectively. The clinical models (CM) included data in Table 29.

Table 29:

We used the likelihood ratio statistics in Cox regression models stratified on center to estimate the added value of the genomic predictor to the previously defined clinical models. We gave p-values of the likelihood ratio test. Results of the assessment of added value of the genomic predictor are shown in Table 30a and b. 95% confidence intervals were obtained using 1000 bootstrap repetitions.

Uno's concordance indices were computed to quantify the capacity of the prediction models in discriminating among subjects with different event times. Two truncation times were considered: 3 years and 5 years. The concordance indices indicate how well the given prediction models work in predicting events that occur in the time range from 0 to 3 years and 0 to 5 years, respectively. The likelihood ratio statistics was used in Cox regression models stratified on series to estimate the added value of the 4-gene signature to the clinical models. 95% confidence intervals were obtained using 1000 bootstrap repetitions. CM, clinical model; C-index, Concordance index; p, p-value; DRFS, distant relapse-free survival; OS, overall survival

Table 31: Summary information about the neoadjuvant studies included in the present analysis.

Table 34: Summary of genes achieving selection criterion (corrected p-value <0.05) in univariate analysis of triple negative patients

Table 35: Summary of genes achieving selection criterion (corrected p-value <0.05) in univariate analysis of all patients stratified on HER status

Table 36: Summary of genes achieving selection criterion (corrected p-value <0.05) in multivariate analysis of triple negative patients

Table 37: Summary of genes achieving selection criterion (corrected p-value <0.05) in multivariate analysis of all patients stratified on HER status

Table 38: One-to-one mapping from gene to 'best' probe sets using 'jetset' package

Association between the four-gene signature and stromal TILs

To assess the prognostic value of the four-gene signature on stromal TILs (Box-cox- transformed), we applied a general linear model for the response variable stromal TIL on the four-gene signature and the clinical covariates series (TOP vs. MDACC), age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3). Results of the general linear model are shown in Table 39. Table 39: General linear model with nonlinear effects - (Box-cox-transformed) Stromal TILs - 4-gene signature

We used restricted cubic splines with 2 degrees of freedom to investigate the non-linear association between stromal TILs and the 4-gene signature. The non-linear effect was found significant. Plot of fitted stromal TILs (Box-cox-transformed) vs. observed stromal TILs (Box-cox-transformed) is shown in Figure 45.

We computed the root mean squared prediction error (RMSE) using 1 000 repetitions of a ten- fold cross validation in the following way; the training dataset is first randomly split into ten previously obtained blocks of approximately equal size. Each of the ten data blocks is left out once to fit the model, and predictions are computed for the observations in the left-out block with the predict method of the fitted model. Thus, a prediction is obtained for each observation. The observed stromal TILs value and the obtained predictions for all observations are then passed to the prediction loss function cost (RMSE) to estimate the prediction error. This process is replicated 1 000 times and the estimated prediction errors from all replications as well as their average are estimated. Assessing the association between the four-gene signature and pathological complete response in the validation set

We explored the association between the probability to achieve pathological complete response (pCR) and the four-gene signature in the validation data set, we computed odds ratios (ORs) and 95% CI using a conditional logistic model that included the four-gene signature and the clinical covariates: age (continuous), cT (0-1-2 vs. 3-4), cN (0 vs. +) and grade (1-2 vs. 3) and was stratified on series (TOP vs. MDACC). Results of the conditional logistic model are shown in Table 40.

Table 40: Results of the conditional logistic regression assessing the association between the probability to achieve pathological complete response and the four-gene signature

Univariate selection (including one gene at a time) with correction for multiple comparisons (secondary analysis)

The univariate selection with correction for multiple comparisons procedure includes three steps:

1. To fit a general linear model to model the continuous level of Stromal TILs in the post chemotherapy samples using complete cases. Stromal TILs is transformed using Box- Cox transformation.

2. To correct for multiple comparisons using False Discovery Rate (FDR) method

(Bonferroni p-values are reported for information purposes only).

3. To report genes that achieved the selection criterion of a corrected p-value < 0 05.

Table 41: Summary of univariate selection with correction for multiple comparisons

EXAMPLE

The starting biological material is a sample from patient having a TNBC, such as as tumor biopsy, fine needle aspiration or blood sample.

Said sample is taken before any treatment. mRNA are extracted from said sample by well-known technics by a person skilled in the art.

These mRNA are used to quantify the expression of the 4 genes GBPl, HLF, CXCL13 and SULT1E1 by a RT-PCR technic or similar technics, using 4 pairs of primers corresponding to the 4 genes of interest.

At least one housekeeping gene selected from the group comprising 18S rRNA, ACTB, HPRT1, HSPCB, PPIA, PUM1, RPS13, SDHA and TBP, is used to performed RT-PCR.

The measured expressions of the 4 genes GBPl, HLF, CXCL13 and SULT1E1 are then incorporated in the following equation in order to obtain the genomic predictor:

Genomic predictor = 0.288 * GBPl expression + 0.392 * CXCL13 expression -1.027 * HLF expression -1.726 * SULT1E1 expression Coefficients applied to each of the gene expressions have been determined according to Table 5.

A distant relapse free and overall survival probability is calculated based on an equation that integrates the expression measurements of the 4 genes through the genomic predictor and the patient clinicopathological characteristics such as age, tumour size, tumour grade and tumour stage.

If the predicted survival probabilities are deemed high enough by the treating physician, the patient will receive a NACT.

If the predicted survival probabilities are deemed too low enough by the treating physician, the patient will receive more aggressive treatments (that can either by new experimental treatments in clinical trials or established therapy regimens for early breast cancer). Another aspect of the invention is the study of HLF (Hepatic Leukemia Factor) gene.

As previously shown by our unit, treatment with chemotherapeutic agents induced an antitumor immune response in TNBC and this high infiltration with TILs was connected to favourable outcome (Dieci et al, 2014). By large scale study, the prognostic role of TILs in early TNBC patients was confirmed, since the ten-year overall survival rates were 89% and 68%o for TNBC with high TILs and low TILs, respectively (Dieci et al, 2015). Another study, performed on primary TNBC patients of international FinHER trial, showed high TIL levels at a time of diagnosis associated with decreased distant recurrence rates (Loi et al, 2014).

In our group, in order to develop a genomic predictor of TILs after neoadjuvant ChT and to validate the possible prognostic value of this tool, post-ChT levels of TILs were quantified in series of TNBC patients that did not achieve pathological complete remission after surgery, and for which a genomic profile was already available. For the analysis, TILs have been evaluated after ChT in 113 samples from TNBC patients; 44 samples from TOP trial of Institut Jules Bordet (Brussels, Belgium) and 69 samples from MD Anderson Cancer Center (Houston, TX, USA) series. Our biostaticians proceeded to model the continuous level of stromal TILs in the post-ChT samples as a function of gene expression. Analyses led to the selection of four genes sharing a triggered gene expression levels in connection to high stromal TILs. One of these signature genes is HLF (Hepatic Leukemia Factor) that was found in negative relation with stromal TILs presence. In other words, the increasing HLF expression levels within tumor cells decreased the presence of stromal TILs and probably the lymphocytic infiltration in tumor in general.

Gene HLF is located on chromosome 17 (17q22), encodes for proline and acidic-rich (PAR) protein family member, and represents a bZIP (basic leucine zipper) transcription factor, as DBP (Albumin D Box-Binding Protein) and TEF (Thyrotrophic Embryonic Factor). Gene HLF was originally identified in a chromosomal translocation with the gene E2A, linked to acute lymphoblastic leukemia (ALL) (Inaba et al, 1992). This led to its aberrant expression as a fusion protein (E2A-HLF), and to a form of ALL connected to poor prognosis due to the resistance to ChT (Jabbour et al, 2015).

However, high impact was given to HLF in connection to circadian rhythms and the mammalian timing system. Transcription factor HLF, as one of the PAR bZIP proteins involved in circadian behaviour, is a regulatory protein that clearly varies with high amplitudes during circadian rhythms and is accepted as an output regulator of this process. The circadian genes have been implicated in the regulation of cell cycle, stress response and drug toxicity (Waters et al, 2013).

The chronotherapy and circadian rhythms consideration in cancer and metabolism will probably play more important role in drug development and therapeutic efficacy (Ferrell and Chiang, 2015). The potential importance of HLF functional analyses in cancer is underlined by certain studies of fatigue-related safety issues and shift work impact on human body. The rotating night shift work has been associated with increased risk of breast carcinoma (Schernhammer et al, 2001). Additionally, in the example of colon cancer, the improved chronopharmacology in 5-fluorouracil night time administration reduced the therapy toxicity and improved the tumor size reduction (Levi et al, 2001).

It has been shown that HLF regulates the expression of numerous genes involved in the metabolism of endobiotics and xenobiotics (Gachon et al, 2006). In this study, mouse models with PAR bZIP proteins triple knock-out (for Hlf, Dbp and Tef genes) were hypersensitive to xenobiotics and their early aging was detected as a consequence of the deficiency in xenobiotics detoxification properties. Recent studies with knock-out mice deficient in both alleles of mouse HLF showed that HSCs in these mice become more sensitive to 5- fluorouracil and that HLF is essential for maintaining the function of HSCs (Komorowska et al, 2015). Furthermore, the literature-based data are clearly connecting the HLF expression in cancer with reduced tumor cells apoptosis and improved cancer cell survival (Waters et al , 2013).

Given these previously published data, we decided to focus on HLF functional analysis, in order to study the role of the post-ChT lymphocytic attraction within tumor. For this objective, we decided to downregulate the expression of HLF in TNBC cell lines. Cells used for siHLF experiments were chosen according to literature-based data of HLF expression levels in various available BC cell lines (Kao et al, 2009).

Breast carcinoma cell lines SUM-52-PE, MDA-MB-468 and MDA-MB-231 were chosen for their respective high, moderate, or low, HLF expression levels (Figure 41A). The HLF mRNA levels were also tested in our laboratory conditions (Figure 41B) and compared to the ones obtained in literature, and immunoblot of HLF protein expression levels was performed in parallel (Figure 41C). According to literature-based and our conditions-based findings of HLF expression levels, we decided to consider both SUM-52-PE and MDA-MB-468 as cell lines with high HLF expression level for downregulation experiments.

For the initial experiments, the HLF gene expression in TNBC cell lines was inhibited by specific siRNA (ON-TARGETplus HLF siRNA, Dharmacon), using Lipofectamine RNAiMAX transfection agent (Figure 42 A). The CTRL siRNA (ON-TARGETplus Non- targeting siRNA, Dharmacon) was used as a negative control of transfection and further experiments were performed by comparing the HLF-knocked-down effect in siHLF cells vs siCTRL cells.

The previous genomic data of our group declared that HLF expression level was reduced in patients' samples with high post-ChT TILs presence. This supports the hypothesis of the low HLF expression levels being connected to ChT-sensitivity of cancer cells, so the first step of our experiments was to verify the possible effect of HLF knock-down on cellular viability under ChT treatment, 24 hours after transient transfection with siHLF and siCTRL. We performed a set of experiments using doxorubicin as a ChT treatment during 48 hours in various concentrations, and cell viability was determined using CellTiter Glow Luminescent Cell Viability Assay (Promega) according to the manufacturer's recommendations. As shown in Figure 42B, no significant difference in cell viability was detected between siHLF cells, when compared to siCTRL counterparts, in any of tested doxorubicin concentrations. Two other time points (24h, 72h) were applied and did not show any further effect (data not shown). Furthermore, the HLF expression level decrease, performed by transient transfection on both cell lines, did not significantly affect cellular viability or morphology (data not shown). This initial set of siRNA experiments of HLF downregulation has shown unclear results, so for the next experiments of HLF activity in ChT treated cells and to analyze various gene expression levels due to HLF downregulation, we decided to perform the HLF knock-down using CRISPR/Cas9 (Clustered Regularly Interspaced Palindromic Repeats Associated Protein 9) system. The CRISPR/Cas9 is a system of targeted genome editing that works with a principle of short guide RNA sequence that recognizes the target DNA with very limited off-target effect (Barrangou et ah, 2015). Subsequently, the endonuclease Cas9 is responsible for target DNA cleavage (DNA flanked by a protospacer-adjacent motif), and the DNA repair of both cleaved parts follows by the machineries of non-homologous end joining or homo logy-directed repair (Hsu et al., 2014).

Cell lines SUM-52-PE and MDA-MB-468 were transfected by Lipofectamine 2000 using plasmid pX278 with HLF-recognizing sequence developed by our collaborators from IGBMC, Strasbourg (Figure 43). The plasmid bears a sequence specific for human HLF gene, and another plasmid was developed for mouse Hlf editing, which is planned to use on murine models.

Transfection effectivity of plasmid pX278 carrying the GFP-tag was verified by IF and transfected cells selection in puromycin-containing culture medium was performed for 48 hours, in order to select only clones bearing a knock-down of HLF together with puromycin resistance cassette. Further subcloning of resistant clones was done and the HLF expression levels were tested in each of potentially HLF-knocked-down clones. Three clones for each of SUM-52-PE and MDA-MB-468 cell line were established and can serve as a model for studies of HLF knock-down in stable manner. The analyses of HLF knock-down effect on cells treated with doxorubicin are ongoing, and the preliminary data show the decreasing tendency in cancer cell viability as a direct effect of HLF knock-down. This trend is not yet clear and needs a further confirmation, although it is in line with literature-based information.

Additionally, the microarray-based gene expression analysis of genes affected by HLF expression level decrease is programmed in parallel in those cells carrying the HLF knockdown. Microarray gene expression analysis in HLF transduced cell models suggested the upregulation of cytochrome P450 enzymes, often associated with circadian rhythms and drug metabolism, as well as the upregulation of genes influencing chemical toxicity (Waters et ah, 2013). Gene expression analysis in our laboratory aims to compare the genomic profiles of TNBC cells with HLF knock-down vs control cells, and will inform us about the impact of HLF on breast carcinoma cells. The possible implication of HLF downregulation in apoptotic pathways, in drug metabolism, or in genes implicated in lymphocytic attraction will be studied intensively.

Breast mouse cell lines transfections by CRISPR/Cas9 method based plasmid to knock-out mouse HLF are ongoing in our laboratory. In this future project direction, mouse models are intended to be established, in order to be able to study the direct impact of HLF knock-out in the tumor development in vivo and to monitor the lymphocytic infiltration of these tumors. Since the carcinomas of TNBC subtype cannot be treated using ET-based agents or by anti- HER2 targeted therapy, the majority of these tumors are treated by ChT. The presence of TILs in tumor after neoadjuvant ChT is associated with good prognosis and therefore it is of major interest to find out the mechanisms of this lymphocytic infiltration. Potential therapeutic targets, involved in this mechanism, could serve for new therapies development and could improve the prognosis, when combined with standard ChT, applied on TNBC patients. Additionally, the role of potential predictive biomarkers of response to neoadjuvant ChT, such possibly HLF, could be very important, in order to avoid the over-dose of chemotherapeutic agents in potentially non-responding patients, or contrarily, to select those patients with high benefit of ChT in neoadjuvant settings.

SEQUENCE LISTING

SEQ ID NO:l

Homo sapiens guanylate binding protein 1 (GBP1), DNA NCBI Reference Sequence: NM 002053.2

SEQ ID NO:2

Homo sapiens HLF, PAR bZIP transcription factor (HLF), DNA NCBI Reference Sequence: NM 002126.4

SEQ ID NO:3

Homo sapiens C-X-C motif chemokine ligand 13 (CXCL13), DNA NCBI Reference Sequence: NM 006419.2

SEQ ID NO:4

Homo sapiens sulfotransferase family IE member 1 (SULT1E1), DNA NCBI Reference Sequence: NM 005420.2

FIGURES

Figures la and lb: Participants' flow chart in the training phase

Figure 2 :Box plots of raw data

Figure 3 : Density plots of raw data

Figure 4: Box plots after separate frozen normalization

Figure 5: Density plots after separate frozen normalization

Figure 6: Box plots after cross-platform normalization

Figure 7: Density plots after cross-platform normalization

Figure 8: Histograms of stromal TIL in TOP samples, MDACC samples and overall

Figure 9: Cross validated likelihood as a function of the tuning parameter

Figure 10: Cross validated likelihood as a function of the tuning parameter in the neighborhood of the maxima

Figure 11 : Histograms of the genomic predictor in TOP samples, MDACC sample and overall Figure 12: Histograms of the transformed genomic predictor in TOP samples, MDACC sample and overall

Figure 13: Check for non-log-linear effect of the predictor on distant relapse-free survival

Figure 14: Check for non-log-linear effect of the predictor on overall survival

Figure 15: Distant relapse-free survival of different risk groups - TER

Figure 16: Distant relapse-free survival of different risk groups - MED

Figure 17: Distant relapse-free survival of different risk groups - COX

Figure 18: Overall survival of different risk groups - TER

Figure 19: Overall survival of different risk groups - MED

Figure 20: Overall survival of different risk groups - COX

Figure 21 : Spearman pairwise correlation of genes - Training

Figure 22: Profiles of stromal TIL - Grey lines: individual profiles - Green line: mean profile

Figure 23: Check for non-log-linear effect of stromal TIL on distant relapse-free survival

Figure 24: Kaplan-Meier distant-relapse free survival curves according to stromal TIL cut-off

(50%)

Figure 25: Check for non-log-linear effect of stromal TIL on overall survival

Figure 26: Kaplan-Meier overall survival curves according to stromal TIL cut-off (50%)

Figure 27: Participants' flow chart of the validation dataset

Figure 28: Histograms of the genomic predictor in the validation dataset

Figure 29: Histograms of the transformed genomic predictor in the validation dataset Figure 30: Check for non- log- linear effect of the genomic predictor on distant relapse-free survival - Validation dataset - Patients achieving pCR

Figure 31 : Check for non- log- linear effect of the genomic predictor on distant relapse-free survival - Validation dataset

Figure 32: Distant relapse-free survival of different risk groups - No pCR - TER

Figure 33: Distant relapse-free survival of different risk groups - No pCR - MED

Figure 34: Distant relapse-free survival of different risk groups - No pCR - COX

Figure 35: Distant relapse-free survival of different risk groups - All patients - TER

Figure 36: Distant relapse-free survival of different risk groups - All patients - MED

Figure 37: Distant relapse-free survival of different risk groups - All patients - COX

Figure 38: Spearman pairwise correlation of genes - Validation

Figure 39: Histograms of the genomic predictor - Training vs. validation

Figure 40: Histograms of the transformed genomic predictor - Training vs. validation Figure 41 : Comparison of HLF expression levels in three breast carcinoma cell lines

Literature microarray-based log2 ratio of HLF mRNA levels, compared to "universal reference RNA" that represents a mixture of RNAs of 11 well described BC cell lines, on SUM-52-PE, MDA-MB-468 and MD-MB-231 cell lines (A) (Kao et al, 2009). HLF mRNA content showed as a ddCT with 18S expression levels as an internal control in our laboratory conditions (B). Western blot of HLF protein expression on three cell lines. Beta-tubulin was used as loading control. Used antibodies: rabbit monoclonal anti-HLF (Genetex), mouse monoclonal anti-P-tubulin (Sigma Aldrich). (C)

Figure 42: Cell lines SUM-52-PE and MDA-MB-468 with HLF siRNA knock-down

Cell lines SUM-52-PE and MDA-MB-468 were transfected with siRNA specific for HLF (or non-targeting control). The HLF mRNA expression level was tested for each experiment and was summarized in one graph. The amount of 18S mRNA was used as internal reference for normalizing qPCR. (A). Cell viability was tested between siHLF and siCTRL cells when treated with doxorubicin (B).

Figure 43: Plasmid for CRISPR/Cas9 human HLF targeted genome editing

The structure of plasmid pX278 (carrying the GFP-tag) for CRISPR/Cas9 human HLF targeted genome editing designed by Bernardo Reina San Martin, IGBMC, Strasbourg. Figure 44: Effect of changing the tuning parameter on the values of fitted regression coefficients

Figure 45: Fitted stromal TILs (Box-cox-transformed) vs. observed stromal TILs (Box-cox- transformed)