Method for predicting the risk of a cancer subject for developing radiation-induced skin fibrosis

Abstract

The present invention relates to a method for predicting the risk of a cancer subject for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the method comprises (a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or (b) genotyping the SNP rs1805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the major homozygous genotype (CC) has a lower risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

Description

[0001] Method for predicting the risk of a cancer subject for developing radiation-induced skin fibrosis

[0002] The present invention relates to a method for predicting the risk of a cancer subject for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the method comprises (a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or (b) genotyping the SNP rsl805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the major homozygous genotype (CC) has a lower risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

[0003] In this specification, a number of documents including patent applications and manufacturer's manuals are cited. The disclosure of these documents, while not considered relevant for the patentability of this invention, is herewith incorporated by reference in its entirety. More specifically, all referenced documents are incorporated by reference to the same extent as if each individual document was specifically and individually indicated to be incorporated by reference.

[0004] Radiation therapy (RT) is an important part of multimodal cancer therapy, and with rising numbers of long-term survivors, dose-limiting late reactions in healthy tissues become increasingly important. There is considerable variation in risk between individual patients with up to 80% being due to nonstochastic factors associated with genetics, clinical characteristics and lifestyle (1). In order to implement personalized treatments, reliable methods for predicting individual risk will be required. Furthermore, a better mechanistic understanding of risk factors might lead to novel interventional approaches. While early association studies on candidate gene single-nucleotide polymorphisms (SNPs) did not always validate in subsequent studies (reviewed in (2)) some associations have been confirmed in large discovery-validation studies or meta-analyses (3-7). Genome-wide association studies (GWAS) are considered more promising but require very large patient numbers from different centers, which inevitably increases heterogeneity within the clinical data. Thus, relatively few SNPs associated with late toxicity have been identified in medium-to-large GWAS (7-11).

[0005] Although significant associations of the TGFB1 SNP rsl800469 with fibrosis were reported in early studies this could not be confirmed in the larger RAPPER study (12) nor in a very large meta-analysis (13). By contrast, a SNP (rs2682585) in the base excision repair gene, XRCC1, which had shown variable significance in early studies was found to be significant in a very large metaphase study (5). Similarly, a SNP (rsl801516) in the ATM gene was confirmed to be associated with fibrosis (6).

[0006] It is frequently assumed, that normal-tissue reactions are determined by a large number of genetic factors each conferring a modest odds ratio (e.g. OR=1.1) to the overall risk (14). This would imply that individual factors act independently of each other and that overall risk may be modelled by logistic regression, mixed models or similar statistical approaches. Furthermore, a common type of damage (e.g. DNA or chromosome damage) is often assumed to be responsible for different toxicities, implicating genetic variants in DNA damage repair genes as risk factors. This assumption underlies the standardized 'STAT' score (15). However, the absence of a correlation between different endpoints in individual patients in some (16,17) though not all studies (18) support an alternative view that different endpoints may develop via different pathways. Thus, it may be hypothesized that subgroups of patients with different risk factors exist, which might be distinguished by a combination of 'omics' and functional assays (2).

[0007] Among functional predictive assays, the radiation-induced lymphocyte apoptosis (RILA) assay on CD8+and / or CD4+T cell populations ('killer' and 'helper' T cells, respectively) is currently the most promising (reviewed in (2)). It showed associations with mixed late reactions, breast fibrosis, and telangiectasia (19-22) and is being validated in the large prospective REQUITE study (23).

[0008] Whereas standard statistical inference tests may be used for simple associations, cross-validation in predictive models is important in order to reduce the influence of sample selection. Advanced machine learning techniques (e.g. logistic regression, random forest, deep learning) provide a high degree of stability to the results by incorporating repeated cross-validations and have been used to develop predictive models based on different combinations of omics, functional and clinical variables (11,24- 27).

[0009] Skin fibrosis (in particular subcutaneous fibrosis) after RT cancers such as breast cancer can seriously impact tissue function, cosmesis, and patients' quality of life, and are useful clinical models for the late reaction because of their prevalence and observability. Prediction of patients' risk of adverse late reactions in healthy tissue after radiation therapy is a prerequisite for personalized treatment to improve the quality of life for long-term survivors.

[0010] Hence, an unmet need for the identification of factors that confer increased or decreased risk of skin fibrosis after RT (and optionally also after tumor surgery) currently exists. This need is addressed by the present invention.

[0011] The present invention therefore relates in a first aspect to a method for predicting the risk of a cancer subject for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the method comprises (a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or (b) genotyping the SNP rsl805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the major homozygous genotype (CC) has a lower risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

[0012] The cancer subject refers to a subject having a cancer. The term "after radiation therapy and optionally after cancer surgery" means that the cancer of the subject was treated by a radiation therapy and may also have been treated by surgery.

[0013] The subject is a human, preferably a female human.

[0014] Cancer is an abnormal malignant new growth of tissue that possesses no physiological function and arises from uncontrolled usually rapid cellular proliferation. The cancer is preferably selected from the group consisting of breast cancer, lung cancer, ovarian cancer, endometrial cancer, vaginal cancer, vulva cancer, bladder cancer, salivary gland cancer, pancreatic cancer, thyroid cancer, kidney cancer, cancer concerning the upper gastrointestinal tract, colon cancer, colorectal cancer, prostate cancer, squamous-cell carcinoma of the head and neck, cervical cancer, glioblastomas, malignant ascites, lymphomas and leukemias. Among this list breast cancer is preferred, in particular female breast cancer. The cancer is preferably a solid cancer. A solid cancer is an abnormal mass of tissue that usually does not contain cysts or liquid areas by contrast to a liquid tumor.

[0015] Radiation therapy is a treatment where radiation is used to kill cancer cells in the subject being treated. Radiation therapy may be used in the early stages of cancer or after it has started to spread. It can be used to: (i) try to cure the cancer completely (curative radiation therapy), (ii) make other treatments more effective, for example, it can be combined with chemotherapy or used before surgery (neoadjuvant radiation therapy), (iii) to reduce the risk of the cancer coming back after surgery (adjuvant radiation therapy), and (iv) relieve symptoms if a cure is not possible (palliative radiation therapy). The radiation therapy is preferably an adjuvant radiation therapy and less preferably intraoperative radiation therapy.

[0016] Cancer surgery is an operation or procedure to remove cancer cells from the body. Sometimes, some surrounding healthy tissue is also removed.

[0017] As well as killing cancer cells, radiation therapy can damage some healthy cells in the area being treated. This can cause some side effects, such as: sore skin that might change colour to red, lighter or darker than the usual skin tone, feeling tired, hair loss in the area being treated, feeling sick, losing your appetite, a sore mouth and diarrhoea. These side effects are all early side effects.

[0018] Long-term side effects or late side effects are by definition effects that occur at least three months after radiation therapy. One long-term side effect or late side effect of radiation therapy for cancer patients is radiation-induced fibrosis, more commonly referred to as radiation fibrosis syndrome (RFS). RFS can manifest in various tissues including skin, lung, gastrointestinal tract, muscle, and numerous other tissues and organs depending on the treatment. In accordance with the present invention, the RFS is skin RFS.

[0019] Symptoms of this disease can begin appearing at least three months after treatment and the severity of the disease generally correlates with increased and prolonged doses of targeted radiation. The general features of this disease include the presence of infiltrating inflammatory cells into tissue, the differentiation of fibroblasts into myofibroblasts, and excessive production of extracellular matrix components including collagen. As with other forms of fibrosis, TGFP plays a key role in this process and leads to increased levels of growth factors and pro-inflammatory mediators. While this may present an interesting target for drug development, many other factors must also be considered, as radiation fibrosis is a complex disease. Radiation can induce injury to the target tissue by both direct action on important biomolecules or indirect action via free radicals formed in water surrounding the biomolecules. Radiation causes excessive production of reactive oxygen species (ROS) and reactive nitrogen species (RNS) which in turn leads to both vascular and parenchymal cell death and tissue damage. Radiation also leads to upregulation of a number of pro-inflammatory cytokines and chemokines. The increased inflammation that occurs after irradiation also drives the recruitment of bone marrow-derived cells to the site of injury, including mesenchymal stem cells, endothelial progenitor cells, and myelomonocytic cells, all of which may play a role in the progression of fibrosis. The skin is often affected, with both acute and chronic phases of skin injury. Early effects of irradiation of the skin include increased erythema, desquamation, and possibly ulceration. Later effects include dermal fibrosis, sebaceous and sweat gland atrophy, hair loss, and telangiectasia. Further information on RFS can be taken from Straub et al. and Nogueira et al. (28,29).

[0020] A single nucleotide polymorphism (SNP) is a genomic variant at a single base position in the DNA. The difference between point mutations and SNPs is defined by the minor allele frequency (MAF) in the population being 1% or higher for SNPs whereas point mutations have frequencies <1%. For example, a SNP may replace the nucleotide cytosine (C) with the nucleotide thymine (T) in a certain stretch of DNA. More than 85 million SNPs have been identified in the 1000 Genomes project. In each human subject, they occur almost once in every 1,000 nucleotides on average, which means there are roughly 4 to 5 million SNPs in a human's genome. Most SNPs have no detectable effect on health or development. However, some of these genetic differences have proven to be very important in the study of human health. SNPs can help in predicting an individual's response to certain drugs, susceptibility to environmental factors such as toxins, and risk of developing diseases. SNPs can also be used to track the inheritance of disease-associated genetic variants within families. Research is ongoing to identify SNPs associated with complex diseases such as heart disease, diabetes, and cancer.

[0021] The SNP rs9399005 can be found on human chromosome 6 at position 131947824 (reference genome built GRCh38 38.1 / 141). The alleles of this SNP are (C) (major allele) and (T) (minor allele) and the genotypes are (CC), (CT) and (TT).

[0022] The SNP rsl805794 can be found on human chromosome 8 at position 89978251 (reference genome built GRCh38 38.1 / 141). The alleles of this SNP are (C) (major allele) and (G) (minor allele) and the genotypes are (CC), (CG) and (GG). The SNPs rs9399005 and rsl805794 can be found at the genomic loci of the Connective Tissue Growth Factor (CTGF) and the Nijmegen Breakage Syndrome 1 (NBS1) gene, respectively. CTGF (also known as Cellular Communication Network Factor 1, CCN2) is a matricellular protein of the CCN family of extracellular matrix-associated heparin-binding proteins. CTGF has important roles in many biological processes, including cell adhesion, migration, proliferation, angiogenesis, skeletal development, and tissue wound repair, and is critically involved in fibrotic disease and several forms of cancers. NBS1 (also known as Nibrin or NBN) is a protein, being associated with the repair of DNA double strand breaks (DSBs) which pose a serious threat to a genome. It is a 754 amino acid protein identified as a part of the Mrell / RAD5O / NBS1 (MRN) repair complex. This complex recognizes DNA damage and rapidly relocates to DSB sites and forms nuclear foci.

[0023] Table 1 : SNPs whose genotypes correlate with rs1805794 in at least of 80% of the cases, according to 1000 Genomes CEU dataset

[0024]

[0025] The SNPs in Table 1 were selected based on the analysis of the 1000 Genome project dataset; see 1000 Genomes Project Consortium; Auton et al (1025), Nature; 526(7571):68-74 (30). The SNPs in Table 1 have high linkage disequilibrium with rsl805794 (i.e. high correlation level). Since the genotypes of the SNPs in Table 1 correlate with rs1805794 in at least of 80% of the cases (r2= between 0.8 and 1), it can be assumed that any of these SNPs can be used instead of rs1805794. The term “SNP rs1805794 or a SNP of Table 1 ” as used therefore means that either the rs1805794 is used or any one of the SNPs of T able 1 , but also covers the possibility that two or more (such as three or more, four or more, five or more, or ten or more) of the SNPs of rs1805794 and the SNPs of Table 1 as used.

[0026] In the above Table 1 those SNPs whose genotypes correlate with rsl805794 in 100% of the cases (r2= 1) are preferred. The most preferred SNPs among the SNPs of rs1805794 and in Table 1 - to be used alone or in combination - rs1805794.

[0027] The "risk" to be determined by the method of the invention may also be designated as the likelihood of a cancer subject for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery.

[0028] SNP genotyping is the measurement of genetic variations of single nucleotide polymorphisms (SNPs), whereby the SNP is in accordance with the present invention SNP rs9399005 and / or SNP rsl805794 or a SNP of Table 1. Means and methods for SNP genotyping are available in the prior art and include, for example, dynamic allele-specific hybridization, molecular beacons, SNP microarrays, allelic discrimination or Sanger sequencing.

[0029] Allelic discrimination is preferred method of determining SNP genotype (i.e. genotyping). Allelic discrimination are methods to discriminate between 2 or more different alleles. The method for allelic discrimination is preferably real-time polymerase chain reaction (PCR) allelic discrimination. In realtime PCR, different probes or DNA intercalating dye systems are used to distinguish between the different genotypes and / or alleles.

[0030] The sample obtained from the subject can be a tissue sample or body fluid sample. The tissue sample can be cancerous or non-cancerous tissue. The body fluid can be blood, serum plasma, saliva, semen, vaginal fluids, mucus, urine, sweat, cerebrospinal fluid that bathes the brain and spinal cord, lymph, the synovial fluid in joints, the pleural fluid in the pleural cavities, the pericardial fluid in the cardiac sac, the peritoneal fluid in the peritoneal cavity, and the aqueous humor of the eye.

[0031] As can be taken from the appended examples it was surprisingly found that (a) genotyping the SNP rs9399005 and / or SNP rsl805794 in a sample obtained from a cancer subject can be used to assess the risk of the cancer subject of developing skin fibrosis after radiation therapy and optionally after cancer surgery. In connection with the rs9399005, the minor homozygous genotype (TT) has a low risk for fibrosis, the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk. In connection with the SNP rsl805794, the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk and the major homozygous genotype (CC) has a lower risk.

[0032] The choice of SNP rs9399005 and / or SNP rsl805794 among the about 85 million SNPs in a human genome as the two best SNPs for the risk assessment according to the method of the invention is the result of a pilot study with Mass Array SNP analysis and the analysis of the three most promising candidate SNPs (noting that one candidate failed in further experiments). A statistically significant association was found between rs9399005 and fibrosis (p=0.044) (Figure 1). Further analysis showed that homozygosity for the minor allele (TT) was associated with a markedly reduced risk of fibrosis (OR= 0.23, p=0.043; Table 2). By contrast, the rate of fibrosis in major homozygotes was close to the mean value for the whole cohort whereas it appeared somewhat higher in the heterozygotes but the difference was not significant (p=0.10). Taken together, this shows that the minor allele is protective in a recessive fashion.

[0033] While the association between rsl805794 and fibrosis initially yielded only a non-significant trend (p=0.106) in the Pearson x2 test (Figure 2a), Fisher's exact test showed a moderately reduced risk for major homozygotes compared with the increased risk of the two other genotypes (OR= 0.55, p=0.047; Table 2). Furthermore, the OR appeared to be higher for minor homozygotes than for heterozygotes but the MAF was lower than expected from the 1000 Genome project and the group of minor homozygotes was too small for a difference to be excluded. However, logistic regression confirmed a significant positive correlation between the number of minor alleles and increased risk of fibrosis (p=0.035; Figure 2b). Therefore, minor homozygotes were combined with heterozygotes into a single group (GG, CG) versus major homozygotes (CC) in predictive modeling.

[0034] It follows that both SNPs rs9399005 and rsl805794 as single SNPs and a fortiori in combination are biomarkers indicating the risk of a cancer patient for developing skin fibrosis after radiation therapy and optionally after cancer surgery. These genotypes are highly relevant to assess this risk.

[0035] In accordance with a preferred embodiment of the first aspect of the invention, the method further comprises (c) performing a radiation-induced lymphocyte apoptosis (RILA) assay on a lymphocyte population of the cancer patient, wherein RILA is the percentage of lymphocyte apoptosis induced by a certain radiation dose or a DNA damaging agent minus the spontaneous lymphocyte apoptosis, and is preferably the percentage of lymphocyte cell death induced by 8 Gy minus the percentage of apoptosis at 0 Gy, and wherein a cancer patient having a high RILA value has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having a low RILA value has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

[0036] The radiation-induced lymphocyte apoptosis (RILA) assay was developed by Ozsahin et al.; Menz et.; Ordonez et al. (31-33). RILA can be defined as the percentage of peripheral blood lymphocyte (PBL) apoptosis induced by a certain radiation dose or a DNA damaging agent minus the spontaneous apoptosis.

[0037] Among the options of a radiation dose and a DNA damaging agent the radiation dose is preferred. The radiation is preferably ionizing radiation. The DNA damaging agent is preferably a cytostatic agent and most preferably Bleomycin. Bleomycin has CAS No. 11056-06-7 and the formula C55H84N17O21S3.

[0038] RILA is a versatile, rapid and reproducible assay of radiosensitivity on PBL. Lymphocytes, in particular, appear to be the ideal candidates for such analyses, owing to both their accessibility and high concentrations in the peripheral blood (34). Lymphocytes have several advantages compared with fibroblasts, they are highly radiosensitive and easy to handle, and results are obtained quickly as lymphocytes do not need to be cultured. Thus RILA is non-invasive and potentially a fast and cheap functional method for predicting fibrosis (35).

[0039] The lymphocytes are or comprise preferably T cells, B cells or NK cells, more preferably CD4+and / or CD8+T cells, and most preferably CD4+T cells. In the most preferred case, the RILA can also be referred to as CD4+RILA.

[0040] A RILA assay may comprise diluting peripheral blood from a subject in a culture medium and incubating the same for 24 hours. Thereafter, a single dose of radiation is delivered to the sample using a linear accelerator. After irradiation, the sample is immediately incubated at 37°C. The sample is left for a period of time and then prepared for measurement of the apoptotic fraction by flow cytometry. RILA is determined as the percentage of lymphocyte apoptosis in the non-irradiated sample subtracted from the percentage of lymphocyte apoptosis in the irradiated sample (preferably about 8 Gy).

[0041] Because RILA is the percentage of lymphocyte apoptosis induced by a certain radiation dose or a DNA damaging agent minus the spontaneous lymphocyte apoptosis, RILA can also be determined, for example, as the percentage apoptosis, as a function of the radiation dose, D, or a transformation of D, e.g. log(D) with D = 0, 1, 2, and 8 Gy. The slope of percentage apoptosis as a function of log(D) is potentially a more accurate way to quantify RILA as compared to just using 8 Gy as outlined above.

[0042] Threshold values for low and high CD4+RILA are preferably in the range between 5% and 12% and more preferably is defined by the machine learning algorithm (see below). CD8+RILA shows a positive correlation with CD4+RILA (Spearman correlation p=0.79, p<0.001) but CD8+RILA values are significantly higher. Threshold values for low and high CD8+RILA are preferably 1.5 to 2-fold higher than the range of CD4+RILA threshold values and more preferably is defined by the machine learning algorithm (see below).

[0043] It is demonstrated in the appended examples that homozygosity for the minor allele of rs9399005 was associated with low risk within the low-RILA group, which by itself was associated with increased risk. Notably, the TT genotype was protective against fibrosis in spite of the low RILA values which predicts an increased risk of fibrosis, implying that the effects of this genotype on fibrosis and RILA are independent. It follows that the genotyping of the SNPs with RILA further improved the risk assessment by the method of the invention.

[0044] In accordance with a further preferred embodiment of the first aspect of the invention, the method further comprises (d) determining the body mass index (BMI) of the cancer patient, wherein a cancer patient having a low BMI has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having a high BMI has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or (e) determining the blood pressure of the cancer patient, wherein a cancer patient having no hypertension has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having hypertension has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

[0045] Body mass index (BMI) is a value derived from the mass (weight) and height of a person. The BMI is defined as the body mass divided by the square of the body height, and is expressed in units of kg / m2, resulting from mass in kilograms (kg) and height in metres (m).

[0046] The cut-off between a low BMI and a high BMI is preferably 24.5. A high BMI is with increasing higher preference a BMI of 28 or above, 29 or above, 30 or above, 31 or above and 32 or above. A low BMI is with increasing higher preference a BMI of 23 or below, 22 or below, 21 or below and 20 or below.

[0047] Hypertension (high blood pressure) is when the pressure in a person's blood vessels is too high.

[0048] Hypertension is preferably defined as a blood pressure of 140 / 90 mmHg or higher. As can be taken from the appended examples a number of clinical parameters were evaluated and only BMI and hypertension showed significant associations with fibrosis and were included in predictive modeling. Although patients with hypertension had significantly higher BMI scores (p<0.0001), the two-way ANOVA analysis showed independent effects of the two parameters on fibrosis (BMI p=0.009; Hypertension p=0.018; BMI x Hypertension p=0.21). Hence, the BMI and hypertension can further improve the risk assessment by the method of the invention, independently from each other a fortiori together.

[0049] In summary, two SNPs, RILA and two clinical parameters showed significant associations with skin fibrosis. In particular, the combination of these five distinct biomarkers provides an outstandingly well- suited method for identifying those cancer patients who develop skin fibrosis after radiation therapy and optional cancer surgery.

[0050] In accordance with a yet further preferred embodiment of the first aspect the method further comprises (f) determining or having determined two or more of the genotype of the SNP rs9399005, the genotype of the SNP rsl805794 or a SNP of Table 1, RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for two or more of the genotype of the SNP rs9399005, the genotype of the SNP rsl805794 or a SNP of Table 1, RILA value, the BMI and the blood pressure in the population of samples; and (f') subjecting or having subjected the data sets of (f) to one or more machine learning models in order to establish a predictive model within the two more data sets that indicates the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, and (f") employing the predictive model of (f') in predicting the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery in the cancer patient for whom the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery is to be determined.

[0051] In accordance with a more preferred embodiment of the first aspect the two or more of (f) comprise the genotype of the SNP rs9399005 and / or the genotype of the SNP rsl805794 or a SNP of Table 1.

[0052] The cancer subjects on which the population of samples is based have preferably the same kind of cancer (e.g. breast cancer, in particular female breast cancer) as the cancer subject to the risk of which is to be determined by the method of the invention. The data sets for two or more of the genotype of the SNP rs9399005, the genotype of the SNP rsl805794 or a SNP of Table 1, RILA value, the BMI and the blood pressure in the population of samples comprise the individual values for the two or more of the five possible parameters for sample / cancer patient of the population.

[0053] The options "having determined" and "having subjected" indicate that these method steps are not carried out in the context of the claimed method but were carried out before. In this connection, it is of note that once a predictive model has been obtained this model can be repeatedly used in connection with the method of the invention. For this purpose, the predictive model can be stored on a storage medium, e.g. a computer, hard disk, CD-ROM, cloud, server or a USB stick.

[0054] A machine learning model is a program that can find patterns or make decisions from a previously unseen dataset. For example, in natural language processing, machine learning models can parse and correctly recognize the intent behind previously unheard sentences or combinations of words. In image recognition, a machine learning model can be taught to recognize objects - such as cars or dogs. A machine learning model can perform such tasks by having it 'trained' with a large dataset. During training, the machine learning algorithm is optimized to find certain patterns or outputs from the dataset, depending on the task. The output of this process - often a computer program with specific rules and data structures - is called a machine learning model. Machine learning models use machine learning algorithms. A machine learning algorithm is a mathematical method to find patterns in a set of data. Machine Learning algorithms are often drawn from statistics, calculus, and linear algebra.

[0055] Predictive modelling generally uses statistics to predict outcomes. Most often the event one wants to predict is in the future, but predictive modelling can be applied to any type of unknown event, regardless of when it occurred. For example, predictive models are often used to detect crimes and identify suspects, after the crime has taken place. In the present case, the future skin fibrosis in accordance with the method of the invention is to be assessed by help of the predictive model.

[0056] Predictive modeling is often performed using curve and surface fitting, time series regression, or machine learning approaches. Regardless of the approach used, the process of creating a predictive model is basically the same across methods. The steps may be one or more, preferably all of:

[0057] 1) Clean the data by removing outliers and treating missing data

[0058] 2) Identify a regression or classification predictive modeling approach to be used

[0059] 3) Preprocess the data into a form suitable for the chosen modeling algorithm 4) Specify a subset of the data to be used for training the model

[0060] 5) Train, or estimate, model parameters from the training data set

[0061] 6) Conduct model performance or goodness-of-fit tests to check model adequacy

[0062] 7) Validate predictive modeling accuracy on data not used for calibrating the model

[0063] 8) Use the model for prediction if satisfied with its performance

[0064] The present invention relates in a second aspect to a method for establishing a predictive model that indicates the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery in a test cancer patient, wherein the method comprises (a) determining the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the following: RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the RILA value, the BMI and the blood pressure in the population of samples; and (b) subjecting the data sets of (a) to one or more machine learning models in order to establish the predictive model within the two or more data sets that indicates the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery.

[0065] The above definitions and preferred embodiments of the first aspect of the invention apply mutatis mutandis to the second aspect of the invention as far as being amenable with the second aspect.

[0066] As explained above, once a predictive model has been obtained this model can be repeatedly used in connection with the method of the first aspect of the invention. For this purpose, the predictive model can be stored on a computer, system or storage medium (e.g. a hard disk, CD-ROM, cloud, server or a USB stick).

[0067] It is therefore also not necessary to establish the predictive model in direct connection with the method of the first aspect of the invention. For this reason, the second aspect of the invention is directed to a method for establishing a predictive model that indicates the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery in a test cancer patient that can then be used again and again in connection with the first aspect of the invention.

[0068] The technical advantage of such a predictive model is that the predictive qualities of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the RILA value, the BMI and the blood pressure in the population of samples are optimal weights and compares with each other in a fully automated manner, so that when the predictive model is applied in connection with the first aspect of the invention the risk assessment can be done fast and with optimized accuracy.

[0069] The present invention relates in a related third aspect to a computer, system or storage medium being configured for predicting the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the computer, system or storage medium stores a predictive model that indicates the risk for developing skin fibrosis in a test cancer patient after radiation therapy and optionally after cancer surgery, wherein the predictive model has been obtained by a method comprising (a) determining the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1, and optionally at least one of the following: RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the following: RILA value, the BMI and the blood pressure in the population of samples; and (b) subjecting the data sets of (a) to one or more machine learning models in order to establish a predictive model within the two more data sets that indicates the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery.

[0070] The above definitions and preferred embodiments of the first and second aspects of the invention apply mutatis mutandis to the third aspect of the invention as far as being amenable with the third aspect.

[0071] As discussed above, until its use the predictive model can be stored on a computer, system or storage medium. The computer, system or storage medium can be configured for predicting the risk of developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery and used in the method of the fist aspect of the invention.

[0072] In accordance with a preferred embodiment of the first to third aspect of the invention one or more machine learning models comprise an ensemble machine learning algorithm and / or a support vector machine (SVM) algorithm and / or neural network machine learning algorithms.

[0073] In statistics and machine learning, ensemble methods use multiple learning algorithms to obtain better predictive performance than could be obtained from any of the constituent learning algorithms alone. Unlike a statistical ensemble in statistical mechanics, which is usually infinite, a machine learning ensemble consists of only a concrete finite set of alternative models, but typically allows for much more flexible structure to exist among those alternatives.

[0074] In machine learning, support vector machines (SVMs) are supervised learning models with associated learning algorithms that analyze data for classification and regression analysis. SVMs are one of the most robust prediction methods, being based on statistical learning frameworks or VC theory proposed by Vapnik and Chervonenkis (so-called Vapnik-Chervonenkis or VC theory). Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that assigns new examples to one category or the other, making it a non-probabilistic binary linear classifier (although methods such as Platt scaling exist to use SVM in a probabilistic classification setting). SVM maps training examples to points in space so as to maximise the width of the gap between the two categories. New examples are then mapped into that same space and predicted to belong to a category based on which side of the gap they fall. In addition to performing linear classification, SVMs can efficiently perform a non-linear classification using what is called the kernel trick, implicitly mapping their inputs into high-dimensional feature spaces.

[0075] Neural networks in machine learning refer to a set of algorithms designed to help machines recognize patterns without being explicitly programmed. They consist of a group of interconnected nodes. These nodes represent the neurons of the biological brain. The basic neural network consists of: The input layer, the hidden layers and the output layer. Neural networks in machine learning use mathematical or computational models to process information. These neural networks are usually non-linear, which allows them to model complex relationships between data input and output and find patterns in a dataset. The application of neural networks in machine learning tends to take one of these three broad categories: a) Classification whereby a neural network can recognize patterns and sequences; b) Functional approximation and regression analysis, and c) Data processing including clustering and filtering data.

[0076] In accordance with a more preferred embodiment of the first to the third aspect of the invention, ensemble machine learning algorithm is random forest (RF), which is preferably based on creating a number of decision trees on subgroups of the data sets, and the predictive model is generated based on votes of majority of the trees.

[0077] The random forest is an ensemble machine learning algorithm consisting of many decision trees. It can be used for classification, regression and other tasks that operate by constructing a multitude of decision trees at training time. For classification tasks, the output of the random forest is the class selected by most trees. For regression tasks, the mean or average prediction of the individual trees is returned. Random decision forests correct for decision trees' habit of overfitting to their training set. Forests generally outperform decision trees, but their accuracy is often lower than gradient boosted trees. However, data characteristics can affect their performance. The appliance of random forest is illustrated in the appended examples.

[0078] In accordance with another more preferred embodiment of the first to third aspect of the invention, the SVM algorithm is developed with the Radial Basis Function (RBF) kernel function, and the predictive model is generated based on the differences of inner outliers, or samples that are most similar to the opposite class, preferably fibrosis negative samples that are similar to fibrosis positive samples vs. fibrosis positive samples that are similar to fibrosis negative samples.

[0079] The application of the SVM algorithm developed with the RBF kernel function is illustrated by the appended examples. The RBF kernel works by mapping the input data to a higher dimensional space and then computing the similarity between them using the kernel function. The mapping is done by applying a non-linear transformation to the input data, which makes the data separable in the higher dimensional space.

[0080] Preferably the machine learning algorithm in this invention is one of the following: ensemble models, SVM, neural networks models. More preferably is the combination of ensemble models and SVM and most preferably it is the combination of RF and SVM, wherein SVM defines low-risk groups and RF separates residual samples into intermediate risk-group and high-risk group.

[0081] The SVM algorithm is preferably trained using the following values for the hyperparameters: Regularization parameter (C) = 20;

[0082] Gamma = 0.05;

[0083] Class weights - balanced (automatically balance according to the ratio between the classes); Kernel - Radial Basis Function (RBF).

[0084] The RF algorithm is preferably trained using the following values for the hyperparameters:

[0085] Max depth (of the trees) = 5;

[0086] Max leaf nodes = 20;

[0087] Max samples used per tree = 0.9 (of the whole cohort);

[0088] Min samples per leaf = 3; Min samples per split = 8;

[0089] Number of estimators (trees) = 500.

[0090] In accordance with a preferred embodiment of the first to third aspect of the invention the population of samples comprises at least 25 samples, preferably at least 50 samples, more preferably at least 100 samples, and most preferably at least 200 samples.

[0091] More generally the population of samples comprises with increasing preference samples from at least 2, at least 10, at least 25, at least 50, at least 100 and at least 200 samples from cancer patients wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery.

[0092] In this connection, the population of samples additionally preferably comprises with increasing preference samples from at least 1, at least 5, at least 10, at least 25, at least 50, and at least 100 samples from cancer patients that developed a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery and / or at least 1, at least 5, at least 10, at least 25, at least 50, and at least 100 samples from cancer patients that did not develop a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery.

[0093] In accordance with a preferred embodiment of the first to third aspect of the invention, the cancer is breast cancer or head and neck cancer.

[0094] Breast cancer (in particular female breast cancer) is preferred over head and neck cancer. Breast cancer is illustrated by the appended examples. Breast cancer often metastasizes to the head and neck.

[0095] Head and neck cancer develops from tissues in the lip and oral cavity (mouth), larynx (throat), salivary glands, nose, sinuses, or skin of the face. The most common types of head and neck cancer occur in the lips, mouth, and larynx.

[0096] In accordance with a further preferred embodiment of the first to third aspect of the invention, the skin fibrosis is subcutaneous fibrosis.

[0097] Radiation-induced skin fibrosis in particular occurs subcutaneously; i.e. just under the skin.

[0098] In accordance with a preferred embodiment of the first to the third aspect of the invention the sample is a blood sample, preferably whole blood, plasma or serum. Blood samples were used in the examples. For the case RILA is performed, it is to be understood that the samples need to comprise lymphocytes (such as CD4+T cells). Hence, for example, plasma and serum cannot be used for RILA but for SNP genotyping.

[0099] In accordance with a preferred embodiment of the first to the third aspect of the invention the sample of the (test) patient and / or the samples in the population of samples has / have been obtained before, during or after radiation therapy and optionally before, during, or after cancer surgery.

[0100] It is to be understood that the timing of sampling for the SNP genotyping is of no matter since the SNP genotype is an inherent genomic property of each individual. Hence, the samples can be obtained before, during or after radiation therapy and optionally before, during, or after cancer surgery.

[0101] The present invention relates in a fourth aspect to a kit for predicting the risk of developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the kit comprises means for (a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for radiation- induced skin fibrosis, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for radiation-induced skin fibrosis, and / or (b) genotyping the SNP rsl805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for radiation-induced skin fibrosis a cancer, and a cancer patient having the major homozygous genotype (CC) has a lower risk for radiation-induced skin fibrosis.

[0102] The above definitions and preferred embodiments of the first to third aspect of the invention apply mutatis mutandis to the fourth aspect of the invention as far as being amenable with the third aspect.

[0103] In accordance with a preferred embodiment of the fourth aspect of the invention, the means comprise one or more hybridization oligonucleotide probe(s) that can discriminate(s) between the minor and the major allele of one or both of SNPs of (a) and (b), wherein the hybridization oligonucleotide probe(s) is / are preferably arrayed on a small chip, or are molecular beacon(s).

[0104] As mentioned above, SNP genotyping is the measurement of genetic variations of single nucleotide polymorphisms (SNPs), whereby the SNP is in accordance with the present invention SNP rs9399005 and / or SNP rsl805794 or a SNP of Table 1. Means for SNP genotyping are available in the prior art and these means can be packed into a kit and such kit is covered by the fifth aspect of the invention.

[0105] The means are one or more hybridization oligonucleotide probe(s) that can discriminate(s) between the minor and the major allele of one or both of SNPs of (a) and (b). The hybridization oligonucleotide probe(s) can be molecular beacons. A molecular beacon is a single-stranded bi-labeled fluorescent probe held in a hairpin-loop conformation (around 20 to 25 nt) by complementary stem sequences (around 4 to 6 nt) at both ends of the probe. The 5' and 3' ends of the probe contain a reporter and a quencher molecule, respectively. The loop is a single-stranded DNA sequence complementary to the target sequence. The proximity of the reporter and quencher causes the quenching of the natural fluorescence emission of the reporter. The structure and mechanism of a molecular beacon is shown below.

[0106] The hybridization oligonucleotide probe(s) can also be on a small chip and such chips are known as SNP microarrays. The basic principles of SNP microarray are the same as the DNA microarray. These are the convergence of DNA hybridization, fluorescence microscopy, and solid surface DNA capture. The three mandatory components of the SNP arrays are: A) An array containing immobilized allelespecific oligonucleotide (ASO) probes, B) Fragmented nucleic acid sequences of a target, labelled with fluorescent dyes, and C) A detection system that records and interprets the hybridization signal.

[0107] The present invention relates in a fifth aspect to a method of preventing radiation-induced skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the method comprises identifying the cancer patient as being at risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery by the method of the first aspect of the invention, and administering to the cancer patient an agent for preventing the skin fibrosis.

[0108] In accordance with a preferred embodiment of the fifth aspect of the invention the agent for preventing radiation-induced skin fibrosis is selected from imatinib, silymarin, pentoxifylline optionally with alpha-tocopherol (vitamin E), corticosteroid optionally with dexpanthenol, small-molecule inhibitors, microRNAs and antibodies against CTGF or the minor allele of NBS1.

[0109] The above definitions and preferred embodiments of the first to fourth aspect of the invention apply mutatis mutandis to the fifth aspect of the invention as far as being amenable with the fifth aspect. Once a cancer patient has been identified as being at risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery by the method of the first aspect of the invention, of course countermeasures should be taken in this patient in order to prevent the skin fibrosis. Such countermeasures are the subject of the sixth aspect of the invention.

[0110] A therapeutic countermeasure might be to replace whole-breast radiation therapy (RT) with partial breast RT (e.g. post-surgical external beam RT, brachytherapy or intraoperative RT).

[0111] In the prior art means and methods for the prevention and treatment of radiation-induced skin fibrosis are known. Preferred examples are imatinib, silymarin, pentoxifylline optionally with alpha-tocopherol (vitamin E), corticosteroid optionally with dexpanthenol.

[0112] Also, small-molecule inhibitors, microRNAs and antibodies against CTGF or the minor allele of NBS1 can be used based on the finding that the SNPs rs9399005 and rsl805794 (or a SNP of Table 1) are associated with the risk of radiation-induced skin fibrosis, noting again that these SNPs can be found at the genomic loci of the CTGF and the NBS1 gene.

[0113] As regards the embodiments characterized in this specification, in particular in the claims, it is intended that each embodiment mentioned in a dependent claim is combined with each embodiment of each claim (independent or dependent) said dependent claim depends from. For example, in case of an independent claim 1 reciting 3 alternatives A, B and C, a dependent claim 2 reciting 3 alternatives D, E and F and a claim 3 depending from claims 1 and 2 and reciting 3 alternatives G, H and I, it is to be understood that the specification unambiguously discloses embodiments corresponding to combinations A, D, G; A, D, H; A, D, I; A, E, G; A, E, H; A, E, I; A, F, G; A, F, H; A, F, I; B, D, G; B, D, H; B, D, I; B, E, G; B, E, H; B, E, I; B, F, G; B, F, H; B, F, I; C, D, G; C, D, H; C, D, I; C, E, G; C, E, H; C, E, I; C, F, G; C, F, H; C, F, I, unless specifically mentioned otherwise.

[0114] Similarly, and also in those cases where independent and / or dependent claims do not recite alternatives, it is understood that if dependent claims refer back to a plurality of preceding claims, any combination of subject-matter covered thereby is considered to be explicitly disclosed. For example, in case of an independent claim 1, a dependent claim 2 referring back to claim 1, and a dependent claim 3 referring back to both claims 2 and 1, it follows that the combination of the subject-matter of claims 3 and 1 is clearly and unambiguously disclosed as is the combination of the subject-matter of claims 3, 2 and 1. In case a further dependent claim 4 is present which refers to any one of claims 1 to 3, it follows that the combination of the subject-matter of claims 4 and 1, of claims 4, 2 and 1, of claims

[0115] 4, 3 and 1, as well as of claims 4, 3, 2 and 1 is clearly and unambiguously disclosed.

[0116] The figures show.

[0117] Figure 1: The association between rs9399005 (CTGF) and fibrosis development (p=0.044)

[0118] Figure 2: a) The association between rsl805794 (NBS1) and fibrosis development (p=0.106); b) Logistic regression analysis showed the positive correlation between the number of minor alleles for rsl805794 and the probability for fibrosis development (p=0.035).

[0119] Figure 3: Power analysis showing the Odds Ratios (OR) that should be detectable with a power of 80% at a=0.05 as a function of the MAF, according to the additive genetic model.

[0120] Figure 4: Schematic representation of development of the RF model based on LOOCV. After splitting and scale-standardizing the dataset, the RF algorithm creates 500 random subgroups from the training dataset with sample sizes from 10 to 218 and performs the decision tree analysis on each of them. By taking majority votes from the individual trees, it predicts and classifies the test sample (n=l) as fibrosis positive or negative.

[0121] Figure 5: Schematic representation of development SVM model based on LOOCV. After splitting and data scale standardization, SVM algorithm uses the Radial Basis Function kernel trick to calculate the distances between the samples as they would be in a multidimensional coordinate system and consider only inner outliers to separate two classes (fibrosis positive and negative). Based on that, the model predicts the fibrosis status and classifies the test sample (n=l).

[0122] Figure 6: The association between rs9399005 (CTGF) genotype and CD4+RILA levels.

[0123] Figure 7: The CDF plots show the distributions of CD4+RILA values in regard to the rs9399005 (CTGF) genotype.

[0124] Figure 8: The association between rsl805794 (NBS1) genotype and CD4+RILA levels.

[0125] Figure 9: The CDF plots show the distributions of CD4+RILA values in regards to the rsl805794 (NBS1) genotyp. Figure 10: The distribution of fibrosis positive and negative samples after partition analysis (PA).

[0126] Figure 11: Decision Tree (DT) for predictive partition analysis (PA). The first split was according to a CD4+RILA with lower values (< 7.48; n=65) associated with increased risk (50.1% fibrosis) and higher values (> 7.48; n=154) with relative resistance (21.4% fibrosis). The second split separated the low- RILA group based on the rs9399005 (CTGF) genotype in a high-risk group (CC+CT; n = 56) with 57.1% fibrosis and a highly resistant subgroup (TT; n = 9) with 11.1% fibrosis in spite of being in the low-RILA group associated with higher risk. The third split separated the high-risk group according to BMI where 75.8% of the patients with BMI > 24.4 kg / m2 (n=33) developed fibrosis whereas the rate was only 30.4% among patients with BMI < 24.4 kg / m2 (n=23). The fourth split divided the high-risk group based on the rsl805794 (NBS1) genotype into a very high-risk group of heterozygous and minor homozygous genotypes (CG+GG) with 87% fibrosis (n=24) against 44% risk in major homozygotes (CC; n=9). Notably, this also indicated that the CC genotype was under-represented in the low-RILA / high-BMI group. Finally, the fifth split divided the high-RILA group based on hypertension status at diagnostic, where 35% of patients with hypertension (n=39) developed fibrosis but only 16.5% of patients with normal blood pressure. The features contributed as follows: RILA CD4 (36.6%), BMI (23.8%), Hypertension status (17.3%), CTGF minor TT (14.9%), and NBS1 major CC (12.4%). The low-risk and high-risk subgroups are indicated. The vertical broken line separates risk groups according to RILA CD4+.

[0127] Figure 12: Receiver Operating Characteristic (ROC) analysis for RF and SVM models.

[0128] Figure 13: Cumulative Accuracy Profile (CAP) analysis for RF and SVM.

[0129] Figure 14: Evaluation of the feature importance for RF and SVM model based on the Permutation feature importance method.

[0130] Figure 15: Permutation significance test, a) The actual accuracy for the RF model is 0.78 (on the original dataset), while the mean accuracy for the permuted datasets was 0.67±0.02 (p<0.001). b) The accuracy for the SVM model is 0.74 while the mean accuracy for the permuted model was 0.51±0.07 (p<0.001).

[0131] Figure 16: Stratified 10-fold cross-validation. The mean error rate for a) RF model was 0.17±0.01 for the training set and 0.24±0.07 for the test set. The mean error rate for b) SVM model was 0.25±0.015 for the training set and 0.30±0.09 for the test set. Figure 17: Risk group classification based on Normal Mixture Clustering. The SVM model separates low-risk group while the RF model divides the rest of the samples into intermediate and high-risk groups.

[0132] Figure 18: The association between risk groups and the fibrosis status.

[0133] Figure 19: The association between combined machine learning (ML: RF+SVM) and partition analysis (PA) risk groups.

[0134] Figure 20: ROC analysis for combined machine learning (ML: RF+SVM) and partition analysis (PA) risk groups.

[0135] Figure 21: Association of the SNPs with fibrosis development in each of the subgroups RILA-high (>7.48; reduced risk) and RILA-low (<7.48; increased risk). Upper row: rs9399005 (CTGF), lower row: rsl805794 (NBS1).

[0136] Figure 22: Association of the clinical features with fibrosis development in each of the subgroups RILA- high (reduced risk) and RILA-low (increased risk). Upper row: BMI, lower row: hypertension.

[0137] Figure 23: Association of the SNPs with fibrosis development in each of the subgroups according to BMI-low (<24.4; reduced risk) and BMI-high (>24.4; increased risk). Upper row: rs9399005 (CTGF), lower row: rsl805794 (NBS1).

[0138] Figure 24: Distribution of probabilities for the number of fibrosis-positive samples calculated for each risk group (high risk (a), low risk (b) and intermediate risk (c)) using the hypergeometric distribution formula and the frequency of fibrosis for the entire cohort.

[0139] Figure 25: In-silico analysis of transcription factor binding profiles using the JASPAR database (Rauluseviciute et al., Nucl. Acids Res 52(D1):D174-D182, 2023). a) cumulative distribution of relative binding scores of transcription factors to the rs9399005 C- and T-alleles; b) distribution of the difference in relative binding scores between the rs9399005 T- and C-alleles for individual transcription factors. Figure 26: Effect of the rs9399005 T-allele on gene expression of CTGF in fibroblasts, a) Normalized CTGF gene expression levels as function of the number of minor alleles (T): 0=CC, 1=CT, 2=TT. Changes were similar at baseline (dO), and after two days (d2) incubation of unirradiated (0 Gy and irradiated (4 Gy) cells; b) Univariate analysis of the association between the irradiation status and CTGF expression levels in the different genotypes, (c) Number of differentially expressed genes (DEGs; false discovery rate: FDR = 0.05) relative to baseline levels (0 Gy, day 0) in fibroblasts with CC genotype.

[0140] The examples illustrate the invention.

[0141] Example 1 - Materials and Method

[0142] Patients, treatments and endpoints. All patients belonged to the German ISE cohort recruited prospectively during 1998-2001 from the Rhine-Neckar region in Germany (Women's Clinic Heidelberg, St. Vincentius Clinic Karlsruhe, City Hospital Karlsruhe, and Universitatsmedizin Mannheim). The patients had been treated with breast conserving surgery (BCS) followed by RT, but not chemotherapy. Adjuvant RT was given to the whole breast (50.0 Gy in 2.0 Gy / fraction or 50.4 Gy in 1.8 Gy / fraction) followed by a tumor bed boost of 6 to 16 Gy (median: 10 Gy, 2.0 Gy / fx; n = 83), or with 56.0 Gy given in 2.0 Gy fractions without a boost. Blood samples were collected with informed consent at late followup and normal-tissue reaction was scored as part of an observational study (22) on the association between radiation-induced lymphocyte apoptosis (RILA) and late adverse effects (breast fibrosis). The median follow-up was 11.6 years [10.3-12.8 years] with a mean age of the patients of 59 years at the diagnosis, and fibrosis was scored according to the LENT-SOMA scoring system (36,37). Patient scores were dichotomized with scores 0-1 considered fibrosis negative, and patients with scores 2-3 fibrosis positive. Genomic DNA from n=238 patients was isolated from peripheral blood monocytic cells (PBMC) in the original blood samples.

[0143] Candidate SNP selection and genotyping. Based on data from a pilot study with MassArray SNP analysis, three candidate SNPs, rs9399005 in the CTGF gene, rsl805794 in NBS1, and rs373759 in ATM, were chosen for genotyping in the ISE cohort (n=238). A power analysis of odds ratios as a function of the MAF indicated that associations with OR=1.75-1.85 could be detected with a power of 80% for a=0.05 assuming a 30% fibrosis rate and an additive effect of individual alleles (Figure 3). A power analysis for additive genetic models was performed in R, using the Power Genetics library (38) Genotyping was performed by allele discrimination with real-time quantitative polymerase chain reaction (qPCR) on the 238 samples for which genomic DNA was available. Statistical analysis

[0144] Significant differences between the three genotypes were tested with Pearson's x2 test while Fisher's Exact Test (2-Tail) was used for testing the associations between two categorical parameters. For nonparametric testing of continuous parameters between two groups the Mann Whitney U test was used, and for comparing more than two groups Kruskal-Wallis test was used.

[0145] Screening predictive modeling was done with partition analysis (PA) which creates a decision tree (DT) by a number of consecutive dichotomous splits. All statistical analyses were performed with JMP statistical software (SAS, Boblingen, Germany).

[0146] Machine learning (ML) models.

[0147] The development of Random Forest (RF) and Support Vector Machine (SVM) classifiers is presented schematically in figure 4 for RF and figure 5 for SVM. The RF algorithm is based on creating a number of decision trees on subgroups of a data set, and making the final prediction by taking votes of majority of the trees. The SVM model was developed with Radial Basis Function (RBF) kernel, classification is based on the differences of inner outliers, or samples that are most similar to the opposite class (e.g. fibrosis negative samples that are similar to fibrosis positive samples vs fibrosis positive samples that are similar to fibrosis negative samples).

[0148] Because of the limited number of samples available, the Leave-One-Out Cross Validation (LOOCV) approach was used instead of the standard method where the cohort is split into training and test datasets. The LOOCV implies using only one sample at a time as a test subject (test dataset, n=l sample), while all other samples are used for training the model (training dataset, n=218 samples). This means that predictions were estimated for each sample, based on a model trained on all other residual samples. In practice, the prediction of fibrosis status was made on every single sample separately, without data leakage from the training dataset to the test dataset. The model evaluation and performance analysis were calculated for the test set composed of single predictions for each sample. The characteristics of the models were evaluated by the cumulative accuracy profile (CAP) analysis and accuracy. CAP analysis plots the fraction of positive observations predicted by the classifier on the y- axis as a function of the fraction of total (positive and negative) observations on the x-axis. For a perfect model, the y-values increase linearly from 0 to 1 over the range of x-values from 0 to the overall rate of positive observations among all observations. The Accuracy Rate (AR) is defined as the ratio of the area under the curves for the actual model and the perfect model.

[0149] Feature importance was evaluated using the Permutation Feature Importance technique. This implies shuffling values in each feature a number of times, everytime repeating the whole model construction process and measuring the accuracy of the model. Finally, the importance of each feature is presented as an average decrease of model accuracy after shuffling the values for that feature.

[0150] To evaluate the statistical significance of the final models, permutation tests were used (39). The permutation test is performed by shuffling the dataset a number of times (e.g. 1000 times), each time making predictions based on the developed models and calculating accuracies. This way, the distribution of accuracies for randomized datasets was created. The p-value is calculated as the fraction of randomized datasets where the accuracy is higher than in the original dataset.

[0151] As leave-one-out cross-validation (LOOCV) is a low-bias cross-validation method with high variance, additional stratified 10-fold cross-validation was applied and repeated 10 times in order to check the stability of the models. For calculating the error rate with stratified 10-fold cross-validation, the whole dataset was randomly split into 10 subgroups (folds), preserving the ratio of fibrosis positive and fibrosis negative samples (~30% of fibrosis positive samples) in each subgroup. For each subgroup, one was used as a test set using the other nine subgroups to train the model, calculating error rates of predicted fibrosis status in the test and training sets. Thus, using each subgroup as a test set, ten pairs of error rates were obtained. The whole process was repeated 10 times, making ten new subgroups each time, and error rates and standard deviations for training and test datasets were calculated from the 100 pairs of error rates. The data are presented as mean ± standard deviation (SD), unless otherwise noted.

[0152] The final models were compared with other commonly used machine learning models: Logistic Regression (LR: a linear model), Extreme Gradient Boosting (XGB: an ensemble model and alternative to RF), K-Nearest Neighbor (KNN) and Naive Bayes Classifier (NBC). Model performance was evaluated by precision (positive and negative predictive values), accuracy, Receiver Operating Characteristic (ROC) Area Under the Curve (AUC), and Cohen's kappa coefficient. Cohen's kappa coefficient helps to evaluate the reliability of the model in case of significant inequality in the frequency of endpoint values, considering the base frequencies of those values. All machine learning modeling and related analyses were performed in Python programming language.

[0153] Example 2 - Results

[0154] Association between candidate SNPs and late tissue reaction. Hardy-Weinberg equilibrium was confirmed for rs9399005 (CTGF) (p=0.08) and rs373759 (ATM) (p=0.97) but not rsl805794 (NBS1) (p=0.006) (n=238). A statistically significant association was found between rs9399005 and fibrosis (p=0.044) (Figure 1). Further analysis showed that homozygosity for the minor allele (TT) was associated with a markedly reduced risk of fibrosis (OR= 0.23, p=0.043; Table 2). By contrast, the rate of fibrosis in major homozygotes was close to the mean value for the whole cohort whereas it appeared somewhat higher in the heterozygotes but the difference was not significant (p=0.10). Taken together, this suggests that the minor allele was protective in a recessive fashion.

[0155] Although the association between rsl805794 (NBS1) and fibrosis yielded only a non-significant trend (p=0.106) in the Pearson x2 test (Figure 2a), Fisher's exact test showed a moderately reduced risk for major homozygotes (OR= 0.55, p=0.047; Table 2). Furthermore, the OR appeared to be higher for minor homozygotes than for heterozygotes but the MAF was lower than expected from the 1000 genomes project and the group of minor homozygotes was too small for a difference to be excluded. Indeed, logistic regression confirmed a significant correlation between the number of minor alleles and increased risk of fibrosis (p=0.035; Figure 2b). Therefore, minor homozygotes were combined with heterozygotes into a single group (GG, CG) versus major homozygotes (CC) in predictive modeling.

[0156] No significant association between rs373759 (ATM) and fibrosis was detected with Pearson's x2 test (p=0.412) nor with Fisher's exact test (Table 2).

[0157] Table 2. The association between fibrosis development and SNP genotypes in the ISE cohort, p-values (Fisher's Exact Test, 2-tail) for each genotype against the two others and odds ratios (OR) with confidence intervals (Cl) are shown. Evaluation of clinical parameters. Among all the patients, 30% developed fibrosis grade 2-3 (i.e. fibrosis positive). Only body mass index (BM I) and hypertension showed significant associations with fibrosis and were included in predictive modeling (Table 3). Although patients with hypertension had significantly higher BMI scores (p<0.0001), the two-way ANOVA analysis showed independent effects of the two parameters on fibrosis (BMI p=0.009; Hypertension p=0.018; BMI x Hypertension p=0.21).

[0158] Table 3. The association between clinical parameters and fibrosis. Some patients did not have complete clinical data.

[0159] Association between candidate SNPs and RILA values. Next, the association between the candidate SNPs and CD4+RILA was tested. A statistical trend was detected between rs9399005 (CTGF) and RILA CD4+levels (p=0.086; Figure 6). The CD4+RILA levels in minor homozygotes (TT) were significantly lower than in the major homozygote samples (CC) (p=0.025) whereas a strong trend was observed between minor homozygote genotype (TT) and heterozygotes (CT) (p=0.058). No statistically significant difference was observed between heterozygotes (CT) and major homozygotes (CC) (p=0.86) and this was corroborated by Cumulative Distribution Function (CDF) plots (figure 7). Notably, the TT genotype was protective against fibrosis in spite of the low RILA values, which predicted increased risk of fibrosis, implying that the effects of this genotype on fibrosis and RILA are independent.

[0160] A borderline statistically significant association was found between rsl805794 (NBS1) and CD4+RILA (p=0.05). The minor homozygote genotype (GG) had significantly lower CD4+RILA values compared to major homozygous (CC) samples (p=0.022), while a statistical trend was observed between minor homozygotes (GG) and heterozygous samples (CG), (p=0.098; Figure 8). CDF plots supported an additive effect of the minor allele (figure 9). There was no statistically significant association between rs373759 (ATM) and CD4+RILA values (p=0.80).

[0161] Partition analysis (PA). In order to test the hypothesis of patient subgroups with different risks for adverse late reaction, a predictive PA based on the two significant SNPs (TT vs CT / CC in rs9399005 (CTGF); CG / GG vs CC vs in rsl805794 (NBS1)), CD4+RILA values, and the clinical parameters, BMI and hypertension. To avoid errors caused by imputed missing data, only patients with complete data sets for the five features were included, resulting in n=219 samples.

[0162] The distribution of fibrosis in the subgroups resulting from the first five splits is shown in Figure 10. The corresponding decision tree (DT) including the number of patients in each group is shown in Figure 11. The subgroups identified from the PA could be grouped into three main groups according to their risk: a high-risk, an intermediate-risk, and a low-risk group (Table 4). This analysis is consistent with and supports the hypothesis that some risk factors are important only in certain subgroups defined by other risk factors or combinations thereof.

[0163] Table 4. Categorization and evaluation of the risk groups based on decision tree (DT) from partition analysis (PA).

[0164] Machine learning (ML) modeling. In the absence of a validation cohort with comparable long-term follow-up and available RILA data, the partition model carries a risk of overfitting the data. Therefore, in order to validate the importance of the five features, two different ML models were developed, Random Forest (RF) and Support Vector Machine (SVM). The RF model was more precise in predicting fibrosis positive samples (PPV of 78% for RF compared to 56% for SVM), while SVM was more precise in predicting fibrosis negative patients (NPV of 83% compared to 78% for RF) (Table 5). The AUC from ROC analysis was 0.67 for RF and 0.71 for SVM (Figure 12). Cumulative accuracy profile (CAP) analysis showed a higher CAP score for SVM compared to RF (RF CAP = 0.34, SVM CAP = 0.42) (Figure 13) whereas the accuracy score was higher for RF than for SVM (0.78 and 0.74, respectively). The predictions for RF and SVM are presented as a confusion matrix showing true and false positive and negative predictions for RF and SVM models (Table 6). The performance of the RF and SVM models was compared with common ML models LR, XGB, KNN, and NBC (Table 5). Table 5. Evaluation and comparison of machine learning model performances.

[0165] Table 6. Confusion matrix presenting positive and negative (true and false) predictions for random forest (RF) and support vector machine (SVM) models.

[0166] CAP curve analysis (Figure 13) shows that RF predicted fibrosis better than SVM in approximately 15% of patients whereas SVM was better than RF in predicting absence of fibrosis in approximately 65% of the patients. The importance of the five features showed that, overall, CD4+RILA and BMI were the two most important features with a trend for being more important in RF (Figure 14). Hypertension reached a similar level of importance in SVM but was much less important in RF. Selected CTGF and NBS1 SNPs were similarly important in both RF and SVM models. Notably, the decrease in accuracy was lower for every feature in RF compared to SVM. This might indicate the differences in the stability of the models.

[0167] A permutation significance test with 1000 iterations showed both models to be statistically highly significant with a very low probability that the accuracy values were a chance result (RF: p<0.001, SVM: p<0.001) (Figure 15a for RF, Figure 15b for SVM).

[0168] To evaluate the stability and reliability of the models, stratified 10-fold cross-validation was performed. The mean error rates were 0.17±0.01 (s.d.) for the training data set and 0.24±0.07 in the test set for the RF model (Figure 16a), and 0.25±0.015 and 0.30±0.09, respectively, for the SVM model (Figure 16b). Thus, although RF model was slightly more stable than SVM, the difference in accuracy between training and test datasets was less than 10% in both models.

[0169] A comparison of predictions by the two models (Figure 17 showed that SVM discriminated a low-risk group (n=146, 66.7%) with 17.1% fibrosis (i.e. false negatives). The remaining 33.3% of the patients were divided into a high-risk group predicted to be fibrosis-positive by both models (n=30, 13.7%) with 83.3% fibrosis and a group predicted to be fibrosis-positive by SVM and fibrosis-negative by RF (n=43, 19.6%) showing 37.2% fibrosis (i.e. intermediate risk) (Table 7). The differences in fibrosis risk between the three groups were highly significant (Figure 18) and highly correlated with the risk groups defined by PA (p<0.0001) (Figure 19). Thus, none of the patients were classified as low-risk by one method and high-risk by the other method. The intermediate-risk group in the combined machine learning (ML: RF+SVM) method included very few patients classified as high-risk and only a minority classified as low-risk by the PA method. ROC analysis indicated that the combined ML model was slightly better at discriminating intermediate-risk patients, resulting in a lower proportion of patients in the intermediate-risk group (19.6% vs 32.4%; cf Tables 3 and 6) and a slightly higher AUC value than for the risk groups based on PA (Figure 20).

[0170] The 3-risk group models (combined ML:RF+SVM, and PA, models) were superior to all individual machine learning models. According to the Kappa coefficient and AUC (ROC), RF and SVM showed better performances compared to the other ML models.

[0171] To test the significance of the three risk groups from the combined ML model (RF+SVF) , the distribution of probabilities for the number of fibrosis-positive samples was calculated for each risk group using the hypergeometric distribution formula and the frequency of fibrosis for the entire cohort. The results showed that the high-risk group had a significantly higher number of patients with fibrosis than expected by chance (p= 9.4xl0-11) and that the low risk group had significantly fewer patients with fibrosis (p= 5.9xl0-9). In the intermediate risk group, the number of patients was not significantly different from that in the cohort as a whole (p=l) (Fig. 24). This corroborates the significance of the combined ML model.

[0172] Table 7. Categorization and evaluation of the risk groups based on combined machine learning (ML:RF+SVM). Basically, SVM classifies low risk versus intermediate+high risk and within the latter group, RF distinguishes between intermediate and high risk. Example 3 - the SNPs rs9399005 and 1805794SNP as features in predictive partition analysis (PA) and machine learning (ML) models

[0173] The two SNPs included as features in the predictive partition analysis (PA) and machine learning (ML) models were significant in the RILA-low subgroup (higher risk of fibrosis) of the PA model. However, the analysis did not show the effect in the complementary, CD4+RILA-high (lower risk of fibrosis) subgroup. Therefore, the effects of the two SNPs and the two clinical features in both RILA subgroups were compared.

[0174] The CTGF SNP (rs 9399005) minor homozygous TT genotype showed the same low rate of fibrosis in both RILA subgroups which was significant (p=0.0135) in the RILA-low subgroup with higher 'background' risk compared with the RILA-high subgroup where it was not significant (p=0.46) and not important in the model. Similarly, the NBS1 SNP (rsl805794) was only significant (p=0.04) and important in the RILA-low but not in the RILA-high (p=0.69) subgroup (Figure 21). The clinical feature BMI was significant (p=0.001) in the RILA-low subgroup while still showing a clear trend (p=0.11) in the RILA-high subgroup. Hypertension showed the opposite pattern, being significant (p=0.02) in the RILA- high but not in the RILA-low (p=0.34) subgroup (Figure 22). These results supported the subgroup hypothesis that some risk factors are only important in certain subgroups.

[0175] When BMI was used to form subgroups with higher and lower risk of fibrosis without invoking RILA, the effects of the two SNPs showed the same pattern as for the RILA subgroups: The CTGF SNP was equally protective in both subgroups but only significant(p=0.048) in the BMI-high subgroup (higher risk of fibrosis) (Figure 23). The NBS1 SNP almost reached significance (p=0.06) in the BMI-high subgroup but showed little effect in the complementary BMI-low subgroup. In fact, further analysis showed that the NBS1 SNP was only significant and important when RILA and BMI both indicated high risk. Thus, the protective effect of the CTGF TT genotype with fibrosis was the same in different subgroups as in the whole cohort while the effect of the NBS1 SNP was seen only in higher-risk subgroups in spite of the smaller group sizes compared with the whole cohort.

[0176] This strongly supports the robustness of the associations with fibrosis and the subgroup hypothesis while also emphasizing that each risk factor may behave in a specific way which is probably related to its function. Example 4 - Discussion

[0177] Development and characteristics of the predictive model

[0178] Prediction of patients' risk of adverse late reactions in healthy tissue after radiation therapy is a prerequisite for personalized treatment to improve the quality of life for long-term survivors. In this invention, a novel method for dividing patients into three risk groups with high, intermediate and low risk of developing subcutaneous fibrosis after breast-conserving and adjuvant radiation therapy was developed. Based on just five features (genomic, functional and clinical), the model was able to discriminate patients with up to a five-fold difference in risk of fibrosis at long-term follow-up.

[0179] Two candidate SNPs (rs9399005 and rsl805794) in the CTGF and NBS1 genes, respectively, showed significant associations with fibrosis whereas the candidate SNP (rs373759) in the ATM gene was not significant. The recessive effect of the CTGF minor allele (T) and the additive effect of the NBS1 risk allele (G) were similar for fibrosis and RILA, supporting this interpretation. Importantly, the lower RILA values of the CTGF minor homozygote would predict an increased rate of fibrosis according to the previous study (22) whereas in fact the TT genotype conferred strong protection, assigning it to the low-risk group. This implies that the direct effect of the TT genotype on fibrosis dominates over the association with reduced RILA values, strongly suggesting that independent effects of a single SNP may operate at different levels and supporting the decision to include both significant SNPs together with RILA values in the predictive model. Among the clinical factors, only BMI and hypertension were significantly associated with fibrosis in univariate analysis and were included as features in predictive modeling.

[0180] The strong correlation of the three risk groups defined by the decision tree (DT) of the partition analysis (PA) and by the combined ML models (RF+SVM) provided evidence and internal validation that some risk factors (features) were important only in the context of certain other risk factors and makes it highly unlikely that the DT was a result of overfitting. These results strongly support the hypothesis that functional assays in combination with 'omics' may define subgroups representing different aspects of the mechanistic pathways leading to a specific endpoint (2). The AUC was similar for the two approaches but the more robust combined machine learning models (ML: RF+SVM) seemed to perform marginally better than PA with respect to identifying fibrosis-positive patients in the high- and intermediate-risk groups. At the same time, the DT from PA provides mechanistic insight into the predictive model. The performance of RF and SVM in comparison with logistic regression (LR) was very favourable, further supporting modeling with subgroup structure. Three other machine learning models also showed lower AUC values.

[0181] Mechanistic aspects of the predictive model

[0182] CTGF plays an important role in the regulation of ECM homeostasis. Along with TGF-pi and downstream signaling proteins, CTGF controls the production of collagen and other components of ECM (40-42). Studies reported that blocking CTGF may stop or even reverse the process of fibrosis development in some cases (43-45).

[0183] A previous study showing an association of the rs9399005 SNP with systemic sclerosis reported a higher affinity of regulatory proteins for the minor allele (T) compared to the major (C) allele and detected an alternative isoform of CTGF mRNA (46). An in-silico analysis of transcription factors binding profiles using the most recent version of the JASPAR database (Rauluseviciute et al., Nucl. Acids Res 52(D1):D174-D182, 2023) confirmed that the T-allele has affinity for significantly more transcription factor binding motifs compared to the C-allele (Figure 25a). Furthermore, bindings affinities for individual transcription factors were heavily skewed towards the T-allele (Figure 25b).

[0184] Additional RNAseq experiments with fibroblasts in vitro showed a strong effect of the T-allele on gene expression. Early-passage fibroblast cultures from donors with different rs9399005 genotypes (n=5-6 of each) available from a local biobank were used. The expression of the CTGF gene itself was significantly increased by up to 2-fold in the order CC<CT<TT irrespective of timepoint and irradiation dose (Figure 26a, b). Furthermore, gene expression profiles showed that the number of differentially expressed genes (DEGs; false discovery rate, FDR = 0.05) after two days in culture was approximately 10-fold lower for the TT compared with the CC genotype and approximately 2.5-3-fold lower if the fibroblasts had been irradiated (Figure 26c). The heterozygote (CT) genotype was intermediate between the two. These results confirm that the rs9399005 SNP genotype has a profound functional effect on the phenotype of fibroblasts.

[0185] The NBS1 gene product, nibrin, is a key component of the MRN complex (MRE11-RAD50-NBS1) which plays a critical role in detecting and stabilizing DSBs during repair. The SNP (rsl805794) is located in the exon region responsible for interaction with SplOO nuclear antigen (SP100), near the BRCA1 C- terminal (BRCT) domain, and a nearby mutation (176:Tyr to Ala) was reported to block radiation- induced phosphorylation (47). In a previous study, no association was found with acute normal-tissue reaction, (48) nor with telangiectasia at 4.3 years mean follow-up after radiation therapy (49). To the best of our knowledge, the present study is the first to show an association of this SNP with fibrosis.

[0186] Among six clinical characteristics, only BMI and hypertension were associated with fibrosis. Logistic regression analysis showed that the effects of the two parameters were independent. A high BMI has been associated with an increased risk of liver fibrosis which may be related to an enhanced inflammatory state associated with visceral fat (47,50). Large breast size and hypertension were associated with fibrosis after hypofractionated radiation therapy (51). Furthermore, hypertension is a risk factor for myocardial fibrosis (52-54) and has been implicated in the development of liver and kidney fibrosis (55).

[0187] The association of low RILA values with an increased risk of fibrosis may suggest an involvement of the immune system (2). A low rate of apoptosis in irradiated lymphocytes might enhance the production of cytokines and attract inflammatory immune cells to the irradiated tissue (20,56). Alternatively, a low rate of apoptosis in CD4+lymphocytes may drive polarization towards a pro-fibrotic response (22). The CD4+subtypes Th2 and Thl7 are pro-fibrogenic via secreted cytokines and their effects on fibroblasts and ECM deposition (reviewed in (57)). Tregs control Thl7 and the Thl7 / Treg ratio has been associated with fibrosis but Tregs also secrete TGF-pi and IL-10, which stimulate and inhibit ECM deposition, respectively, and have been reported to have pro- or anti-fibrogenic effects (57).

[0188] Clinical potential and comparison with other models.

[0189] The combined machine learning model (ML: RF+SVM) was able to correctly predict 25 / 30 (83%) fibrosis-positive patients with fibrosis in the high-risk group (13.7% of the cohort) and 121 / 146 (83%) fibrosis-negative patients in the low-risk group (67% of the cohort), representing a five-fold difference in risk levels (83% and 17.1%, respectively). Thus, more than 80% of the cohort could be predicted with a precision of 83% by just five features. In a clinical setting, this would potentially allow clinicians to reduce the frequency of fibrosis by up to 37.9% (25 / 66) by modifying radiation therapy in the high-risk group. An example of an alternative, personalized treatment, that might be offered to patients in the high-risk group, is partial breast radiotherapy. Thus TARGIT-A trial on early breast cancer has shown non-inferiority regarding local recurrence and breast cancer-specific survival, of risk-adapted intraoperative RT (IORT) to the tumor bed compared with conventional external beam RT (EBRT) to the whole breast (Vaidya et al. (2020) BMJ; 370:m2836. doi: 10.1136 / bmj.m2836 and Vaidya et al. (2023), Int J Radiat Oncol Biol Phys, 115(l):77-82). A previous paper from one of the participating centers suggested that treatment with IORT without EBRT reduced the risk of fibrosis by approximately two thirds from 18.4% for the EBRT-only control arm (n=55) to 5.9% (n=34) in the lORT-only patients (Sperk et al. (2012), Breast Cancer Res Treat, 135:253-60). Although fibrosis was scored at 36 months and may continue to develop at a decreasing yearly rate for up to ten years, the ratio of the fibrosis rates between the two treatment groups is not expected to change with time. Assuming a rate of fibrosis of 30% 10 years after EBRT, and that 37.9% of these will be in the high-risk group (i.e. 11.4% of all patients), and that two thirds of these may be spared by treating the high-risk group with IORT, a potential reduction in fibrosis by 7.6% (= 2 / 3 of 11.4%) of all patients from 30% to 22.4% may be expected, which would be clinically relevant.

[0190] In a previous study, six different machine learning models were used to predict radiation therapy outcome and three toxicities in patients treated for different cancers ((non-)small-cell lung cancer, head-and-neck cancer, meningioma) (58). The models were developed and evaluated on 12 separate datasets but found no best model for all datasets. AUC values for the RF model applied to the 12 data sets were in the range of 0.54-0.82 with a median of 0.725, i.e. comparable to the simpler and more specific model in the present study.

[0191] A number of recent studies also aimed at predicting late normal tissue reactions in breast cancer patients, using machine learning modeling. A logistic regression model incorporating several clinical and genomic features reported a misclassification error of 34% (27) while a pre-conditioned random forest regression (PRFR) with SNPs from a GWAS on prostate cancer patients yielded AUC values of 0.55-0.70 for four genitourinary toxicities (26) in ROC analysis with AUC values for the other three endpoints in the range 0.55-0.64. The REQUITE cohort has been used to validate candidate SNPs identified from the literature (11) with modest results. A similar approach selecting sets of 13 interacting SNPs yielded AUC values of 0.78 for late urinary frequency >gr.2 and 0.71 for late hematuria >gr.l but just 0.63-0.68 for three other endpoints (25).

[0192] Conclusion

[0193] Significant associations of two SNPs, rs9399005 in CTGF and rsl805794 SNP in NBS1 were found with subcutaneous fibrosis in breast cancer patients after RT. Predictive modeling based on just five features yielded subgroups with up to five-fold different risk of fibrosis. This shows that certain risk factors only have a major impact in subgroups defined by a combination of certain other risk factors. The novel combined machine learning (ML: RF+SVM) model was able to assign 80% of the patients to the high- and low-risk groups with a 17% error rate and has the potential to spare 37.9% of the patients who would go on to develop fibrosis.

[0194] References

[0195] 1. Safwat A, Bentzen SM, Turesson I, Hendry JH. Deterministic rather than stochastic factors explain most of the variation in the expression of skin telangiectasia after radiotherapy. IntJ Radiat Oncol Biol Phys 2002;52(l):198-204 doi S0360301601026906 [pii],

[0196] 2. Herskind C, Talbot CJ, Kerns SL, Veldwijk MR, Rosenstein BS, West CM. Radiogenomics: A systems biology approach to understanding genetic risk factors for radiotherapy toxicity? Cancer Lett 2016;382(l):95-109 doi 10.1016 / j.canlet.2016.02.035.

[0197] 3. Talbot CJ, Tanteles GA, Barnett GC, Burnet NG, Chang-Claude J, Coles CE, et al. A replicated association between polymorphisms near TNFalpha and risk for adverse reactions to radiotherapy. BrJ Cancer 2012;107(4):748-53 doi 10.1038 / bjc.2012.290.

[0198] 4. Edvardsen H, Landmark-Hoyvik H, Reinertsen KV, Zhao X, Grenaker-Alnaes Gl, Nebdal D, et al. SNP in TXNRD2 associated with radiation-induced fibrosis: a study of genetic variation in reactive oxygen species metabolism and signaling. IntJ Radiat Oncol Biol Phys 2013;86(4):791-9 doi 10.1016 / j.ijrobp.2013.02.025.

[0199] 5. Seibold P, Behrens S, Schmezer P, Helmbold I, Barnett G, Coles C, et al. XRCC1 Polymorphism Associated With Late Toxicity After Radiation Therapy in Breast Cancer Patients. IntJ Radiat Oncol Biol Phys 2015;92(5):1084-92 doi 10.1016 / j.ijrobp.2015.04.011.

[0200] 6. Andreassen CN, Rosenstein BS, Kerns SL, Ostrer H, De Ruysscher D, Cesaretti JA, et al. Individual patient data meta-analysis shows a significant association between the ATM rsl801516 SNP and toxicity after radiotherapy in 5456 breast and prostate cancer patients. Radiother Oncol 2016;121(3):431-9 doi 10.1016 / j.radonc.2016.06.017.

[0201] 7. Kerns SL, Dorling L, Fachal L, Bentzen S, Pharoah PD, Barnes DR, et al. Meta-analysis of Genome Wide Association Studies Identifies Genetic Markers of Late Toxicity Following Radiotherapy for Prostate Cancer. EBioMedicine 2016;10:150-63 doi 10.1016 / j.ebiom.2016.07.022.

[0202] 8. Barnett GC, Thompson D, Fachal L, Kerns S, Talbot C, Elliott RM, et al. A genome wide association study (GWAS) providing evidence of an association between common genetic variants and late radiotherapy toxicity. Radiother Oncol 2014;lll(2):178-85 doi 10.1016 / j.radonc.2014.02.012.

[0203] 9. Fachal L, Gomez-Caamano A, Barnett GC, Peleteiro P, Carballo AM, Calvo-Crespo P, et al. A three-stage genome-wide association study identifies a susceptibility locus for late radiotherapy toxicity at 2q24.1. Nat Genet 2014;46(8):891-4 doi 10.1038 / ng.3020.

[0204] 10. Kerns SL, Fachal L, Dorling L, Barnett GC, Baran A, Peterson DR, et al. Radiogenomics Consortium Genome-Wide Association Study Meta-Analysis of Late Toxicity After Prostate Cancer Radiotherapy. J Natl Cancer Inst 2020;112(2):179-90 doi 10.1093 / jnci / djz075. 11. Massi MC, Gasperoni F, leva F, Paganoni AM, Zunino P, Manzoni A, et al. A Deep Learning Approach Validates Genetic Risk Factors for Late Toxicity After Prostate Cancer Radiotherapy in a REQUITE Multi-National Cohort. Front Oncol 2020;10:541281 doi 10.3389 / fonc.2020.541281.

[0205] 12. Barnett GC, Coles CE, Burnet NG, Pharoah PD, Wilkinson J, West CM, et al. No association between SNPs regulating TGF-betal secretion and late radiotherapy toxicity to the breast: results from the RAPPER study. Radiother Oncol 2010;97(l):9-14 doi 10.1016 / j.radonc.2009.12.006.

[0206] 13. Barnett GC, Elliott RM, Alsner J, Andreassen CN, Abdelhay O, Burnet NG, et al. Individual patient data meta-analysis shows no association between the SNP rsl800469 in TGFB and late radiotherapy toxicity. Radiother Oncol 2012;105(3):289-95 doi

[0207] 10.1016 / j.radonc.2012.10.017.

[0208] 14. Andreassen CN, Alsner J. Genetic variants and normal tissue toxicity after radiotherapy: a systematic review. Radiother Oncol 2009;92(3):299-309 doi S0167-8140(09)00325-9 [pii] 10.1016 / j.radonc.2009.06.015.

[0209] 15. Barnett GC, West CM, Coles CE, Pharoah PD, Talbot CJ, Elliott RM, et al. Standardized Total Average Toxicity score: a scale- and grade-independent measure of late radiotherapy toxicity to facilitate pooling of data from different studies. IntJ Radiat Oncol Biol Phys 2012;82(3):1065-74 doi 10.1016 / j.ijrobp.2011.03.015.

[0210] 16. Tucker SL, Turesson I, Thames HD. Evidence for individual differences in the radiosensitivity of human skin. Ear J Cancer 1992;28A(11):1783-91.

[0211] 17. Bentzen SM, Overgaard M, Overgaard J. Clinical correlations between late normal tissue endpoints after radiotherapy: implications for predictive assays of radiosensitivity. Eur J Cancer 1993;29A(10):1373-6.

[0212] 18. Lilia C, Ambrosone CB, Kropp S, Helmbold I, Schmezer P, von Fournier D, et al. Predictive factors for late normal tissue complications following radiotherapy for breast cancer. Breast Cancer Res Treat 2007;106(l):143-50.

[0213] 19. Azria D, Riou O, Castan F, Nguyen TD, Peignaux K, Lemanski C, et al. Radiation-induced CD8 T- lymphocyte Apoptosis as a Predictor of Breast Fibrosis After Radiotherapy: Results of the Prospective Multicenter French Trial. EBioMedicine 2015;2(12):1965-73 doi 10.1016 / j.ebiom.2015.10.024.

[0214] 20. Ozsahin M, Crompton NE, Gourgou S, Kramar A, Li L, Shi Y, et al. CD4 and CD8 T-lymphocyte apoptosis can predict radiation-induced late toxicity: a prospective study in 399 patients. Clin Cancer Res 2005;ll(20):7426-33 doi 10.1158 / 1078-0432.CCR-04-2634.

[0215] 21. Bourgier C, Castan F, Riou O, Nguyen TD, Peignaux K, Lemanski C, et al. Impact of adjuvant hormonotherapy on radiation-induced breast fibrosis according to the individual radiosensitivity: results of a multicenter prospective French trial. Oncotarget 2018;9(21):15757-65 doi 10.18632 / oncotarget.24606.

[0216] 22. Veldwijk MR, Seibold P, Botma A, Helmbold I, Sperk E, Giordano FA, et al. Association of CD4(+) Radiation-Induced Lymphocyte Apoptosis with Fibrosis and Telangiectasia after Radiotherapy in 272 Breast Cancer Patients with >10-Year Follow-up. Clin Cancer Res 2019;25(2):562-72 doi 10.1158 / 1078-0432.CCR-18-0777.

[0217] 23. Seibold P, Webb A, Aguado-Barrera ME, Azria D, Bourgier C, Brengues M, et al. REQUITE: A prospective multicentre cohort study of patients undergoing radiotherapy for breast, lung or prostate cancer. Radiother Oncol 2019;138:59-67 doi 10.1016 / j.radonc.2019.04.034.

[0218] 24. Aldraimli M, Soria D, Grishchuck D, Ingram S, Lyon R, Mistry A, et al. A data science approach for early-stage prediction of Patient's susceptibility to acute side effects of advanced radiotherapy. Comput Biol Med 2021;135:104624 doi 10.1016 / j. compbiomed.2021.104624.

[0219] 25. Franco NR, Massi MC, leva F, Manzoni A, Paganoni AM, Zunino P, et al. Development of a method for generating SNP interaction-aware polygenic risk scores for radiotherapy toxicity. Radiother Oncol 2021;159:241-8 doi 10.1016 / j.radonc.2021.03.024. 26. Lee S, Kerns S, Ostrer H, Rosenstein B, Deasy JO, Oh JH. Machine Learning on a Genome-wide Association Study to Predict Late Genitourinary Toxicity After Prostate Radiation Therapy. Int J Radiat Oncol Biol Phys 2018;101(l):128-35 doi 10.1016 / j.ijrobp.2018.01.054.

[0220] 27. Mbah C, De Ruyck K, De Schrijver S, De Sutter C, Schiettecatte K, Monten C, et al. A new approach for modeling patient overall radiosensitivity and predicting multiple toxicity endpoints for breast cancer patients. Acta Oncol 2018;57(5):604-12 doi

[0221] 10.1080 / 0284186X.2017.1417633.

[0222] 28. Straub JM, New J, Hamilton CD, Lominska C, Shnayder Y, Thomas SM. Radiation-induced fibrosis: mechanisms and implications for therapy. J Cancer Res Clin Oncol 2015;141(ll):1985-94 doi 10.1007 / s00432-015-1974-6.

[0223] 29. Nogueira RMP, Vital FMR, Bernabe DG, Carvalho MB. Interventions for Radiation-Induced Fibrosis in Patients With Breast Cancer: Systematic Review and Meta-analyses. Adv Radiat Oncol 2022;7(3):100912 doi 10.1016 / j.adro.2022.100912.

[0224] 30. Auton A, Brooks LD, Durbin RM, Garrison EP, Kang HM, Korbel JO, et al. A global reference for human genetic variation. Nature 2015;526(7571):68-74 doi 10.1038 / naturel5393.

[0225] 31. Ozsahin M, Ozsahin H, Shi Y, Larsson B, Wurgler FE, Crompton NE. Rapid assay of intrinsic radiosensitivity based on apoptosis in human CD4 and CD8 T-lymphocytes. Int J Radiat Oncol Biol Phys 1997;38(2):429-40 doi 10.1016 / s0360-3016(97)00038-2.

[0226] 32. Menz R, Andres R, Larsson B, Ozsahin M, Trott K, Crompton NE. Biological dosimetry: the potential use of radiation-induced apoptosis in human T-lymphocytes. Radiat Environ Biophys 1997;36(3):175-81 doi 10.1007 / s004110050069.

[0227] 33. Ordonez R, Henriquez-Hernandez LA, Federico M, Valenciano A, Pinar B, Lloret M, et al. Radio-induced apoptosis of peripheral blood CD8 T lymphocytes is a novel prognostic factor for survival in cervical carcinoma patients. Strahlenther Onkol 2014;190(2):210-6 doi

[0228] 10.1007 / s00066-013-0488-x.

[0229] 34. Cozzarini C. Radiation Induced Lymphocyte Apoptosis: An Effective Way of "Tailoring" Radiotherapy to the Right Patients Only? EBioMedicine 2015;2(12):1852-3 doi 10.1016 / j.ebiom.2015.11.014.

[0230] 35. Fhoghlu MN, Barrett S. A Review of Radiation-Induced Lymphocyte Apoptosis as a Predictor of Late Toxicity After Breast Radiotherapy. J Med Imaging Radiat Sci 2019;50(2):337-44 doi 10.1016 / j.jmir.2019.02.004.

[0231] 36. LENT SOMA tables. Radiother Oncol 1995;35(l):17-60.

[0232] 37. LENT SOMA scales for all anatomic sites. Int J Radiat Oncol Biol Phys 1995;31(5):1049-91 doi 10.1016 / 0360-3016(95)90159-0.

[0233] 38. Moore CM, Jacobson SA, Fingerlin TE. Power and Sample Size Calculations for Genetic Association Studies in the Presence of Genetic Model Misspecification. Hum Hered 2019;84(6):256-71 doi 10.1159 / 000508558.

[0234] 39. Ojala M, Garriga GC. Permutation Tests for Studying Classifier Performance. 20096-9 Dec. 2009. p 908-13.

[0235] 40. Duncan MR, Frazier KS, Abramson S, Williams S, Klapper H, Huang X, et al. Connective tissue growth factor mediates transforming growth factor beta-induced collagen synthesis: downregulation by CAMP. FASEBJ 1999;13(13):1774-86.

[0236] 41. Garrett Q, Khaw PT, Blalock TD, Schultz GS, Grotendorst GR, Daniels JT. Involvement of CTGF in TGF-betal-stimulation of myofibroblast differentiation and collagen matrix contraction in the presence of mechanical stress. Invest Ophthalmol Vis Sci 2004;45(4):1109-16 doi

[0237] 10.1167 / iovs.03-0660.

[0238] 42. Quan T, Shao Y, He T, Voorhees JJ, Fisher GJ. Reduced expression of connective tissue growth factor (CTGF / CCN2) mediates collagen loss in chronologically aged human skin. J Invest Dermatol 2010;130(2):415-24 doi 10.1038 / jid.2009.224.

[0239] 43. Barbe MF, Hilliard BA, Amin M, Harris MY, Hobson U, Cruz GE, et al. Blocking CTGF / CCN2 reverses neural fibrosis and sensorimotor declines in a rat model of overuse-induced median mononeuropathy. J Orthop Res 2020;38(ll):2396-408 doi 10.1002 / jor.24709. 44. Bickelhaupt S, Erbel C, Timke C, Wirkner U, Dadrich M, Flechsig P, et al. Effects of CTGF Blockade on Attenuation and Reversal of Radiation-Induced Pulmonary Fibrosis. J Natl Cancer Inst 2017;109(8) doi 10.1093 / jnci / djw339.

[0240] 45. Lipson KE, Wong C, Teng Y, Spong S. CTGF is a central mediator of tissue remodeling and fibrosis and its inhibition can reverse the process of fibrosis. Fibrogenesis Tissue Repair 2012;5(Suppl 1):S24 doi 10.1186 / 1755-1536-5-S1-S24.

[0241] 46. Granel B, Argiro L, Hachulla E, Fajardy I, Weiller PJ, Durand JM, et al. Association between a CTGF gene polymorphism and systemic sclerosis in a French population. J Rheumatol 2010;37(2):351-8 doi 10.3899 / jrheum.090290.

[0242] 47. Cerosaletti KM, Concannon P. Nibrin forkhead-associated domain and breast cancer C- terminal domain are both required for nuclear focus formation and phosphorylation. J Biol Chem 2003;278(24):21944-51 doi 10.1074 / jbc.M211689200.

[0243] 48. Popanda O, Tan XL, Ambrosone CB, Kropp S, Helmbold I, von Fournier D, et al. Genetic polymorphisms in the DNA double-strand break repair genes XRCC3, XRCC2, and NBS1 are not associated with acute side effects of radiotherapy in breast cancer patients. Cancer Epidemiol Biomarkers Prev 2006;15(5):1048-50.

[0244] 49. Chang-Claude J, Ambrosone CB, Lilia C, Kropp S, Helmbold I, von Fournier D, et al. Genetic polymorphisms in DNA repair and damage response genes and late normal tissue complications of radiotherapy for breast cancer. Br J Cancer 2009;100(10):1680-6 doi 6605036 [pii] 10.1038 / sj.bjc.6605036.

[0245] 50. Walker RW, Allayee H, Inserra A, Fruhwirth R, Alisi A, Devito R, et al. Macrophages and fibrosis in adipose tissue are linked to liver damage and metabolic risk in obese children. Obesity (Silver Spring) 2014;22(6):1512-9 doi 10.1002 / oby.20730.

[0246] 51. Colciago RR, Cavallo A, Magri MC, Vitullo A, La Rocca E, Giandini C, et al. Hypofractionated whole-breast radiotherapy in large breast size patients: is it really a resolved issue? Med Oncol 2021;38(9):107 doi 10.1007 / sl2032-021-01550-6.

[0247] 52. Diez J. Mechanisms of cardiac fibrosis in hypertension. J Clin Hypertens (Greenwich) 2007;9(7):546-50 doi 10.1111 / j.l524-6175.2007.06626.x.

[0248] 53. Iyer NR, Le TT, Kui MSL, Tang HC, Chin CT, Phua SK, et al. Markers of Focal and Diffuse Nonischemic Myocardial Fibrosis Are Associated With Adverse Cardiac Remodeling and Prognosis in Patients With Hypertension: The REMODEL Study. Hypertension 2022;79(8):1804-13 doi 10.1161 / HYPERTENSIONAHA.122.19225.

[0249] 54. Tomek J, Bub G. Hypertension-induced remodelling: on the interactions of cardiac risk factors. J Physiol 2017;595(12):4027-36 doi 10.1113 / JP273043.

[0250] 55. Sun HJ. Current Opinion for Hypertension in Renal Fibrosis. Adv Exp Med Biol 2019;1165:37- 47 doi 10.1007 / 978-981-13-8871-2_3.

[0251] 56. Azria D, Ozsahin M, Kramar A, Peters S, Atencio DP, Crompton NE, et al. Single nucleotide polymorphisms, apoptosis, and the development of severe late adverse effects after radiotherapy. Clin Cancer Res 2008;14(19):6284-8 doi 10.1158 / 1078-0432.CCR-08-0700.

[0252] 57. Zhang M, Zhang S. T Cells in Fibrosis and Fibrotic Diseases. Front Immunol 2020;ll:1142 doi 10.3389 / fimmu.2020.01142.

[0253] 58. Deist TM, Dankers F, Valdes G, Wijsman R, Hsu IC, Oberije C, et al. Machine learning algorithms for outcome prediction in (chemo)radiotherapy: An empirical comparison of classifiers. Med Phys 2018;45(7):3449-59 doi 10.1002 / mp.l2967.

Claims

CLAIMS1. Method for predicting the risk of a cancer subject for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the method comprises(a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or(b) genotyping the SNP rsl805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having the major homozygous genotype (CC) has a lower risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

2. The method of claim 1, wherein the method further comprises(c) performing a radiation-induced lymphocyte apoptosis (RILA) assay on a lymphocyte population of the cancer patient, wherein RILA is the percentage of lymphocyte apoptosis induced by a certain radiation dose or a DNA damaging agent minus the spontaneous lymphocyte apoptosis, and is preferably the percentage of lymphocyte apoptosis induced by 8 Gy minus the percentage of apoptosis at 0 Gy, and wherein a cancer patient having a high RILA value has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having a low RILA value has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery.

3. The method of any one of claim 1 or 2, wherein the method further comprises(d) determining the body mass index (BMI) of the cancer patient, wherein a cancer patient having a low BMI has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having a high BMI has an increased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and / or(e) determining the blood pressure of the cancer patient, wherein a cancer patient having no hypertension has a low risk for skin fibrosis after radiation therapy and optionally after cancer surgery, and a cancer patient having hypertension has anincreased risk for skin fibrosis after radiation therapy and optionally after cancer surgery, wherein hypertension is preferably a blood pressure of 140 / 90 mmHg or higher.

4. The method of any one of claims 1 to 3, wherein the method further comprises(f) determining or having determined two or more of the genotype of the SNP rs9399005, the genotype of the SNP rsl805794 or a SNP of Table 1, RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for two or more of the genotype of the SNP rs9399005, the genotype of the SNP rsl805794 or a SNP of Table 1, RILA value, the BMI and the blood pressure in the population of samples; and(f') subjecting or having subjected the data sets of (f) to one or more machine learning models in order to establish a predictive model within the two more data sets that indicate the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, and(f") employing the predictive model of (f') in predicting the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery in the cancer patient for whom the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery is to be determined.

5. The method of claim 4, wherein the two or more of (f) comprise the genotype of the SNP rs9399005 and / or the genotype of the SNP rsl805794 or a SNP of Table 1.

6. A method for establishing a predictive model that indicates the risk for developing skin fibrosis after radiation therapy and optionally after cancer surgery in a test cancer patient, wherein the method comprises(a) determining the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the following: RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for the genotype(s) of the SNP rs9399005 and / or the SNPrsl805794 or a SNP of Table 1 and optionally at least one of the following: RILA value, theBMI and the blood pressure in the population of samples; and(b) subjecting the data sets of (a) to one or more machine learning models in order to establish the predictive model within the two or more data sets that indicates the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery.

7. A computer, system or storage medium being configured for predicting the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the computer, system or storage medium stores a predictive model that indicates the risk for developing skin fibrosis in a test cancer patient after radiation therapy and optionally after cancer surgery, wherein the predictive model has been obtained by a method comprising(a) determining the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1, and optionally at least one of the following: RILA value, the BMI and the blood pressure in a population of samples, wherein each sample has been obtained from a cancer patient, and wherein it is known for each cancer patient whether a skin fibrosis occurred after the radiation therapy and optionally after the cancer surgery, thereby obtaining data sets for the genotype(s) of the SNP rs9399005 and / or the SNP rsl805794 or a SNP of Table 1 and optionally at least one of the following: RILA value, the BMI and the blood pressure in the population of samples; and(b) subjecting the data sets of (a) to one or more machine learning models in order to establish a predictive model within the two more data sets that indicate the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery.

8. The method of any one of claims 4 to 6 or the computer, system or storage medium of claim 7, wherein the one or more machine learning models comprise an ensemble machine learning algorithm and / or a support vector machine (SVM) algorithm and / or neural networks (NN) machine learning algorithms.

9. The method, computer, system or storage medium of claim 8, wherein the ensemble machine learning algorithm is random forest (RF), which is preferably based on creating a number of decision trees on subgroups of the data sets, and the predictive model is generated based on votes of majority of the trees.

10. The method, computer, system or storage medium of claim 7 or 8, wherein the SVM algorithm is developed with the Radial Basis Function (RBF) kernel function, and the predictive model is generated based on the differences of inner outliers, or samples that are most similar to the opposite class, preferably fibrosis negative samples that are similar to fibrosis positive samples vs. fibrosis positive samples that are similar to fibrosis negative samples.

11. The method, computer, system or storage medium of any one of claims 4 to 10, wherein the population of samples comprises at least 25 samples, preferably at least 50 samples, more preferably at least 100 samples, and most preferably at least 200 samples.

12. The method, computer, system or storage medium of any one of claims 1 to 11, wherein the cancer is breast cancer and head and neck cancer.

13. The method, computer, system or storage medium of any one of claims 1 to 12, wherein the skin fibrosis is subcutaneous fibrosis.

14. The method, computer, system or storage medium of any one of claims 1 to 13, wherein the sample of the (test) patient and / or the samples in the population of samples has / have been obtained before, during or after radiation therapy and optionally before, during, or after cancer surgery.

15. A kit for predicting the risk for developing skin fibrosis in a cancer patient after radiation therapy and optionally after cancer surgery, wherein the kit comprises means for(a) genotyping the SNP rs9399005 in a sample obtained from the cancer patient, wherein a cancer patient having the minor homozygous genotype (TT) has a low risk for radiation-induced skin fibrosis, and a cancer patient having the heterozygous genotype (CT) or major homozygous genotype (CC) has an increased risk for radiation-induced skin fibrosis, and / or(b) genotyping the SNP rsl805794 or a SNP of Table 1 in a sample obtained from the cancer patient, wherein a cancer patient having the heterozygous genotype (CG) or minor homozygous genotype (GG) has a high risk for radiation-induced skin fibrosis a cancer, and a cancer patient having the major homozygous genotype (CC) has a lower risk for radiation-induced skin fibrosis.

Images