SUPERIOR BIOMARKER SIGNATURE TO PREDICT THE RESPONSE OF A BREAST CANCER PATIENT TO CHEMOTHERAPY

Abstract

The present invention relates to methods for predicting the response of a breast cancer patient to a chemotherapy. The present invention further relates to a method of determining whether to treat a breast cancer patient with a chemotherapy. The present invention also relates to a kit for predicting the response of a breast cancer patient to a chemotherapy.

Claims

1.-15. (canceled)

16. A method of predicting the response of a breast cancer patient to a chemotherapy based on a combination of levels determined from at least two biomarkers in a biological sample of the breast cancer patient, wherein the at least two biomarkers are selected from three groups, the at least two biomarkers belonging to different groups and differ from each other, wherein the three groups comprise a first group, a second group, and a third group, wherein the first group comprises SCUBE2, CA12, and ANXA9, the second group comprises ELF5, ROPN1, ROPN1B, SOX10, TMEM158, FAM171A1, and SFRP1, and the third group comprises NFIB and SFRP1.

17. The method of claim 16, wherein the combination of levels is determined from at least three biomarkers, at least one first biomarker, at least one second biomarker, and at least one third biomarker, wherein the at least one first biomarker is selected from the first group consisting of SCUBE2, CA12, and ANXA9, the at least one second biomarker is selected from the second group consisting of ELF5, ROPN1, ROPN1B, SOX10, TMEM158, FAM171A1, and SFRP1, and the at least one third biomarker is selected from the third group consisting of NFIB and SFRP1.

18. The method of claim 17, wherein the at least one first biomarker is SCUBE2, the at least one second biomarker is ELF5, and the at least one third biomarker is NFIB.

19. The method of claim 16, wherein the chemotherapy comprises the administration of a taxane.

20. The method of claim 19, wherein the taxane is paclitaxel or docetaxel.

21. The method of claim 16, wherein the response is a pathological complete response (pCR).

22. The method of claim 16, wherein the biological sample is a breast tumor sample.

23. The method of claim 22, wherein the breast tumor sample is a pre-treatment breast tumor sample.

24. The method of claim 16, wherein the breast cancer is HER2-negative breast cancer.

25. A method of determining whether to treat a breast cancer patient with a chemotherapy comprising the steps of: (i) carrying out the method of claim 16 to obtain patient specific data, (ii) determining whether to treat the breast cancer patient with a chemotherapy based on comparing the patient-specific data with at least one reference criterion, and (iii) if the patient-specific data meets the at least one reference criterion recommending treatment of the patient with a chemotherapy.

26. The method of claim 25, wherein the chemotherapy comprises the administration of a taxane.

27. The method of claim 26, wherein the taxane is paclitaxel or docetaxel.

28. The method of claim 25, wherein the breast cancer is HER2-negative breast cancer.

29. A kit for predicting the response of a breast cancer patient to a chemotherapy comprising means for determining the level of at least two biomarkers in a biological sample of a breast cancer patient, wherein the at least two biomarkers are selected from three groups, the at least two biomarkers belonging to different groups and differ from each other, wherein the three groups comprise a first group, a second group, and a third group, wherein the first group comprises SCUBE2, CA12, and ANXA9, the second group comprises ELF5, ROPN1, ROPN1B, SOX10, TMEM158, FAM171A1, and SFRP1, and the third group comprises NFIB and SFRP1.

Description

BRIEF DESCRIPTION OF THE FIGURES

[0327] FIG. 1: Applied biomarker workflow.

[0328] FIG. 2: A PCA of the already normalized data (RD group) of the discovery cohort showing a site effect between the data obtained at the I-SPY-1 and MDACC sites, respectively (upper panel (A)). The same figure applying quantile normalization to the discovery set (lower panel, (B)).

[0329] FIG. 3: Normalized data of discovery and validation cohorts according to the normalization prescription of the present inventors, see text for details (upper panel (A)). Signal distribution of the discovery cohort (lower panel (B)). Relative frequencies are shown such that the histogram shapes may be compared more easily. pCR=predict pathologic complete response, RD=non-responder.

[0330] FIG. 4: Bar graphs of the genes contained in the gene signature of the classification model comparing the responder class (pCR=predict pathologic complete response) with the non-responder class (RD). The case of the validation cohort is shown.

[0331] FIG. 5: Histograms of the genes contained in the gene signature of the classification model comparing the distributions within the responder class with the distributions within the non-responder class within the validation cohort.

[0332] FIG. 6: The receiver-operator characteristics curve of our model comparing the performances on the discovery and validation sets (upper panel (A)). Various performance characteristics depending on the chosen model-internal threshold used for predicting responders (lower panel (B)).

[0333] FIG. 7: The receiver-operator characteristics curves of our model comparing the site specific performances. The LBJ (panel (A)) and USO (panel (B)) site curves demonstrate the cross site validation performance of our classification model, while the MDACC site result (1 panel) is shown for completeness (panel (C)).

[0334] FIG. 8: Comparison of response rate of the model of the present inventors (OakLabs) to the cases without Companion diagnostic (CDx) assay and with the model by Hatzis et al.

EXAMPLES

1. Methods

[0335] CDx Workflow

[0336] In FIG. 1 summarizes the steps that are indispensable for a successful predictive biomarker study with maximum probability for obtaining a market-ready, approved companion diagnostic. Starting at taking (genome-wide) molecular data—already in the early stages of the study—and after proper data preprocessing (whitening) signals in the registered data can be readily analyzed and a robust gene signature for discerning non-responders from responders to the drug under study is developed. Further, a model is built on the basis of the underlying gene signature data and maximized for best performance. The optimal model is then refined and validated on an independent set of samples.

[0337] Choice of Data Set for Case Study

[0338] As prerequisite for a reliable and robust machine learning analysis the underlying data sets should fulfil the following points: a high enough total number of samples (a); from independent study sites (b); with global gene expression profiles publicly available (e.g. by ArrayExpress) (c); a single technology platform used for data acquisition (d); data taken at baseline of the study (e); sufficient meta data to access and work with (f); published in a high ranking journal (g). The first point is crucial in order to be able to split up the data into a discovery and a validation set, which is inevitable for a reliable machine learning analysis, since any algorithm needs to be validated with data that has not been used in the training phase in order to allow to assess its performance and in order to circumvent the well-known overfitting problem. Points (b) and (c) allow free access to the complete gene profile of patients while point (d) ensures that data is intrinsically comparable. Point (e) is necessary to ensure that differences in the responders and non-responders are not caused by the applied medical treatment. Point (f): it is important to have the means to understand possible patterns found while analyzing the data that are not caused by the treatment—think of lab effects, effects due to patient sex, age dependence etc.

[0339] The data obtained in the study undertaken by Hatzis et al. [11] of HER2 negative breast cancer patients meets all of the above listed criteria. The pretreatment characteristics of the discovery cohort of 310 patients and the validation cohort of 198 patients have been reported. The data has been taken prior to taxane-anthracycline based chemotherapy. All data has been acquired using the U133A GeneChip by Affymetrix. No details are available for the sequential taxane and anthracycline based chemotherapy for individual patients as well as on the question which patients of the validation cohort received entirely neoadjuvant (N=165), partial neoadjuvant (N=18) or entirely adjuvant (N=16) chemotherapy.

[0340] Evaluation of Data Quality and Data Processing

[0341] Gene expression raw data were obtained from the ArrayExpress online pages (IDs: E-GEOD-25055[13] and E-GEOD-25065[14]). They provide both the already normalized sets of the data as have been provided by the authors of [11] as well as the respective raw data.

[0342] It is widely accepted that gene expression based data requires a proper normalization of the recorded data for comparing samples among each other. On the other hand, it should be assured that the normalization method does not introduce clusters when normalizing data from different sites, different cohorts etc. This is even more true if the data is used to train a classification algorithm since it might learn a pattern that has been artificially introduced by the normalization.

2. Results

[0343] Unbiased Sample Normalization

[0344] As a possible starting point for model selection and for training the machine learning algorithms, the already normalized data provided via ArrayExpress were considered. However, it turned out that this data contains patterns revealed using a primary component analysis (PCA) when comparing the two sites I-SPY-1 and MDACC data, cf. the upper panel (A) of FIG. 2, where the RD group of the data is shown, as more samples are available in this group. The data projected on the two major axes in the PCA space is shown.

[0345] In an attempt to possibly eliminate or at least reduce the splitting between the two labs, the data were normalized using the Affymetrix package (“affy”) available for the R programming language to normalize the discovery cohort. Many options are available and provided by the “affy” package to process the input gene chip data using R. The “rma” method for background correction, quantile normalization using the “quantile” option, “pmonly” for only using the signals of the pm channel as well as “medianpolish” for data summary were employed. Although the effect does not completely vanish even when using this alternative normalization, the site-specific clustering in the PCA of the discovery cohort was significantly reduced as can be seen from the lower panel (B) of FIG. 2.

[0346] A known disadvantage of the quantile normalization is that it does not treat a single sample on its own, but the signals of all samples together are used to adapt the signal of each individual sample. In a machine learning context, this behavior of the normalization is disfavorable since a clear separation of the data used for training and tuning the model from the data used to validate the model is always mandatory in order to prevent overfitting. The validation of a model should be done using data that was unknown during the learning phase. In order to have normalized validation cohort samples which moreover are maximally independent of one another, it was chosen to normalize each validation cohort sample together with all available discovery cohort samples. In this way, it was possible to retain the advantages of quantile normalization also for the validation cohort, but avoid spilling validation cohort information into the discovery cohort, thus, guaranteeing independence of the training phase from the validation data. As an additional benefit, this method makes it possible to easily validate new samples with the already trained model and setup. The upper panel (A) of FIG. 3 shows that using this way of normalizing the validation cohort samples, no clustering is visible comparing the normalized discovery data to the normalized validation data in a PCA analysis. Thus, the two cohorts of data after the chosen normalization are compatible and are ready to be used as input to subsequent machine learning analysis.

[0347] Finally, it was also checked for the discovery cohort that after normalization a similar distribution of signals for the pCR and RD groups is present, cf. the lower panel (B) of FIG. 3 (pCR=pathological complete response, RD=residual invasive disease/non-responder).

[0348] Feature Selection and Classification Model

[0349] A major part in the development of a machine learning algorithm is the choice of an appropriate feature set (gene signature). A notorious problem in life sciences as compared to other fields, where AI algorithms are commonly applied, is the limited amount of samples (N) and the high costs related to each sample. On the other hand, the number of recorded features (genes, proteins etc.) M is usually much higher than N. Since after all, a machine learning algorithm is based in one way or another on a fitting i.e. regression technique, it is crucial to reduce the gene set sufficiently in order to have some degrees of freedom left within the fit. The feature set can be reduced using filters (for instance filtering on p-value or the mean signal), removing redundant i.e. highly correlated genes, using classifier based methods (such as RFE [15]) or using L1-norm based lasso techniques [16]. In addition, biological input concerning the mode of action of the drug and associated pathways is valuable to reduce the set further. A brute-force search over all possible gene subsets may also be done if starting from a sufficiently small feature set and using parallel computing techniques in order to overcome the associated 2.sup.M asymptotics.

[0350] It was made use of all above mentioned techniques to single out the optimal gene signature that allows the training of a performant classification algorithm and is yet small in size. It is however of utter importance to use proper resampling methods to obtain in average performance comparisons of different gene signatures. Resampling as well as cross-validation techniques on the discovery cohort were used to obtain robust metrics.

[0351] One may choose among a plethora of classification algorithms that are available on the market, such as linear models, tree-based models (with or without boosting), different kinds of support vector machines, network based classifiers etc. Each of them has its own right of existence and comes with its strengths and weaknesses. For example, some are good at capturing non-linear effects, others perform worse in such cases. Furthermore, there are algorithms that tend to more easily overfit than others. It is therefore crucial to understand the underlying data (noise, variance, reproducibility between labs) to choose the best classification model. Several classifiers were tested on our candidate gene signatures in order to choose the best performing algorithm. Thereby, the mean value of the achieved performance metric as well as strive for a low variation i.e. a low standard deviation, were taken into account. A summary of the tested classification models and their achieved performances in shown in Table 1.

[0352] Table 1 Examples of classifier performances in order to illustrate the need to test several algorithms. The achieved mean ROC area under the curve score and its standard error (68% CI) are shown.

TABLE-US-00001 Candidate Algorithm AUC score Decision Tree 0.77 (0.06) Logistic Regression 0.82 (0.05) Radial basis function SVM 0.82 (0.04) LogitBoost 0.80 (0.06)

[0353] The area under the curve (AUC) of the receiver-operator characteristics (ROC) was chosen as the most appropriate score method for the unbalanced data sets underlying this study. The commonly used accuracy would require a balancing of the unbalanced data within the two classes. Here, the comparison of the classifiers using a cross-validation technique on the whole discovery set was performed fixing the same ratio of training and test number of samples at 0.7 in each case. The achieved mean and standard error of the AUC were reported.

[0354] Performance of the Predictive Biomarker Signature in the Independent Validation Cohort

[0355] The final model, whose performance is the subject of this paragraph, is based on merely the three genes listed in Table 2.

TABLE-US-00002 TABLE 2 Gene signature Affymetrix Code Gene Symbol X219197_s_at SCUBE2 X220625_s_at ELF5 X209289_at NFIB

[0356] FIG. 4 shows bar graphs of the genes contained in the gene signature (SCUBE 2, ELF5, and NFIB) of the classification model comparing the responder class (pCR=predict pathologic complete response) with the non-responder class (RD). The case of the validation cohort is shown.

[0357] Comparing the histograms of these genes within the responder's group with the histograms within the non-responders group shown in FIG. 5 reveals that there are indeed differences even on the level of a single gene.

[0358] The model's performance on the discovery and validation cohorts is presented in the upper panel (A) of FIG. 6, where the ROC curve is shown. The curve as obtained on the discovery set as well as the curve obtained when predicting the unknown validation cohort data is shown. Various performance characteristics depending on the chosen model-internal threshold used for predicting responders are shown in the lower panel (B) of FIG. 6. It can be observed from the figure, both ROC are compatible, from which it is concluded that a valid classification model was developed, which would not be the case if either of the curve would deviate significantly from the other.

[0359] The performance of the new model for predicting pCR and RD was next compared with the model by Hatzis et al. [11]. The available meta data published with the data has been used to obtain them. In Table 3 the various model performance metrics were accumulated.

[0360] Table 3 Comparison of response prediction algorithm performance on the independent validation cohort (182 samples). The sensitivity of our model has been matched to the value of Hatzis et al. by setting the ROC work point to 0.520. See text for further details. CDx=Companion diagnostics, positive predictive value (PPV), negative predictive value (NPV).

TABLE-US-00003 Without CDx Hatzis et al. Present analysis Response rate 23% 33% 44% Sensitivity — 55% 57% Specificity — 67% 79% PPV — 33% 44% NPV — 83% 86%

[0361] The response rate without Companion diagnostics (CDx) (defined in the usual way as fraction of responders and all patients treated) and with CDx, which is equal to the positive predictive value (PPV) were also computed. Since all performance values are mutually dependent and furthermore depend intrinsically on the operation point of the classifier, i.e. the model-internal threshold for the probability of classifying “responders”, one can only compare two models if either of the performance metrics matches among the models. Our classifier's operation point was set to a value of 0.520 such that our achieved sensitivity matches as closely as possible the sensitivity of the model by Hatzis et al. Doing so, the model is completely fixed and the specificity, the positive predictive value (PPV), and the negative predictive value (NPV) in Table 3 are directly comparable. For reference, the dependence of the performance numbers in the lower panel (B) of FIG. 6 were additionally visualized.

[0362] Similarly, the operation point of our model may be chosen such that instead the specificity is matched to the one obtained by Hatzis et al., which is approximately obtained at a value of 0.437. The other performance characteristics in this case may be read from Table 4 and may be compared among the two models.

[0363] Table 4 Comparison of response prediction algorithm performance on the independent validation cohort (182 samples). The specificity of our model has been matched to the value of Hatzis et al. by setting the ROC work point to 0.437. See text for further details. CDx=Companion diagnostics, positive predictive value (PPV), negative predictive value (NPV).

TABLE-US-00004 Without CDx Hatzis et al. Present analysis Response rate 23% 33% 38% Sensitivity — 55% 68% Specificity — 67% 68% PPV — 33% 38% NPV — 83% 87%

[0364] The response rates of the cases without companion diagnostics, the Hatzis et al. model and the model of the present inventors evaluated at a threshold 0.520 is visualized in FIG. 8.

[0365] Cross-Site Validation

[0366] In order to rule out medical site-related biases of our model, a cross-site validation study was performed. Since data taken at the MDACC site has been included both in the discovery cohort and the validation cohort, while samples from the LBJ/INEN/GEICAM (for brevity called LBJ in what follows) and USO centers are only included in the validation cohort of the original study, the easiest way to accomplish a cross-site validation may be to predict only the data of the latter two sites.

[0367] In this way data originating from the same medical centers in both the learning and application stages of our model could be avoided, while still being able to compare our model's performance on the validation set with values reported in literature, as has been done above. The site-specific performances that are achieved by our model as presented above are summarized in Table 5.

[0368] Table 5 Cross-site study performance of our model at the ROC work point 0.520. The 95% confidence intervals are shown in parenthesis. See text for details. CDx=Companion diagnostics, positive predictive value (PPV), negative predictive value (NPV).

TABLE-US-00005 ROC Site Sensitivity Specificity PPV NPV AUC LBJ/INEN/ 45(27)% 85(11)% 42(22)% 87(7)% 0.73 GEICAM USO 64(22)% 66(16)% 51(15)% 77(11)%  0.65 MDACC 58(25)% 82(10)% 40(18)% 91(5)% 0.82 all sites 57(14)%  78(7)% 44(11)% 86(4)% 0.74

[0369] For reference, the performances obtained for all sites together were also included. Additionally, the 95% confidence interval of our model performance numbers was computed using a bootstrapping scheme over the validation cohort samples. Symmetrical 95% confidence intervals were found. The associated errors are listed in Table 3 in parenthesis. As can be seen, the values obtained for the USO and LBJ sites are compatible with the average obtained including all sites.

[0370] The site-specific ROC curves shown in FIG. 7 were achieved. Comparing the AUC scores, is has been observed that for the LBJ site (panel (A)) a similarly good score was obtained, while for the USO site (panel (B)) a slightly worse value was obtained, which has dropped by 0.1 as compared to the discovery set. Given the typical fluctuations, an AUC of about 0.04-0.06 at the 68% confidence level was observed (cf. Table 1). The drop is insignificant. The performance on the MDACC validation set on the other hand, which was added for completeness, is slightly better than the reference discovery value (panel (C)).

[0371] Example of a Mathematical Calculation with the Novel Signature of the Biomarkers SCUBE2, ELF5, and NFIB in Order to Predict Whether a Patient Will Respond to Chemotherapy or not

[0372] In the following, a mathematical calculation with the novel signature of the biomarkers SCUBE2, ELF5, and NFIB is shown in order to predict whether a patient will respond to chemotherapy or not:

The levels (expression levels) of the biomarkers SCUBE2, ELF5, and NFIB were designated with g:

gi,i∈E1, . . . ,3,

where [0373] g1: SCUBE2 [0374] g2: ELF 5 [0375] g3: NFIB
The levels were normalized and are, thus, dimensionless, i.e. they do not have associated units. The normalization was carried out using the Affymetrix packages (“affy”). More specifically, the packages: affy (Version 1.44.0), Biobase (Version 2.26.0), BiocGenerics (Version 0.12.1) were used.
To map the values of the three biomarkers to the two categories “responder” and “non-responder”, a function ƒ was chosen satisfying any and/or any arbitrary combinations of the following features:

[0376] without loss of generality ƒ takes values between 0.0 and 1.0

[0377] the image of ƒ describes a sigmoid curve

[0378] as an example for a sigmoid curve, a logit function may be used

[0379] ƒ is injective

[0380] ƒ is surjective

[0381] ƒ is bijective

In case of three biomarkers g1; g2; g3, a logit function was selected having the form:

[00002]$f\left(g1,g2,g3\right)=\frac{1}{1+e^{-\left(c0+c1g1+c2g2+c3g3\right)}}c0=1.76,c1=0.31,c2=-0.16,c3=-0.37\begin{array}{cc}c0:& \mathrm{constant}\mathrm{intercept}\\ c1:& \mathrm{coefficient}\mathrm{of}\mathrm{the}\mathrm{level}\mathrm{of}\mathrm{SCUBE}2\\ c2:& \mathrm{coefficient}\mathrm{of}\mathrm{the}\mathrm{level}\mathrm{of}\mathrm{ELF}5\\ c3:& \mathrm{coefficient}\mathrm{of}\mathrm{the}\mathrm{level}\mathrm{of}\mathrm{NFIB}\end{array}$

The coefficients of the levels and c0 were determined by mathematical/statistical evaluation of the reference data which has known clinical responders and known clinical non-responders to the chemotherapy, such that the function ƒ is fitted ideally to the reference data, e.g. by an optimization process such as by linear optimization. The method by Broyden, Fletcher, Goldfarb and Shanno (1-bfgs) was used here. Thus, based on the reference data, a prediction is made by calculating a score using the function ƒ. The score is the value of the function ƒ for the patient specific levels g1, g2, g3 using the coefficients mentioned above.
Consequently, the result of the calculation is a score which allows to predict the response of the breast cancer patient to chemotherapy.
To finally make the decision whether the patient is regarded as a “responder” or “non-responder”, e.g. a patient which should be treated or a patient which should not be treated, a specific threshold parameter ξ is selected within the value range off:

ξ∈[0.0,1.0].

In case of E [0.0, 1.0],

∀gi∈.sup.+,i∈{1, . . . ,3}: ƒ(g1,g2,g3)≥ξ.Math.responder  (1)

ƒ(g1,g2,g3)<ξ.Math.non-responder  (2).

FIG. 6 shows the sensitivity, specificity, positive predictive value (PPV) and the negative predictive value (NPV) for different values of the threshold parameter ξ.
In fact, the new response rate of Taxane restricted to the patients which are predicted responders by this model is identical to the PPV. The sensitivity denotes the fraction of true responders and all responders.
The meaning of the parameter ξ is obvious when selecting the external values. If ξ is set to ξ=0.0 equation, this suggests that all patients are considered as potential responders while no patient is excluded as a potential non-responder. In this case the PPV should match the actual response rate (23%) of the Taxane without any Companion diagnostics (CDx) which is given at PPV=0.23. At this threshold, the sensitivity, which is defined by the fraction of true responders by all responders, is exactly sensitivity=1.0. Increasing the parameter ξ increases the PPV while at the same time the sensitivity decreases as responders are lost which are wrongly classified as non-responders. At ξ=1.0, the model classifies all patients as non-responders which yields the highest specificity. The specificity is the fraction of true predicted non-responders by the number of all non-responders.
Obviously, the extremal points of ξ are not useful. The plot shows that ξ should be chosen between 0.2 and 0.7 as this range has the highest economic impact with respect to the number of patients treated and the achieved response rate.

0.2≤ƒ(g1,g2,g3)≤0.7,gi∈.sup.+,i∈1, . . . ,3

Here specific statistics for a selection of ξ at the limits of the above range ξ∈[0.2, 0.7] assuming a total number of 1.000 patients are given:

Without Companion Diagnostics (CDx):

[0382] Total patients: 1.000
Response rate: 0.23 (=23%)

Responders: 230

Non-Responders: 770

With ξ=0.21:

[0383] Total patients: 1.000
Response rate (PPV): 0.29 (=29%)
Predicted responders: 728

True Responders: 214

Non-Responders: 515

With ξ=0.68:

[0384] Total patients: 1.000
Response rate (PPV): 0.56 (=56%)
Predicted responders: 137

Responders: 77

Non-Responders: 60

[0385] Further, the following ratios could be reached:
SCUBE2: mean (R)/mean (NR)=0.86+/−0.06
ELF5: mean (R)/mean (NR)=1.22+/−0.02
NFIB: mean (R)/mean (NR)=1.18+/−0.02

R=Responder

NR=Non-Responder

[0386] SCUBE2, ELF5, and NFIB and their Correlated Genes

The signature comprising the genes SCUBE2, ELF5, and NFIB was determined. In addition, correlated genes of SCUBE2, ELF5, and NFIB were identified. The level of said genes can alternatively be measured/determined.
SCUBE2 (geneID=57758′) and its correlated genes CA12 (gene_id=771) and ANXA9 (gene_id=8416),
ELF5 (geneID=2001′) and its correlated genes ROPN1 (gene_id=54763), ROPN1B (gene_id=152015), SOX10 (gene_id=6663), TMEM158 (gene_id=25907), FAM171A1 (gene_id=221061), and SFRP1 (gene_id=6422), and
NFIB (geneID=4781′) and its correlated gene SFRP1 (gene_id=6422).
SCUBE2, ELF5, and NFIB are genes.
Two genes are said to be correlated if their variation about their respective mean values is not statistically independent, but mutually and linearly related. The Pearson correlation coefficient, which normalizes the expectation value of the common variation about the mean value of the genes with the product of the standard deviations of the two gene's signals, has been used here.
In addition to the signature(s) described above, the following signatures were determined/calculated:

1. ILF2, CXCR4, and WWP1,

2. IGHG1, IGHG3, IGHM, IGHV4-31, ID4, and CSRP2, or

3. DNAJC12, PRSS23, and TTC39A.

[0387] They allow the prediction of the response of a breast cancer patient to chemotherapy. The prediction response (e.g. with respect to sensitivity and/or specificity) was, however, not as good as for the signature(s) described above.

3. Discussion

[0388] In this study, results in life sciences were improved by using dedicated new AI concepts. A well suited case example of high medical relevance in the field of breast cancer was chosen and demonstrated the superiority of our approach: With a model of just 3 genes the response rate can almost be increased by 33% compared to the benchmark published by Hatzis et al.

[0389] Having evolved in the field of image recognition, artificial intelligence and machine learning algorithms are increasingly employed for tasks in life sciences. While images are highly reproducible and contain several million data points (pixels), life science data are quite different in respect to number of data points and noise for example. Algorithms in image recognition require approximations to deliver results within minutes. In contrast, the major demand on predictive biomarkers is maximum ac-curacy. This can only be achieved by complete avoidance of approximations which in turn increases the computing time. Two months of computing time on a compute cluster with 80 compute cores were necessary for our results.

[0390] The majority of public genome-wide gene expression data is not compatible with an approach to develop reliable predictive biomarkers, mainly due to limitations in sample size. An integrative analysis of raw data from independent studies could improve the situation, but comes with a number of challenges. Differences in the experimental protocols or technology platform used can introduce systematic variation across studies. The focus here was on gene expression data of sufficient samples obtained on a single technology platform with minimal variation in the experimental protocols. Such a setting could easily be implemented as part of a clinical phase 3 and is compatible with a straightforward translation of the developed biomarker signature to a companion diagnostics assay.

[0391] An interdisciplinary team of quantum physicists and life scientists was able to develop and cross-site validate a 3-genes predictive biomarker signature which is capable of nearly doubling the response rate within the group of predicted responders.

[0392] Adding strength to our results is that all three genes are biologically plausible. They all are described in the literature in the context of cancer and breast cancer in particular. SCUBE2 (Signal peptide-complement protein C1r/C1s, Uegf, and Bmp1 [CUB]-epidermal growth factor [EGF] domain-containing protein) is an 807-amino acids protein that belongs to a small family of three members. SCUBE2 is predominantly expressed in vascular endothelial cells [17] and regulates the SHH (Sonic Hedgehog) signaling, acting upstream of ligand binding at the plasma membrane [18]. Mounting evidence suggests that SCUBE2 acts as a tumor suppressor in breast cancer [19,20], NSCLC [21], colorectal cancer [22] and gastric cancer [23].

[0393] ELF5 (E74 Like E26 transformation-specific [ETS] Transcription Factor 5) is a 265-amino acids protein and a member of the ETS family of transcription factors. ETS family proteins regulate a wide spectrum of biological processes and several ETS factors have been implicated with cancer initiation, progression and metastasis [25,26]. For ELF5, both tumor promoting and suppressive roles have been reported in breast cancer [27].

[0394] NFIB belongs to the nuclear factor 1 (NFI) family of transcription factors which control expression of a large number of cellular genes [29,30]. In a hetero and homodimer complex, the four members of the NFI family can activate or repress transcription depending on the context [30]. NFIB has been defined as an oncogene in several reports [31,32]. The chromosomal region encoding NFIB is amplified in TNBC [33].

4. Conclusion

[0395] A novel AI-based approach enabled the development of a predictive biomarker signature that significantly outperforms the benchmark in respect to accuracy, number of features and reproducibility. The small size of the signature allows efficient translation to a CDx assay that is compatible with technology in routine diagnostic laboratories. Especially in view of increasing costs and time for clinical trials, predictive single drug biomarkers combined with modern trial designs offer the opportunity to increase the R&D productivity in healthcare.

REFERENCES

[0396] 1. Learn, P. A., Yeh, I.-T., McNutt, M., Chisholm, G. B., Pollock, B. H., Rousseau Jr, D. L., Sharkey, F. E., Cruz, A. B., Kahlenberg, M. S.: Her-2/neu expression as a predictor of response to neoadjuvant docetaxel in patients with operable breast carcinoma. Cancer: Interdisciplinary International Journal of the American Cancer Society 103(11), 2252-2260 (2005) [0397] 2. Vogel, C., Cobleigh, M., Tripathy, D., Gutheil, J., Harris, L., Fehrenbacher, L., Slamon, D., Murphy, M., Novotny, W., Burchmore, M., et al.: First-line, single-agent herceptin® (trastuzumab) in metastatic breast cancer: a preliminary report. European journal of cancer 37, 25-29 (2001) [0398] 3. Audeh, M. W., Carmichael, J., Penson, R. T., Friedlander, M., Powell, B., Bell-McGuinn, K. M., Scott, C., Weitzel, J. N., Oaknin, A., Loman, N., et al.: Oral poly (adp-ribose) polymerase inhibitor olaparib in patients with brca1 or brca2 mutations and recurrent ovarian cancer: a proof-of-concept trial. The Lancet 376(9737), 245-251 (2010) [0399] 4. Kaufman, B., Shapira-Frommer, R., Schmutzler, R. K., Audeh, M. W., Friedlander, M., Balmaña, J., Mitchell, G., Fried, G., Stemmer, S. M., Hubert, A., et al.: Olaparib monotherapy in patients with advanced cancer and a germline brca1/2 mutation. Journal of clinical oncology: official journal of the American Society of Clinical Oncology 33(3), 244 (2015) [0400] 5. Herbst, R. S., Soria, J.-C., Kowanetz, M., Fine, G. D., Hamid, O., Gordon, M. S., Sosman, J. A., McDermott, D. F., Powderly, J. D., Gettinger, S. N., et al.: Predictive correlates of response to the anti-pd-11 antibody mpd13280a in cancer patients. Nature 515(7528), 563 (2014) [0401] 6. Garon, E. B., Rizvi, N. A., Hui, R., Leighl, N., Balmanoukian, A. S., Eder, J. P., Patnaik, A., Aggarwal, C., Gubens, M., Horn, L., et al.: Pembrolizumab for the treatment of non-small-cell lung cancer. New England Journal of Medicine 372(21), 2018-2028 (2015) [0402] 7. Herbst, R. S., Baas, P., Kim, D.-W., Felip, E., Pérez-Gracia, J. L., Han, J.-Y., Molina, J., Kim, J.-H., Arvis, C. D., Ahn, M.-J., et al.: Pembrolizumab versus docetaxel for previously treated, pd-11-positive, advanced non-small-cell lung cancer (keynote-010): a randomized controlled trial. The Lancet 387(10027), 1540-1550 (2016) [0403] 8. Kim, S., Lin, C.-W., Tseng, G. C.: Metaktsp: a meta-analytic top scoring pair method for robust cross-study validation of omics prediction analysis. Bioinformatics 32(13), 1966-1973 (2016) [0404] 9. Rohart, F., Eslami, A., Matigian, N., Bougeard, S., Le Cao, K.-A.: Mint: a multivariate integrative method to identify reproducible molecular signatures across independent experiments and platforms. BMC bioinformatics 18(1), 128 (2017) [0405] 10. Harris, L. N., Ismaila, N., McShane, L. M., Andre, F., Collyar, D. E., Gonzalez-Angulo, A. M., Hammond, E. H., Kuderer, N. M., Liu, M. C., Mennel, R. G., et al.: Use of biomarkers to guide decisions on adjuvant systemic therapy for women with early-stage invasive breast cancer: American society of clinical oncology clinical practice guideline. Journal of Clinical Oncology 34(10), 1134 (2016) [0406] 11. Hatzis, C., Pusztai, L., Valero, V., Booser, D. J., Esserman, L., Lluch, A., Vidaurre, T., Holmes, F., Souchon, E., Wang, H., et al.: A genomic predictor of response and survival following taxane-anthracycline chemotherapy for invasive breast cancer. Jama 305(18), 1873-1881 (2011) [0407] 12. Bianco, S., Burger, F., Kallarackal, J., Romualdi, A., Schad, M.: Prediction of sensitivity to taxane-antracycline chemotherapy in invasive breast cancer (in preparation). TBA (2019) [0408] 13. Hatzis, C.: Discovery cohort for genomic predictor of response and survival following neoadjuvant taxane-anthracycline chemotherapy in breast cancer. https://www.ebi.ac.uk/arrayexpress/experiments/E-GEOD-25055/?query=GSE25055. [Online; accessed 5 Jun. 2019]” (2011) [0409] 14. Hatzis, C.: Validation cohort for genomic predictor of response and survival following neoadjuvant taxane-anthracycline chemotherapy in breast cancer. https://www.ebi.ac.uk/arrayexpress/experiments/E-GEOD-25065/?query=GSE25065. [Online; accessed 5 Jun. 2019] (2011) [0410] 15. Guyon, I., Weston, J., Barnhill, S., Vapnik, V.: Gene selection for cancer classification using support vector machines. Machine Learning 46(1), 389-422 (2002). doi:10.1023/A:1012487302797 [0411] 16. Tibshirani, R.: Regression shrinkage and selection via the lasso: a retrospective. (2011) [0412] 17. Yang, R.-B., Ng, C. K. D., Wasserman, S. M., Colman, S. D., Shenoy, S., Mehraban, F., Kömüves, L. G., Tomlinson, J. E., Topper, J. N.: Identification of a novel family of cell-surface proteins expressed in human vascular endothelium. Journal of Biological Chemistry 277(48), 46364-46373 (2002) [0413] 18. Tsai, M.-T., Cheng, C.-J., Lin, Y.-C., Chen, C.-C., Wu, A.-R., Wu, M.-T., Hsu, C.-C., Yang, R.-B.: Isolation and characterization of a secreted, cell-surface glycoprotein scube2 from humans. Biochemical Journal 422(1), 119-128 (2009) [0414] 19. Cheng, C.-J., Lin, Y.-C., Tsai, M.-T., Chen, C.-S., Hsieh, M.-C., Chen, C.-L., Yang, R.-B.: Scube2 suppresses breast tumor cell proliferation and confers a favorable prognosis in invasive breast cancer. Cancer Research 69(8), 3634-3641 (2009) [0415] 20. Lin, Y.-C., Chen, C.-C., Cheng, C.-J., Yang, R.-B.: Domain and functional analysis of a novel breast tumor suppressor protein, scube2. Journal of Biological Chemistry 286(30), 27039-27047 (2011) [0416] 21. Yang, B., Miao, S., Li, Y.: Scube2 inhibits the proliferation, migration and invasion of human non-small cell lung cancer cells through regulation of the sonic hedgehog signaling pathway. Gene 672, 143-149 (2018) [0417] 22. Song, Q., Li, C., Feng, X., Yu, A., Tang, H., Peng, Z., Wang, X.: Decreased expression of scube2 is associated with progression and prognosis in colorectal cancer. Oncology reports 33(4), 1956-1964 (2015) [0418] 23. Wang, X., Zhong, R.-Y., Xiang, X.-J.: Reduced expression of scube2 predicts poor prognosis in gastric cancer patients. INTERNATIONAL JOURNAL OF CLINICAL AND EXPERIMENTAL PATHOLOGY 11(2), 972-980 (2018) [0419] 24. Van′t Veer, L. J., Dai, H., Van De Vijver, M. J., He, Y. D., Hart, A. A., Mao, M., Peterse, H. L., Van Der Kooy, K., Marton, M. J., Witteveen, A. T., et al.: Gene expression profiling predicts clinical outcome of breast cancer. nature 415(6871), 530 (2002) [0420] 25. Sharrocks, A. D.: The ets-domain transcription factor family. Nature reviews Molecular cell biology 2(11), 827 (2001) [0421] 26. Hsu, T., Trojanowska, M., Watson, D. K.: Ets proteins in biological control and cancer. Journal of cellular biochemistry 91(5), 896-903 (2004) [0422] 27. Luk, I., Reehorst, C., Mariadason, J.: Elf3, elf5, ehf and spdef transcription factors in tissue homeostasis and cancer. Molecules 23(9), 2191 (2018) [0423] 28. Omata, F., McNamara, K. M., Suzuki, K., Abe, E., Hirakawa, H., Ishida, T., Ohuchi, N., Sasano, H.: Effect of the normal mammary differentiation regulator elf5 upon clinical outcomes of triple negative breast cancers patients. Breast Cancer 25(4), 489-496 (2018) [0424] 29. Gronostaj ski, R. M.: Roles of the nfi/ctf gene family in transcription and development. Gene 249(1-2), 31-45 (2000) [0425] 30. Harris, L., Genovesi, L. A., Gronostaj ski, R. M., Wainwright, B. J., Piper, M.: Nuclear factor one transcription factors: divergent functions in developmental versus adult stem cell populations. Developmental dynamics 244(3), 227-238 (2015) [0426] 31. Dooley, A. L., Winslow, M. M., Chiang, D. Y., Banerji, S., Stransky, N., Dayton, T. L., Snyder, E. L., Senna, S., Whittaker, C. A., Bronson, R. T., et al.: Nuclear factor i/b is an oncogene in small cell lung cancer. Genes & development 25(14), 1470-1475 (2011) [0427] 32. Zhang, Q., Cao, L.-Y., Cheng, S.-J., Zhang, A.-M., Jin, X.-S., Li, Y.: p53-induced microrna-1246 inhibits the cell growth of human hepatocellular carcinoma cells by targeting nfib. Oncology reports 33(3), 1335-1341 (2015) [0428] 33. Han, W., Jung, E.-M., Cho, J., Lee, J. W., Hwang, K.-T., Yang, S.-J., Kang, J. J., Bae, J.-Y., Jeon, Y. K., Park, I.-A., et al.: Dna copy number alterations and expression of relevant genes in triple-negative breast cancer. Genes, Chromosomes and Cancer 47(6), 490-499 (2008)