Identification of biomarkers for breast cancer early diagnosis based on the molecular classification using machine learning algorithms on transcriptomic data and factorial designs for analysis

doi:10.21203/rs.3.rs-2877610/v1

Download PDF

Research Article

Identification of biomarkers for breast cancer early diagnosis based on the molecular classification using machine learning algorithms on transcriptomic data and factorial designs for analysis

https://doi.org/10.21203/rs.3.rs-2877610/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background:

Breast cancer is the second leading cause of global female mortality. Diagnosing and treating breast cancer patients at early stages is relevant for providing successful treatment and increasing the patient's survival rate. The use of new analytical methods for massive data from biological samples, such as Machine Learning Algortithms (MLAs), is necessary for improving cancer diagnosis, especially in patients from low-income countries. A computational methodology for selecting a small number of biomarkers with strong diagnostic capabilities and an accessible cellular location could be useful for developing low-cost diagnostic devices. Hence, this study aimed to develop a computational methodology to find relevant genetic biomarkers and establish a discrete panel of genes capable of classifying breast cancer samples for diagnostic purposes with high accuracy.

Methods:

This study aimed to develop a computational methodology for finding genetic biomarkers and establish a panel with a few genes capable of classifying breast cancer molecularly for diagnostic purposes. Panels with a small number of genes (<10) that can be used for the molecular classification of breast cancer cells through four Machine Learning Algorithms on transcriptomic data. Five gene selection approaches were used for the generation of these panels: factor analysis genes, surfaceome genes, transmembrane genes, combined genes, and network analysis genes. The classification performance and analyzed and validated using seven factorial designs and non-parametric statistical tests.

Results:

The MLAs accuracy was higher than 80% in cell lines and in patient samples for all selection approaches. The combined approach with the best genes of the three approaches (transmembrane, surfaceome, and factor analysis) had better classification performance than each approach alone. Also, the combined genes of this approach (TMEM210, CD44, SPDEF, TENM4, KIRREL, BCAS1, TMEM86A, LRFN2, TFF3) had similar performance than the ones selected by network analysis. The panel of genes identified from the combined approach was completely different from the genes previously described in four commercial panels for breast cancer that were analyzed.

Conclusions

In this study, the panels of selected genes were capable of classify breast cancer cell lines and patient samples according to their molecular characteristics. Two genes of the combined approach (TFF3 and CD44) have been used in cancer biosensors, which suggests a plausible result due to the potential for the development of new diagnostic devices; however, experimental studies are required to corroborate this type of implementation.

Triple Negative Breast Cancer

Molecular Classification

Artificial Intelligence

Gene Expression

Factorial Analysis

Statistical validation

Trans-membrane genes

Surfaceome

Bioinformatics

RNA-seq

Breast cancer is the second leading cause of death in women and has become the most commonly diagnosed cancer worldwide since last year (1, 2). Around 2.3 million new cases were reported, and more than 30% died from this disease only in 2020 (2). Diagnosing and treating the disease in its early stages have been recognized as important factors in improving the survival rate and treatment success (3, 4). Breast cancer mortality in high-income countries is decreasing, owing to the implementation of prevention campaigns, early diagnosis, and improved treatments (4, 5). However, socioeconomic and geographic factors in low-income and middle-income countries still represent a barrier to early detection and an effective healthcare service (6). Nowadays, imaging is the most efficient tool for detecting breast cancer; however, more accessible and precise diagnostic tools are needed to combat this global health problem. Using molecular diagnostic techniques is a relevant alternative to solve this situation.

Breast cancer is a highly heterogeneous disease; thus, it can be classified according to histology and molecular characteristics (7). Histologically, breast cancer could be divided into two categories: In situ Carcinoma and Invasive Carcinoma, according to the tumors’ architectural features and growth patterns (7). Meanwhile, gene expression analysis has allowed the characterization of molecular subtypes of breast cancer according to gene profile. Recent studies proposed six intrinsic subtypes: claudin-low, basal-like, ER2-overexpressing, normal breast-like, luminal subtype A, and luminal subtype B, which are directly related to the expression of Estrogen Receptor (ER) and Human Epidermal Growth Factor Receptor 2 (HER2) genes (7, 8). However, the classification of non-triple negative breast cancer (BC) —with high expression of ER, Progesterone Receptor (PR), or HER2 genes— and in triple-negative breast cancer (TNBC) —with low expression of these genes— continues to play an essential role in the diagnosis and treatment of this disease (9).

Machine Learning (ML), a type of artificial intelligence, is based on algorithms that use patterns in historical data to predict new output values (10). For this reason, Machine Learning techniques have been used widely in breast cancer diagnosis and prognosis over the last two decades (11).

Transcriptomic data with differential gene expression (12) and Machine Learning (13) algorithms were used to find relevant genes for the classification of breast cancer; however, the analytical strategies and validation procedures can be significantly improved. The development of methodologies capable of selecting a small number of biomarkers of next-generation sequencing data from large datasets without compromising the diagnostic performance is of high interest. A small number of biomarkers with adequate diagnostic capabilities and an accessible cellular location (cell surface or plasma membrane) can be useful in developing low-cost diagnostic devices that can be implemented for breast cancer detection in low-income countries. Here, we developed an extensive study on the molecular classification of breast cancer using factorial designs to analyze multiple gene selection strategies, machine learning algorithms, and validation.

Specifically, (1) we developed a computational methodology based on MLAs for identifying genes (biomarker candidates) capable of classifying breast cancer samples and whose proteins are present in cellular locations (cell surface or membranes), which are useful for diagnostic device development. (2) We established a panel with a small number of genes (< 10) that can be used for the early detection (cancer vs. non-cancer) and molecular classification for diagnosis (non-triple negative vs. triple negative) of breast cancer cells through four Machine Learning Algorithms on transcriptomic data using five gene selection approaches: factor analysis genes, surfaceome genes, transmembrane genes, combined genes, and network analysis genes.

Trascriptomic data from cell lines and breast cancer patient biopsies

Transcriptomic data from RNA-seq experiments was obtained from breast cancer related cell lines and patient biopsies using online open repositories. The selection criteria for the datasets collected from these repositories included the following characteristics: 1) the dataset included more than 30 samples with raw data files available; 2) the samples belong to human breast cells; 3) RNA-seq data, quantification of gene expression, and molecular classification of the cells in three groups: Non-malignant (NM), Breast Cancer (BC), and Triple Negative Breast Cancer (TNBC). Three datasets from the Gene Expression Omnibus (GEO) database and one from The Cancer Genome Atlas (TCGA) were studied (Table 1). The raw data from these datasets were publicly available online and independent of each other. Only cell lines that were common in the 1st, 2nd, and 3rd datasets were used to have consistent data. The patient samples of the 3rd and 4th datasets were selected randomly from the data sources. Absolute counts were used for gene expression quantification in the initial analysis, validation on independent cell lines datasets, analysis on patient samples datasets, and comparison with the network analysis approach. In contrast, absolute counts and FPKM were statistically compared (Wilcoxson test) in the validation on patient samples datasets, analysis of the combined genes approach, and validation of the combined genes approach.

Table 1

Datasets used in the study
Dataset	Data source	Database^a	Type of sample	Dataset details^b
1st	GSE48216	GEO	Cell lines	20 BC, 13 TNBC, and 5 NM samples
2nd	GSE152908	GEO	Cell lines	20 BC, 12 TNBC, and 6 NM samples
3rd	GSE58135	GEO	Cell lines and patient samples	Cell lines: 9 BC and 7 TNBC. Patient samples: 10 BC (ER+), 10 TNBC, and 5 NM
4th	TCGA-BRCA	TCGA	Patient samples	52 BC, 50 TNBC, and 11 NM samples
^aGene Expression Omnibus (GEO); The Cancer Genome Atlas (TCGA).^bBreast Cancer (BC); Triple Negative Breast Cancer (TNBC); Non-malignant (NM)

Data pre-processing

Data pre-processing was performed using Access and Excel for dataset management and Tauber Bioinformatics Research Center Server (TBRCS) pipelines for transcriptomic analysis, normalization, and annotation (https://server.t-bio.info/). Two types of quantification for the gene expression were used: Absolute counts and Fragments Per Kilobase Million (FPKM). All datasets were normalized using natural logarithm transformation and quantile normalization with a threshold value of 5, which is commonly used in transcriptomic data (14).

Gene selection approaches

Five gene selection approaches were applied to achieve the objective of this study. These included three initial approaches, which were: 1) Surfaceome genes, 2) Transmembrane genes, and 3) Factor analysis genes; and two subsequent approaches named: 4) Combined genes and 5) Network analysis genes. The gene selection algorithms Factor Regression Analysis was used for the Factor analysis approach, while Random Forest was applied for the Surfaceome genes and Transmembrane genes approaches, because they are easy to implement, provide reliable results (15, 16), and require low computational power. Then for the Combined gene approach instead of using an algorithm the three genes with the best performance obtained from each one initial approaches were selected. Finally, for the Network analysis the algorithm reported by Ren et al. (17, 18) was used.

In the next paragraphs, the steps followed by each gene selection approach are described:

1) Surfaceome genes approach: Genes from Bausch-Fluck’s surfaceome (19) present in the 1st dataset were studied. A total of 2,653 genes were obtained. The Random Forest Algorithm of the TBRCS was applied to reduce the number of genes using the Mean Decrease Accuracy (MDA) as the selection criteria (a higher value represents a higher contribution to the classification performance). The 38 surfaceome genes with MDA > 2.0 (percentile 99). Then, the number of genes was further reduced to manageable quantities for low-cost biosensor panels: ten, five, or three genes (Table 2).

2) Transmembrane genes approach

Genes annotated as transmembrane in the 1st dataset were studied. A total of 460 genes were obtained. The 38 transmembrane genes with MDA > 2.0 (same quantity as previous approach) were selected. Then, the number of genes was further reduced to ten, five, or three genes (Table 2).

3) Factor analysis approach. The cell lines of the 1st dataset were separated into three levels according to their molecular categories (NM, BC, and TNBC). The Factor Analysis pipeline of the TBRCS was used, indicating one factor and these three levels. The mentioned pipeline was run, separating the cell lines in the mentioned levels. The 38 genes (same quantity as previous approaches) with the highest difference in expression between each pair of levels (NM vs. BC, NM vs. TNBC, BC vs. TNBC) were selected. Then, the number of genes was further reduced to ten, five, or three (Table 2).

Table 2

Groups of genes selected in each of the three initial approaches
Number of genes			Transmembrane genes	Surfaceome genes	Factor analysis genes
10 genes	5 genes	3 genes	TMEM210	CD44	SPDEF
			TENM4	KIRREL1	BCAS1
			TMEM86A	LRFN2	TFF3
			TMEM45B	DCBLD1	TFF1
			TMPRSS13	KIAA1324	TRIL
			LRTOMT	CLDN3	AGR2
			TMEM229B	GPR160	KIF12
			LRTM2	GJB3	MUCL1
			TMEM125	PLAUR	DSCAM-AS1
			SIDT1	PTGER3	C1orf64

The list of 38 genes obtained in each of the previous selection approaches is presented in the supplementary material (List of genes from the initial selection approaches).

4) Combined genes approach

A group of nine genes was created by combining the best three genes from the three initial approaches. For gene reduction, a second group was formed from the previous one considering the five genes with MDA > 7.0 (percentile 56), which conserve genes from the three initial approaches and is a more manageable quantity for biosensor design. Both groups are presented in Table 3.

Table 3

Genes from the combined approach
Number of genes		Genes
9 genes	5 genes	TMEM86A
		KIRREL1
		SPDEF
		BCAS1
		TFF3
		TMEM210
		TENM4
		CD44
		LRFN2

5) Network analysis

The previous gene selection approaches have the disadvantage that strong correlations among genes are not appropriately accounted for. Therefore, for this reason, we decided to compare our beast approach (combined) against one that incorporates network structure when selecting genes. The last approach was performed by applying the algorithms developed by Ren et al. (17, 18), using the associated CRAN package called regnet. The classification procedure was reduced to two categories (BC and TNBC) because the regnet algorithms only accept binary or continuous data. The correlation coefficient of each gene from the 4th dataset was calculated using the network regnet algorithm with the parameters and R code presented in the supplementary information (Code and parameters of the Network Analysis). This approach required high computational power to be executed, so it was run in Purdue University Community cluster “Bell” using two nodes, with 40 cores each (AMD Rome CPUs @ 2.0GHz) and 1TB memory per node. The running time of the mentioned algorithm was 13 hours. The selected genes were ordered by the absolute value of the correlation coefficient, and the ones with a value equal to zero were removed, generating a list of 1,279 genes. The nine and five genes with the highest coefficient were selected and are presented in Table 4.

Table 4

Genes from the network analysis approach
Number of genes		Genes
9 genes	5 genes	AL078582
		AC098679
		AP006261
		AC141930
		AL390726
		PRRT1B
		AC093838
		PERCC1
		AC079296

Computational methodology

The computational methodology for the data analysis was constructed following the factorial experimental design to obtain solid statistical support and determine the most relevant factors that affect the classification performance of the Machine Learning Algorithms. Seven factorial designs were run during this study (Fig. 1) to accomplish the mentioned general objectives. The specific objectives of each factorial design are described as follows:

Initial analysis: Determine if the Machine Learning Algorithms (MLA) are adequate to classify breast cancer (accuracy ≥ 70%) with the genes selected from three different approaches: Transmembrane, Surfaceome, and Factor analysis, described below.

Validation on independent cell lines datasets: Validate in two independent cell lines datasets if the MLA can classify breast cancer with similar accuracy to the initial analysis. Reduce the number of genes and establish their effect on the performance of the algorithms.

Analysis of patient sample datasets: Determine if the MLAs trained with cell lines can classify breast cancer adequately in patient samples (accuracy ≥ 70%).

Validation on patient sample datasets: Validate the performance of the MLAs with the selected genes of the three approaches when the algorithms are trained and tested in patient sample datasets.

Analysis of the combined genes approach: Determine if the MLAs’ performance in patient samples is improved when the best genes of the three initial approaches (Transmembrane, Surfaceome, and Factor analysis) are combined in one set.

Validation of the combined genes approach: Validate the MLAs’ performance in an independent dataset with patient samples: The Cancer Genome Atlas (TCGA).

Comparison with network analysis approach: Compare the MLAs’ performance of the combined genes approach against the network analysis approach on TCGA.

Initial analysis: It was an exploratory analysis of the 1st dataset to determine whether the transcriptomic data were suitable for the molecular classification of the breast cancer cell lines using MLAs. The 38 genes of each of the initial gene approaches (transmembrane, surfaceome, and factor analysis) were evaluated by Principal Component Analysis (PCA) on Rstudio using the CRAN packages ggfortify, and cluster to see if the cell lines of each molecular category were separated into different groups. Then, the factorial design was executed using the following factors and levels: Gene selection approach (transmembrane, surfaceome, and factor analysis); classification strategy (categories and subcategories); ML algorithm (Control (CZR), Random Forest (RF), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and Naive Bayes (BYS)). The MLAs were trained and tested in the same dataset.

Validation on independent cell lines datasets: Two independent datasets (2nd and 3rd ) were used to corroborate the performance of the MLAs for the classification of breast cancer cell lines. The factorial design was executed using the following factors and levels: Gene selection approach (transmembrane, surfaceome, and factor analysis); ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (38, 10, 5 and 3). The MLAs were trained on the 1st dataset and tested in the 2nd and 3rd datasets.

Analysis on patient sample datasets: The MLAs trained in cell lines (1st dataset) were tested against breast cancer patient samples from the 3rd dataset. The factorial design was executed using the following factors and levels: Gene selection approach (transmembrane, surfaceome, and factor analysis); ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (10, 5, and 3).

Validation on patient sample datasets: The MLAs were trained and tested with breast cancer patient samples from the 3rd dataset using the same genes of the previous analyses. The factorial design was executed using the following factors and levels: Gene selection approach (transmembrane, surfaceome, and factor analysis); ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (10, 5, and 3); type of quantification (Absolute counts and Fragments Per Kilobase Million (FPKM)).

Analysis of the combined genes approach: The best three genes from the initial approaches were combined, and their capacity for classifying patient samples was evaluated. The heatmap plots and hierarchical clustering were performed in the TBRCS with the following parameters: Transpose: Yes; Scale: None; Row Reordering: Dendrogram; Column Reordering: Dendrogram; Transform Table to LN: No. Then, the factorial design was executed using the following factors and levels: ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (9 and 5); type of quantification (Absolute counts and FPKM). The MLAs were trained and tested in the 3rd dataset.

Validation of the combined genes approach: One independent dataset (4th dataset) was used to corroborate the performance of the MLAs for the classification of breast cancer patient samples. The factorial design was executed using the following factors and levels: ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (9 and 5); type of quantification (Absolute counts and FPKM). The MLAs were trained and tested in the 4th dataset, which was obtained from The Cancer Genome Atlas (TCGA).

Comparison with network analysis approach: The classification performance of genes selected using network analysis was compared against the ones from the combined approach. The factorial design was executed using the following factors and levels: Gene selection approach (Combined and network analysis); ML algorithm (CZR, RF, SVM, LDA, and BYS); the number of genes (9 and 5). The MLAs were trained and tested in the 4th dataset.

The factors and the response variables used in the computational methodology are presented in Table 5.

Table 5

Factors and response variables used in the computational methodology
Factor	Level
Classification strategy	Molecular categories
Classification strategy	Molecular subcategories
Gene selection approach	Factor analysis
	Transmembrane genes
	Surfaceome genes Combined genes Network analysis
Machine Learning Algorithm	Control ZeroR
	Random Forest
	Support Vector Machine
	Linear Discriminant Analysis
	Naive Bayes
Number of genes	38
	10 or 9
	5
	3
Type of quantification	Absolute counts
Type of quantification	Fragments Per Kilobase Million
Algorithm performance (response variable)	Accuracy (%)

Machine learning algorithms

One control algorithm and five supervised MLAs (representative of the main types of MLAs) were used for the classification of the breast cells in molecular categories using transcriptomic data: Control Zero R (CZR), Random Forest (RF), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA) and Naive Bayes (BYS). CZR is a simple algorithm used to determine the baseline performance. It ignores all predictors and employs the proportion of the majority category (20). The classification analyses were performed in the software Weka (Waikato Environment for Knowledge Analysis) 3.8.3, following the data mining recommendations in (20). LibLINEAR (21) and discriminantAnalysis WEKA packages were installed for SVM and LDA, respectively. A 10-fold cross-validation technique was applied to the factorial designs where the training and testing of the MLAs took place on the same dataset: Initial analysis, validation on patient samples datasets, analysis of the combined genes approach, validation of the combined genes approach, and comparison with network analysis approach. In the supplementary materials are presented (the setup parameters of the Machine Learning Algorithms) (20, 21).

The following parameter was used to evaluate the performance of the MLAs on the classification of breast cancer: Accuracy. It is calculated as the percentage of correct predictions, which is shown in Eq. 1. The accuracy of the MLAs on the datasets was compared to F-measure (performance parameter based on precision and recall), and the average difference between them was lower than 1%.

Equation 1. Calculation of the accuracy

$$\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}= \frac{\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t} \text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{i}\text{o}\text{n}\text{s}}{\text{t}\text{o}\text{t}\text{a}\text{l} \text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{i}\text{o}\text{n}\text{s}}*100$$

Classification categories

During the initial analysis, cell lines were studied using two classification strategies: three molecular categories or six molecular subcategories (Table 6). These classifications are based on the study of Daemen et al. (22), which generated the data source of the 1st dataset. The non-malignant category is represented by cell lines similar to non-cancerous cells or patient samples from non-cancerous breast tissue. The second strategy is more challenging to implement in clinical practice, and the data available is more limited (especially for patient samples). Also, the accuracy showed no statistical difference between the strategies. For these reasons, the molecular categories were used in the following factorial designs.

Table 6

Classification strategies used in the study
Classification strategy	Categories
Molecular categories	Non-malignant (NM)
	Breast Cancer (BC)
	Triple Negative Breast Cancer (TNBC)
Molecular subcategories	Non-malignant
	Basal BC
	Luminal BC
	Basal TNBC
	Luminal TNBC
	Claudin-low TNBC

Statistical analyses

The statistical analyses were performed on the six factorial designs using JMP 14 software. The treatments were run with two replicates (using different random seeds when applicable), except for the 2nd validation, where each test against the validation dataset (2nd or 3rd) was considered replicate.

The means and standard deviations were calculated for each treatment and factor. Non-parametric tests were implemented (to determine if the differences between groups were statistically significant) to analyze the algorithm’s accuracy because the data presented distributions that were not normal. These tests were Wilcoxson (two groups), Kruskal-Wallis (more than two groups), and Steel-Dwass (multiple comparisons between more than two groups). Additional information about the statistical analyses is presented in the supplementary information (Additional information on the statistical analysis of each factorial design).

Kaplan-Meier plots

The Kaplan-Meier plots with the prognostic values of the combined genes approach (Fig. 17) were obtained from the RNA-seq breast cancer database with 2,976 patients on the KM plotter platform (https://kmplot.com/) using default parameters (23). The KIRREL1 gene was not plotted because it was not found in the database. The Log Rank test was applied for each gene to determine if the survival distribution of patients with high expression was statistically different from those with low expression. Genes with differences in their survival distributions have a higher potential to be used as biomarkers of prognosis (23).

Venn diagram for the genes of the combined approach and the commercial panels

The Venn diagram was generated with the nine best genes from all the approaches and the list of genes of four commercial panels (supplementary information: List of the genes of the four commercial panels) in the Bioinformatics and Systems Biology platform of Ghent University (https://bioinformatics.psb.ugent.be/webtools/Venn/). The following parameters were used to generate the mentioned diagram: Venn Diagram Shape: Symmetric; Venn Diagram Fill: Colored.

Initial analysis

We selected 38 genes from three approaches: transmembrane, surfaceome, and factor analysis and the PCA graphs of these approaches are shown in Fig. 2, where it can be seen that the breast cancer (BC) category was the most differentiated from the other two categories. Thus, the non-malignant (NM) is seems more difficult to separate from triple negative breast cancer (TNBC), which indicates that these two cell line categories have more similarities in their expression patterns. The PCA results suggest that the 38 genes of each approach will enable the separation of at least the two molecular categories of breast cancer (BC and TNBC) efficiently, which is in line with the objectives of this study. The training and testing of the MLAs was carried out using 20 cell lines isolated from BC, 13 from TNBC, and 5 from NM samples (1st dataset).

The initial analysis for the classification of the cell lines into the 6 molecular subcategories (NM, Basal BC, Luminal BC, Basal TNBC, Luminal TNBC, Claudin-low TNBC), or the 3 molecular categories (NM, BC and TNBC) is shown as the accuracy of each method in Fig. 3. According to these results, the MLAs showed better accuracy than the control algorithm (CZR) in 23 of the 24 treatments. 22 MLA treatments had an average accuracy of around 70% or higher. Transmembrane and Surfaceome approaches had better performance than Factor analysis. SVM was the MLA with the highest accuracy. Also, the two classification strategies had similar performance.

The statistical evaluation of the factors studied in the initial analysis using the percentage of accuracy found differences among approaches and MLAs but not between categories versus subcategories (Fig. 4C). For the gene selection approach, statistical differences were found between Factor analysis and both Surfaceome (P value < 0.05) and Transmembrane (P value < 0.01) (Fig. 4A). For the algorithm, all the MLAs showed statistical differences (P value < 0.01) against the control (CZR) (Fig. 4B).

Validation on independent cell lines datasets

We carry out the validation on independent cell lines datasets for the three initial approaches using two independent datasets (2nd and 3rd ) to corroborate the performance of the MLAs for the classification of breast cancer cell lines (described in the methods section). According to these results, the MLAs showed better accuracy than the control algorithm (CZR) in 45 of the 48 treatments (Fig. 5). These 45 MLA treatments had an average accuracy percentage of around 70% or higher. All the approaches had similar performance. SVM tended to show the highest accuracy. A decrease in accuracy is observed when the number of genes is reduced. However, 11 of the 12 MLA treatments using three genes have an average accuracy percentage equal to or greater than 70%. It suggests that the number of genes can be considerably reduced from 38 to 3, but the classification performance is still acceptable (accuracy ≥ 70%).

The statistical evaluation of the factors studied in the validation on independent cell lines datasets using the percentage of accuracy is shown in Fig. 5. No statistical differences between the gene selection approaches were found (Fig. 6A). All the MLAs showed statistical differences (P value < 0.01) compared to the control (CZR) (Fig. 6B). Statistical differences were found between the 38/10 gene groups versus five and ten gene groups (P value < 0.05) (Fig. 6C).

Analysis on patient samples datasets

We carry out the analysis on patient samples datasets for the three initial approaches using patient samples from 10 BC, 10 TNBC and 5 NM tissues, as well as 9 BC and 7 TNBC cell lines (3rd dataset). According to the analysis of the accuracy of each treatment (Fig. 7), the MLAs showed better accuracy than the control algorithm (CZR) in 32 of the 36 treatments. None of the MLA treatments had an average accuracy percentage equal to or higher than 70%, which is expected because of the significant differences between cell lines and patient samples. It indicates that the MLA trained with cell lines could not classify the patient samples in the stated molecular categories with acceptable accuracy. All the approaches had similar performance, but the surfaceome was the only one with treatments having an accuracy higher than 60%. BYS and LDA were the MLAs with the highest accuracy.

The statistical evaluation of the factors studied in the analysis on patient samples datasets using the percentage of accuracy is shown in Fig. 7. No statistical difference between the type of the gene selection approach was found (Fig. 8A). For the algorithms, all the MLAs showed statistical differences (P value < 0.01) compared to the control (CZR) (Fig. 8B). For the number of genes, a statistical difference between the groups of three and ten genes (P value < 0.01) was found (Fig. 8C), with three genes showing higher accuracy than 10 genes.

Validation on patient samples datasets

For the validation on patient samples datasets, the MLAs were trained and tested with breast cancer patient samples from the 3rd dataset using the same genes of the previous analyses (Fig. 9). According to these results, the MLAs showed better accuracy than the control algorithm (CZR) in all 72 treatments. Of these, 67 MLA treatments had an average accuracy percentage around 70% or higher. In general, absolute counts and FPKM had similar performance.

The statistical evaluation of the factors studied in the validation on patient samples datasets using the percentage of accuracy is shown in Fig. 10). For the gene selection approach, statistical differences were obtained when comparing Factor analysis and Surfaceome (P value < 0.01) and Factor analysis compared to Transmembrane (P value < 0.01) (Fig. 10A). All the MLAs showed statistical differences (P value < 0.01) compared to the control (CZR) (Fig. 10B). No statistical differences were found between the levels of the number of genes (Fig. 10C). No statistical differences between the levels of the type of quantification were found (Fig. 10D).

Analysis of the combined genes approach

For the analysis of the combined genes approach, the best three genes from the initial approaches were combined in one group, and their capacity for classifying patient samples was evaluated. Heat map plots and cluster dendrograms of these gene sets, for all the approaches, are shown in Fig. 11. According to this, the expression levels of some genes, such as TFF3, are different in the samples of each molecular category. As is shown on the horizontal axis, the hierarchical clustering algorithm puts together the patient samples of each molecular category in both gene groups (9 and 5). This indicates that the reduction in the number of genes from nine to five did not affect the capabilities of the mentioned clustering. The main disadvantage of using a heat map method is that it cannot be trained, so the classification performance is usually lower than the MLAs.

The graphs with the accuracy of each treatment are presented in Fig. 12. The MLAs showed better accuracy than the control algorithm (CZR) in all 16 treatments. These treatments had an average accuracy percentage equal to or above 90%. LDA was the MLA with the highest accuracy (100%). The two gene groups (5 and 9) had similar performances. Also, the absolute counts had similar accuracy than FPKM. However, using absolute counts has the advantage of requiring one less step in its calculation than FPKM.

The statistical evaluation of the factors studied in the analysis of the combined genes approach using the percentage of accuracy is shown in Fig. 13. For the algorithm, all the MLAs showed statistical differences (P value < 0.01) against the control (CZR) (Fig. 13A). No statistical difference was found between the levels of the number of genes (Fig. 13B). No statistical difference between the levels of the type of quantification was found (Fig. 13C). Because the obtention of the absolute counts requires fewer steps than the FPKM and both obtained similar performance, the use of absolute counts can be easier to implement in breast cancer diagnosis.

Validation of the combined genes approach

For the validation of the combined genes approach we used one independent dataset (4th dataset: 52 BC, 50 TNBC, and 11 NM samples) to corroborate the performance of the MLAs for the classification of breast cancer patient samples. The graphs with the accuracy of each treatment are presented in Fig. 14. According to these results, the MLAs showed better accuracy than the control algorithm (CZR) in all 16 treatments. All MLA treatments had an average accuracy percentage equal to or above 83%. RF was the MLA with the highest accuracy (94.7%). The two gene groups (5 and 9) and the types of quantification (absolute and FPKM) had similar performances.

The statistical evaluation of the factors studied in the validation of the combined genes approach using the percentage of accuracy is shown in Fig. 15. For the algorithm, all the MLAs showed statistical differences (P value < 0.01) against the control (CZR) (Fig. 15A). For the type of quantification, no statistical difference between absolute counts and FPKM was found (Fig. 15B). No statistical difference was found between the levels of the number of genes (Fig. 15C). Therefore, the group of five genes is a better option, considering that a panel with fewer biomarkers is easier to implement in diagnostic device development, which is especially relevant for cancer detection in places with limited access to healthcare.

Comparison with network analysis approach

For the comparison with network analysis approach, the classification performance of genes selected using network analysis was compared against the ones from the combined approach. The graphs with the accuracy of each treatment are presented in Fig. 16. According to these results, the MLAs showed better accuracy than the control algorithm (CZR) in all 16 treatments. All MLA treatments had an average accuracy close to 90%. RF and LDA were the MLA with the highest accuracy. The two gene groups (5 and 9) had similar performances. The combined genes approach had similar performance in comparison to the network analysis approach.

The statistical evaluation of the factors studied in the validation of the combined genes approach using the percentage of accuracy is shown in Fig. 17. For the gene selection approach, a statistical difference (P value < 0.01) was found (Fig. 17A). The network analysis was better than the combined approach; however, the performance of the MLAs is very good in both approaches (accuracy > 85%). No statistical difference was found between the levels of the number of genes (Fig. 17B).

The name, symbol, function, and implications in cancer of the best genes of all the approaches are presented in Table 7. All the genes except TMEM210 have reports indicating their involvement in breast cancer and other types of cancer. The transmembrane genes TEMEN210 and TMEM86A are not well studied, so the information about their function and participation in cancer is limited. On the other hand, CD44 and TENM4 have been proposed as biomarker candidates in the AACR Project GENIE (24). Also, the nine genes’ functions are heterogeneous, but three (TENM4, LRFN2, and BCAS1) are involved in neural development function.

Table 7

Information about the best nine selected genes of all the approaches
Gene selection approach	Gene symbol	Gene name	Function (25)	Implication in cancer
Transmembrane	TMEM210	Transmembrane protein 210 [HGNC:34059]	A protein that was shown to be expressed in the acrosome during the formation of elongated/late spermatids (26)	Not reported
	TENM4	Teneurin transmembrane protein 4 [HGNC:29945]	Involved in neural development, regulating the establishment of proper connectivity within the nervous system	It is altered in breast invasive ductal carcinoma, invasive breast carcinoma, lung adenocarcinoma, adenocarcinoma of unknown primary, and urothelial bladder carcinoma (24)
	TMEM86A	Transmembrane protein 86A [HGNC:26890]	Enables protein binding	Reported as one of the 200 most expressed genes by microarray analysis in breast cancer for PAM50 and Immunohistochemistry-based subtypes (27)
Surfaceome	CD44	CD44 molecule (Indian blood group) [HGNC:1681]	Cell-surface receptor plays a role in cell-cell interactions, cell adhesion, and migration, helping them to sense and respond to changes in the tissue microenvironment (28–30)	It is involved in breast cancer as well as prostate cancer progression (31)
	KIRREL1	Kin of IRRE like 1 (Drosophila) [HGNC:15734]	Required for the proper function of the glomerular filtration barrier (32)	It is involved in breast and gastric cancer progression (33)
	LRFN2	Leucine-rich repeat and fibronectin type III domain containing 2 [HGNC:21226]	Promotes neurite outgrowth in hippocampal neurons	It is involved in ovarian and breast cancer progression (34)
Factor Analysis	SPDEF	SAM pointed domain containing ETS transcription factor [HGNC:17257]	It may function as an androgen-independent transactivator of the prostate-specific antigen (PSA) promoter	Higher expression of this protein has been reported in brain, breast, lung, and ovarian tumors compared to the corresponding normal tissues (35,36)
	BCAS1	Brain-enriched myelin-associated protein 1 [HGNC:974]	Required for myelination	Significant down-regulation in the expression of the BCAS1 gene results in colorectal cancer. A splice variant of BCAS1 promotes the development of glioblastoma multiforme by suppressing the β-arrestin-2 pathway (37,38)
	TFF3	Trefoil factor 3 [HGNC:11757]	Involved in the maintenance and repair of the intestinal mucosa	It was one of the most underexpressed genes in papillary thyroid carcinoma. It can indicate tumor immune cell infiltration (39,40)

We used the Kaplan-Meier plots with the prognostic values of the combined genes approach to calculate the survival distributions (Fig. 18). All plotted genes have a P value (Log Rank test) lower than 0.05, indicating a statistically significant difference in survival between low and high expression levels. Three genes (SPDEF, BCAS, and TFF3) of the group of five genes (Fig. 18A) and two (TMEM210 and TENM4) of the rest of the group of nine genes (Fig. 18B) have a P value lower than 0.01, so they have a higher prognostic value.

The Oncotype DX Breast Recurrence Score Test (Genomic Health) analyzes, by quantitative TaqMan RT-PCR, the activity of 16 cancer-related genes and five control genes selected from a list of 250 candidate genes from a study of three independent cohorts totaling 447 patients (41). This test predicts the progression of early-stage, hormone-receptor-positive, and HER2-negative breast cancer and its response. Combined with other features of cancer, it can help the patient/doctor better evaluate whether or not to apply chemotherapy (42).

Prosigna® (PAM50) (Breast cancer Gene Signature Assay), a prognostic gene expression assay for breast cancer, simultaneously measures the expression of 50 genes, including additional five housekeeper genes, using RT-qPCR but runs in the proprietary nCounter DX analysis system developed by Nanostring (43). The genes were selected from a microarray and RT-qPCR data study from 189 prototype samples and tested on the prognosis of 761 patients and the prediction of pathological response to treatment in 133 patients (44). This panel identifies the intrinsic subtypes of tumors, particularly the differentiation between the luminal subtypes Luminal A and B, and the Risk of Recurrence (ROR). Wallden et al. (43) tested this platform in 2015 on 514 formalin-fixed, paraffin-embedded (FFPE) breast cancer patient samples after a previous training of the PAM50 algorithm using 304 FFPE patient samples from a well-annotated clinical cohort. They found that patients assigned to a Luminal A subtype would have the best prognosis compared to the other three subtypes. The Prosigna ROR score model was verified to be significantly associated with prognosis (43).

EndoPredict (EP; Myriad Genetics, Cologne, Germany) is a panel of 8 cancer-related and four control genes analyzed by RT-qPCR for FFPE breast cancer tissue samples (45, 46). The commercial test is available as EPclin, which considers clinical variables such as tumor stage and nodal status and generates a yes/no answer to recommend adjuvant chemotherapy and provides the risk of late recurrences (EPclin low-risk < 3.3, EPclin high-risk ≥ 3.3) (47). EPclin was clinically validated in randomized Phase III trials of 1700 post-menopausal BC patients treated with endocrine therapy alone, demonstrating that EP is prognostic for early (years 1–5) and late (years 5–10) distant recurrences in node-negative and node-positive disease (48).

The Breast Cancer Index (BCI, BioTheranostics) selected seven genes and the ratio between HOXB13 and IL17BR to predict late recurrence after treatment (49, 50). Five genes related to cell cycle regulation were chosen from a study of two publicly available microarray data sets implemented in 410 patients whose expression correlates with tumor grade and stage progression. It can assist the treating physician in determining if the patient should extend endocrine therapy in early-stage HR + breast cancer. Ma et al. (2008) (49) did a validation study with two cohorts of 323 patients, using RT-qPCR assay to analyze formalin-fixed paraffin-embedded clinical samples. They found that the combination of the expression analysis (molecular grade index or MGI) and the HOXB13:IL17BR ratio identifies a subgroup (∼30%) of early-stage estrogen receptor-positive breast cancer patients with very poor outcomes despite endocrine therapy. Sgroi et al. (2013) (50) found that a high HOXB13:IL17BR ratio could predict a higher risk for late recurrence in ER-positive patients who were disease-free after five years of tamoxifen when letrozole therapy was not extended, while high HOXB13:IL17BR ratio could predict the decreased probability of late disease recurrence when extended endocrine therapy with letrozole was prescribed.

The expression pattern and ratio have been retrospectively validated using samples from the Arimidex, Tamoxifen, Alone or in Combination (ATAC) Trial, the Adjuvant Tamoxifen-To Offer More? (aTTom) trial, as well as the Investigation of the Duration of Extended Letrozole (IDEAL) trial (51), which is now recommended for use in the 2021 National Comprehensive Cancer Network (NCCN) guideline (52).

MammaPrint (Agendia, The Netherlands) uses FFPE or fresh breast cancer tissue to carry out an RNA microarray assay, evaluating the expression pattern of 70 genes, which was initially validated by a cohort of 295 node-negative and node-positive breast cancer patients (47). Even though it can be used to guide clinical decisions as a prognostic tool in HR-positive, HER2-negative breast cancer with up to 3 positive nodes, it has proven not to be cost-effective and not very useful as a prospective tool to predict the benefits of chemotherapy in any of its risk groups.

To date, Prosigna, Endopredict, and Breast Cancer Index are recommended by several guidelines as prognostic tools in HR-positive, HER2-negative breast cancer with up to 3 positive nodes, including the NCCN.

Sestak et al. (2018) (53) carried out a within-patient comparison of the prognostic value of 6 multigene signatures in 774 post-menopausal women with ER-positive, ERBB2-negative breast cancer patients who received endocrine therapy for five years. They found that the signatures providing the most prognostic information were the PAM50-based Prosigna risk of recurrence, followed by the Breast Cancer Index and EndoPredict-EPclin, showing more information than the Clinical Treatment Score, the Oncotype Dx recurrence score, and the 4-marker immunohistochemical score. All six molecular tests provided substantially less information for the 183 patients with 1 to 3 positive nodes, but the BCI and EPclin provided more additional prognostic information than the other signatures (53).

Buus et al. (2021) (54) conducted another study comparing the Oncotype DX Recurrence Score, PAM50-based Prosigna risk of recurrence, EndoPredict, and Breast Cancer Index to estimate the risk of distant recurrence for patients receiving endocrine therapy. They studied 785 women with ER-positive/ HER2-negative breast cancer without chemotherapy treatment. They found that there were moderate to strong correlations among the four molecular scores (ρ = 0.63–0.74) except for the Oncotype DX Recurrence Score versus both Prosigna risk of recurrence (ρ = 0.32) and Breast Cancer Index (ρ = 0.35). Most of EndoPredict’s and Breast Cancer Index’s variation accounted for by the proliferation module (50.0% and 54.3%, respectively) and much less by the estrogen module (20.2% and 2.7%, respectively) (54).

Fewer studies focus on TNBC, presumably because the indication for adjuvant chemotherapy is less uncertain in most cases. Still, several gene expression studies have helped increase the understanding of this heterogeneous disease, which is not characterized by common features but rather by the absence of the three markers (Qian et al. 2021). For example, Lehmann et al. (2016) (55) have categorized TNBC into four subclasses: basal-like 1, 2, luminal androgen receptor, and mesenchymal subtypes.

When analyzing the genes in the Oncotype DX, PAM50, EndoPredict, Breast Cancer Index, and the top nine genes of the combined approach in this study (74 genes in total), only 13 were common among the different panels (Fig. 19). Also, none of the mentioned genes in this study were included in the mentioned commercial panels (Fig. 19). It shows that the diagnosis and prognosis of TNBC are difficult due to the lack of consistent gene expression panels. However, our analysis has proven to provide a set of genes that could have great potential as biomarkers. All the genes from the four commercial panel are described in supplementary information (List of the genes of the four commercial panels).

All gene selection approaches showed similar classification performance in the factorial designs. As a factor, the gene selection approach presented statistically significant differences between its levels in the initial analysis only. However, transmembrane and surfaceome genes that codify proteins are more likely to be present in cellular locations, which are accessible by capture antibodies or labeled molecules, making the application of diagnostic techniques and devices easier. Therefore, the genes selected by the transmembrane and surfaceome approaches could be useful for the development of cheaper diagnostic devices, which are required for vulnerable populations and developing countries. On the other hand, the combination of the best genes from the three approaches showed higher performance than each approach alone.

The MLAs showed statistically significant differences compared to the control algorithm in all the factorial designs, indicating that MLAs could use patterns in the transcriptomic data to increase the classification performance. Most of the MLAs obtained an acceptable accuracy (> 70%) except in the analysis on patient samples datasets, in which the algorithms were trained on cell lines and tested on patient samples, which is expected considering the significant differences between these types of samples. Some MLAs worked better in specific conditions and datasets, but no clear pattern showed that one MLA was significantly better than the others. On the other hand, the classification strategy showed non-significant differences between its levels in the initial analysis.

The number of genes presented significant differences between its levels in two of the five factorial designs (validation on independent cell lines datasets and analysis on patient samples datasets) where it was analyzed. However, in most cases, even the smallest number of genes had an acceptable accuracy. It was very relevant because it allowed reducing the number of genes without compromising the classification performance. As a result, the panel could be reduced to the top five genes in the combined approach.

On the other hand, the analysis of the type of quantification demonstrated that the absolute expression counts had a similar performance to FPKM. However, absolute counts are not always comparable between samples or datasets because the sequencing depths or library sizes can generate biases in the sequenced data (56), so it is better to use relative expression counts such as FPKM to eliminate possible biases. The results presented in this study suggest that the classification performance of the MLAs was not affected by the mentioned biases. In addition, the advantage of absolute counts is that they require fewer steps in the analysis, so the implementation in diagnostic procedures could be easier.

Although the performance of the combined genes approach is slightly inferior in comparison to the network analysis approach, the first one is preferable in a scenario with limited resources because this approach requires much less computational power.

After searching the Web of Science and Scopus databases for research articles about sensors or biosensors using any of the nine genes of the combined approach, CD44 and TFF3 were the genes that have been reported. CD44 has been applied in cancer detection and immune analysis (57–59), while TFF3 has been applied in an inflammatory response, kidney disease, and cancer (60–62). These results indicate that there is potential for the development of biosensors using the methodologies in this study; however, the majority of the combined approach genes have not been studied with a biosensor perspective. For this reason, experimental studies are required to validate the utility of the rest of the genes for biosensor implementation.

African, Hispanic, and Asian populations are underrepresented in the datasets analyzed in this study. The data availability and technical capabilities limited the number of analyzed patient samples, so testing in larger patient datasets is recommended. All the analyses were performed using public data, so the quality and reliability of the data depended on the policies implemented by the public repositories with transcriptomic data (GEO and TCGA). No experimental validations were performed in the present study, since the main goal here was to provide an efficient Machine Learning analytical method that can help identify new biomarkers.

Experimental and clinical validations are required to confirm that the selected biomarkers and the applied machine learning methods are adequate for early detection and diagnosis of breast cancer in the medical practice. A study focused on detecting and quantifying the selected biomarkers on breast cancer patient samples using RT-PCR, and proteomic techniques are recommended. A diagnostic clinical trial using the selected biomarkers can be performed if the results of the previous study are consistent. Moreover, the development of low-cost biosensors or diagnostic devices can be considered in which early breast cancer diagnostic accuracy is not compromised to reach the wide range of populations of different genetic backgrounds and for different socioeconomic groups.

This study implemented a methodology of machine learning algorithms (MLAs) capable of finding patterns in transcriptomic data with promising results in the classification of breast cancer for the development of low-cost diagnostic devices. Furthermore, it describes an extensive and unique integrative study that involves the classification of breast cancer cells in three groups (non-malignant, non-triple negative, and triple-negative) using four different MLAs and two types of samples (cell lines and patient samples). Factorial designs and non-parametric statistical tests were performed to enrich the analysis and support the validation. Therefore, the results obtained are more reliable and closer to a real situation in comparison to other similar studies. Also, the initial and combined gene selection approaches used algorithms that require low computational power. The results from these approaches were similar to a network analysis algorithm, which required a supercomputer. Using low-cost algorithms decreases the development cost and is valid for scenarios with limited resources frequently occurring in developing countries.

Most of the MLAs showed an accuracy higher than 80% in all the experimental designs except in the analysis on patient samples datasets, where the performance decreased considerably, which was expected because of the significant differences between the cell lines and the patient samples. The results showed that reducing the number of genes decreased the algorithm performance, but it was maintained in an acceptable range (accuracy ≥ 70%). The methodology presented here provides evidence that MLAs are useful for the selection of biomarkers and the classification of breast cancer.

Two panels (9 and 5 genes) of the combined approach performed better than the genes selected by each of the initial approaches, highlighting the importance of enriching the diversity of genes by combining different selection approaches. From these two panels, using five genes seems better because it had a similar performance to nine genes; however, using fewer genes would be easier to implement in the development of diagnostic devices.

Eight of the best nine genes of all the approaches were previously reported to be implicated in cancer. Also, these nine genes were utterly different from those present in the four commercial panels analyzed in the current study panels (Oncotype DX, PAM50, EndoPredict, Breast Cancer Index). The Log Rank test had a P value < 0.01 in five genes of the mentioned group (SPDEF, BCAS, TFF3, TMEM210, and TENM4), so they could be useful in breast cancer prognosis.

The methodology and the biomarkers (genes) selected in this study showed promising characteristics with the potential to be implemented in the future for low-cost early detection and diagnosis of breast cancer. Our methodology was selected to reach a wide population range of different backgrounds without compromising the performance of the diagnostic tools. However, we anticipate that promoting more collaborative and multidisciplinary research in machine learning will contribute to the inclusion of communities from different disciplines in the machine learning areas.

BC: Breast Cancer (non-triple negative)

BYS: Naive Bayes

CZR: Control Zero R

ER: Estrogen receptor

FPKM: Fragments Per Kilobase Million (fragments per kilobase of transcript per million reads mapped)

GEO: Gene Expression Omnibus

HER2: Human Epidermal Growth Factor Receptor 2

LDA: Linear Discriminant Analysis

MDA: Mean Decrease Accuracy

MLA: Machine Learning Algorithm

NM: Non-malignant

PCA: Principal Component Analysis

PR: Progesterone receptor

RF: Random Forest

SVM: Support Vector Machine

TBRCS: Tauber Bioinformatics Research Center Server

TCGA: The Cancer Genome Atlas

TNBC: Triple Negative Breast Cancer

Ethical Approval

Not applicable.

Funding

Not applicable.

Availability of data and materials

Data used in this study are accessible in the public data sources mentioned in the Methods section (Table 2). The analyzed datasets are available at https://zenodo.org/record/7886643.

Acknowledgements

Goldie Oza (CIDETEQ, CONACYT) is acknowledged for his support in the discussion section, while Juan Carlos Verduzco (School of Materials Engineering, Purdue University) for his support in running the Network Analysis Approach. Kalaumari Mayoral-Peña appreciates the CONACYT scholarship funding for his PhD studies.

Author information

Authors and Affiliations

School of Engineering and Sciences, Campus Queretaro, Tecnologico de Monterrey, Queretaro, Mexico

Kalaumari Mayoral-Peña and Marcos de Donato

Department of Medicine, Division of Engineering in Medicine, Brigham and Women’s, Hospital Harvard Medical School, Boston, MA, USA.

Natalie Artzi and Kalaumari Mayoral-Peña

Evidence-Based Medicine Research Unit, Children’s Hospital of Mexico Federico Gómez, National Institute of Health, Mexico City 06720, Mexico

Omar Israel González Peña

Education and Research Department, Hospital Clínica Nova de Monterrey, San Nicolas de los Garza, Mexico

Omar Israel González Peña

International University of La Rioja, Logrono La Rioja, Spain.

Omar Israel González Peña

School of Engineering and Sciences, Campus Monterrey, Tecnologico de Monterrey, Monterrey, Mexico

Omar Israel González Peña

Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, USA

Natalie Artzi

Wyss Institute for Biologically Inspired Engineering Harvard University Boston, MA, USA

Natalie Artzi

Contributions

Conceptualization: K.M.-P. and M.d.D Methodology: K.M.-P., M.d.D, and O.I.G.P. Validation, K.M.-P. Formal analysis, K.M.-P. Investigation: K.M.-P., M.d.D., O.I.G.P., and N.A Data curation, K.M.-P, O.I.G.P. Writing—original draft preparation, K.M.-P. Writing—review and editing, K.M.-P., M.d.D., O.I.G.P., and N.A. Visualization, K.M.-P. Supervision, O.I.G.P. and M.d.D Project administration, O.I.G.P. and M.d.D.

Corresponding author

Correspondence to Marcos de Donato and Omar Israel González Peña

Ethics declarations

Competing interests

The Authors declare no Competing Interests.

McKinney SM, Sieniek M, Godbole V, Godwin J, Antropova N, Ashrafian H, et al. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89–94.
Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians. 2021;71(3):209–49.
Ahmad A. Breast cancer statistics: recent trends. Breast Cancer Metastasis and Drug Resistance. 2019;1–7.
Britt KL, Cuzick J, Phillips KA. Key steps for effective breast cancer prevention. Nature Reviews Cancer. 2020;20(8):417–36.
Francies FZ, Hull R, Khanyile R, Dlamini Z. Breast cancer in low-middle income countries: abnormality in splicing and lack of targeted treatment options. 2020;
Ginsburg O, Yip C, Brooks A, Cabanes A, Caleffi M, Dunstan Yataco JA, et al. Breast cancer early detection: A phased approach to implementation. Cancer. 2020;126:2379–93.
Malhotra GK, Zhao X, Band H, Band V. Histological, molecular and functional subtypes of breast cancers. Cancer biology & therapy. 2010;10(10):955–60.
Tsang J, Tse GM. Molecular classification of breast cancer. Advances in anatomic pathology. 2020;27(1):27–35.
Medina MA, Oza G, Sharma A, Arriaga L, Hernández Hernández JM, Rotello VM, et al. Triple-negative breast cancer: a review of conventional and advanced therapeutic strategies. International journal of environmental research and public health. 2020;17(6):2078.
Alzubi J, Nayyar A, Kumar A. Machine learning from theory to algorithms: an overview. In IOP Publishing; 2018. p. 012012.
Yue W, Wang Z, Chen H, Payne A, Liu X. Machine learning with applications in breast cancer diagnosis and prognosis. Designs. 2018;2(2):13.
Zhai Q, Li H, Sun L, Yuan Y, Wang X. Identification of differentially expressed genes between triple and non-triple-negative breast cancer using bioinformatics analysis. Breast Cancer. 2019;26(6):784–91.
Kothari C, Osseni MA, Agbo L, Ouellette G, Déraspe M, Laviolette F, et al. Machine learning analysis identifies genes differentiating triple negative breast cancers. Scientific reports. 2020;10(1):1–15.
Zhao Y, Wong L, Goh WWB. How to do quantile normalization correctly for gene expression data analyses. Scientific reports. 2020;10(1):1–11.
Kursa MB. Robustness of Random Forest-based gene selection methods. BMC bioinformatics. 2014;15:1–8.
Risso D, Ngai J, Speed TP, Dudoit S. Normalization of RNA-seq data using factor analysis of control genes or samples. Nature biotechnology. 2014;32(9):896–902.
Ren J, He T, Li Y, Liu S, Du Y, Jiang Y, et al. Network-based regularization for high dimensional SNP data in the case–control study of Type 2 diabetes. BMC genetics. 2017;18(1):1–12.
Ren J, Du Y, Li S, Ma S, Jiang Y, Wu C. Robust network‐based regularization and variable selection for high‐dimensional genomic data in cancer prognosis. Genetic epidemiology. 2019;43(3):276–91.
Bausch-Fluck D, Goldmann U, Müller S, van Oostrum M, Müller M, Schubert OT, et al. The in silico human surfaceome. Proceedings of the National Academy of Sciences. 2018;115(46):E10988–97.
Witten IH, Frank E, Hall MA. Data mining: practical machine learning tools and techniques with Java implementations. 3rd ed. Morgan Kaufmann Publishers; 2011.
Fan RE, Chang KW, Hsieh CJ, Wang XR, Lin CJ. LIBLINEAR: A library for large linear classification. the Journal of machine Learning research. 2008;9:1871–4.
Daemen A, Griffith OL, Heiser LM, Wang NJ, Enache OM, Sanborn Z, et al. Modeling precision treatment of breast cancer. Genome biology. 2013;14(10):1–14.
Lánczky A, Győrffy B. Web-based survival analysis tool tailored for medical research (KMplot): development and implementation. Journal of Medical Internet Research. 2021;23(7):e27633.
AACR Project Genie Consortium, AACR Project GENIE Consortium, André F, Arnedos M, Baras AS, Baselga J, et al. AACR Project GENIE: powering precision medicine through an international consortium. Cancer discovery. 2017;7(8):818–31.
Zahn-Zabal M, Michel PA, Gateau A, Nikitin F, Schaeffer M, Audot E, et al. The neXtProt knowledgebase in 2020: data, tools and usability improvements. Nucleic Acids Research. 2020;48(D1):D328–34.
Pineau C, Hikmet F, Zhang C, Oksvold P, Chen S, Fagerberg L, et al. Cell type-specific expression of testis elevated genes based on transcriptomics and antibody-based proteomics. Journal of proteome research. 2019;18(12):4215–30.
Schettini F, Barbao P, Brasó-Maristany F, Galván P, Martínez D, Paré L, et al. Identification of cell surface targets for CAR-T cell therapies and antibody–drug conjugates in breast cancer. ESMO open. 2021;6(3):100102.
Vikesaa J, Hansen TV, Jønson L, Borup R, Wewer UM, Christiansen J, et al. RNA‐binding IMPs promote cell adhesion and invadopodia formation. The EMBO journal. 2006;25(7):1456–68.
Crosby HA, Lalor PF, Ross E, Newsome PN, Adams DH. Adhesion of human haematopoietic (CD34+) stem cells to human liver compartments is integrin and CD44 dependent and modulated by CXCR3 and CXCR4. Journal of Hepatology. 2009;51(4):734–49.
Yoshida T, Matsuda Y, Naito Z, Ishiwata T. CD44 in human glioma correlates with histopathological grade and cell migration. Pathology international. 2012;62(7):463–70.
Al-Othman N, Alhendi A, Ihbaisha M, Barahmeh M, Alqaraleh M, Al-Momany BZ. Role of CD44 in breast cancer. Breast disease. 2020;39(1):1–13.
Solanki AK, Widmeier E, Arif E, Sharma S, Daga A, Srivastava P, et al. Mutations in KIRREL1, a slit diaphragm component, cause steroid-resistant nephrotic syndrome. Kidney international. 2019;96(4):883–9.
Chen K, Zhao R, Yao G, Liu Z, Shi R, Geng J. Overexpression of kin of IRRE-Like protein 1 (KIRREL) as a prognostic biomarker for breast cancer. Pathology-Research and Practice. 2020;216(7):153000.
Verma S, Bakshi D, Sharma V, Sharma I, Shah R, Bhat A, et al. Genetic variants of DNAH 11 and LRFN 2 genes and their association with ovarian and breast cancer. International Journal of Gynecology & Obstetrics. 2020;148(1):118–22.
Cheng XH, Black M, Ustiyan V, Le T, Fulford L, Sridharan A, et al. SPDEF inhibits prostate carcinogenesis by disrupting a positive feedback loop in regulation of the Foxm1 oncogene. PLoS genetics. 2014;10(9):e1004656.
Fararjeh AS, Liu YN. ZBTB46, SPDEF, and ETV6: Novel potential biomarkers and therapeutic targets in castration-resistant prostate cancer. International Journal of Molecular Sciences. 2019;20(11):2802.
Kuo YH, Hung HS, Tsai CW, Chiu SC, Liu SP, Chiang YT, et al. A Novel Splice Variant of BCAS1 Inhibits β-Arrestin 2 to Promote the Proliferation and Migration of Glioblastoma Cells, and This Effect Was Blocked by Maackiain. Cancers. 2022;14(16):3890.
Prakash A, Taylor L, Varkey M, Hoxie N, Mohammed Y, Goo YA, et al. Reinspection of a clinical proteomics tumor analysis consortium (CPTAC) dataset with cloud computing reveals abundant post-translational modifications and protein sequence variants. Cancers. 2021;13(20):5034.
Yang L, Zhang X, Zhang J, Liu Y, Ji T, Mou J, et al. Low expression of TFF3 in papillary thyroid carcinoma may correlate with poor prognosis but high immune cell infiltration. Future Oncology. 2021;18(3):333–48.
Wang X, Qin H. TFF3 promotes pituitary tumor cell migration and angiogenesis via VEGFA. Acta Neurobiol Exp. 2022;82.
Paik S, Shak S, Tang G, Kim C, Baker J, Cronin M, et al. A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. New England Journal of Medicine. 2004;351(27):2817–26.
Davey MG, Richard V, Lowery AJ, Kerin MJ. OncotypeDX© Recurrence Score in BRCA mutation carriers: a systematic review and meta-analysis. European Journal of Cancer. 2021;154:209–16.
Wallden B, Storhoff J, Nielsen T, Dowidar N, Schaper C, Ferree S, et al. Development and verification of the PAM50-based Prosigna breast cancer gene signature assay. BMC medical genomics. 2015;8(1):1–14.
Parker JS, Mullins M, Cheang MC, Leung S, Voduc D, Vickery T, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. Journal of clinical oncology. 2009;27(8):1160.
Warf MB, Rajamani S, Krappmann K, Doedt J, Cassiano J, Brown K, et al. Analytical validation of a 12-gene molecular test for the prediction of distant recurrence in breast cancer. Future science OA. 2017;3(3):FSO221.
Filipits M, Rudas M, Jakesz R, Dubsky P, Fitzal F, Singer CF, et al. A New Molecular Predictor of Distant Recurrence in ER-Positive, HER2-Negative Breast Cancer Adds Independent Information to Conventional Clinical Risk FactorsA Molecular Predictor of Distant Recurrence in Breast Cancer. Clinical Cancer Research. 2011;17(18):6012–20.
Qian Y, Daza J, Itzel T, Betge J, Zhan T, Marmé F, et al. Prognostic cancer gene expression signatures: current status and challenges. Cells. 2021;10(3):648.
Dubsky P, Brase J, Jakesz R, Rudas M, Singer C, Greil R, et al. The EndoPredict score provides prognostic information on late distant metastases in ER+/HER2− breast cancer patients. British journal of cancer. 2013;109(12):2959–64.
Ma XJ, Salunga R, Dahiya S, Wang W, Carney E, Durbecq V, et al. A five-gene molecular grade index and HOXB13: IL17BR are complementary prognostic factors in early stage breast cancer. Clinical cancer research. 2008;14(9):2601–8.
Sgroi DC, Carney E, Zarrella E, Steffel L, Binns SN, Finkelstein DM, et al. Prediction of late disease recurrence and extended adjuvant letrozole benefit by the HOXB13/IL17BR biomarker. Journal of the National Cancer Institute. 2013;105(14):1036–42.
Noordhoek I, Treuner K, Putter H, Zhang Y, Wong J, Meershoek-Klein Kranenbarg E, et al. Breast Cancer Index Predicts Extended Endocrine Benefit to Individualize Selection of Patients with HR+ Early-stage Breast Cancer for 10 Years of Endocrine TherapyBCI Predicts Benefit from Extended Endocrine Therapy. Clinical Cancer Research. 2021;27(1):311–9.
Bartlett J, Sgroi D, Treuner K, Zhang Y, Ahmed I, Piper T, et al. Breast Cancer Index and prediction of benefit from extended endocrine therapy in breast cancer patients treated in the Adjuvant Tamoxifen—To Offer More?(aTTom) trial. Annals of Oncology. 2019;30(11):1776–83.
Sestak I, Buus R, Cuzick J, Dubsky P, Kronenwett R, Denkert C, et al. Comparison of the performance of 6 prognostic signatures for estrogen receptor–positive breast cancer: a secondary analysis of a randomized clinical trial. JAMA oncology. 2018;4(4):545–53.
Buus R, Sestak I, Kronenwett R, Ferree S, Schnabel CA, Baehner FL, et al. Molecular Drivers of Oncotype DX, Prosigna, EndoPredict, and the Breast Cancer Index: A TransATAC Study. J Clin Oncol. 2020;
Lehmann BD, Jovanović B, Chen X, Estrada MV, Johnson KN, Shyr Y, et al. Refinement of triple-negative breast cancer molecular subtypes: implications for neoadjuvant chemotherapy selection. PloS one. 2016;11(6):e0157368.
Zhao S, Ye Z, Stanton R. Misuse of RPKM or TPM normalization when comparing across samples and sequencing protocols. Rna. 2020;26(8):903–9.
Monteiro AC, Schmitz V, Morrot A, De Arruda LB, Nagajyothi F, Granato A, et al. Bradykinin B2 Receptors of dendritic cells, acting as sensors of kinins proteolytically released by Trypanosoma cruzi, are critical for the development of protective type-1 responses. PLoS pathogens. 2007;3(11):e185.
Bekmurzayeva A, Ashikbayeva Z, Myrkhiyeva Z, Nugmanova A, Shaimerdenova M, Ayupova T, et al. Label-free fiber-optic spherical tip biosensor to enable picomolar-level detection of CD44 protein. Scientific reports. 2021;11(1):19583.
Zhang R, Rejeeth C, Xu W, Zhu C, Liu X, Wan J, et al. Label-free electrochemical sensor for cd44 by ligand-protein interaction. Analytical chemistry. 2019;91(11):7078–85.
Massoner P, Kugler KG, Unterberger K, Kuner R, Mueller LA, Fälth M, et al. Characterization of transcriptional changes in ERG rearrangement-positive prostate cancer identifies the regulation of metabolic sensors such as neuropeptide Y. PloS one. 2013;8(2):e55207.
DeGregory PR, Tsai YJ, Scida K, Richards I, Crooks RM. Quantitative electrochemical metalloimmunoassay for TFF3 in urine using a paper analytical device. Analyst. 2016;141(5):1734–44.
Xue L, Aihara E, Podolsky DK, Wang TC, Montrose MH. In vivo action of trefoil factor 2 (TFF2) to speed gastric repair is independent of cyclooxygenase. Gut. 2010;59(9):1184–91.

No competing interests reported.

TableS1Listofgenesfromtheinitialselectionapproaches.docx
S1: List of genes of each one of the three initial approaches;
S2CodeandparametersoftheNetworkAnalysis.docx
S2: Setup parameters of the Machine Learning Algorithms;
TableS3SetupparametersoftheMachineLearningAlgorithmsinWeka.docx
S3: Code and parameters of the Network Analysis;
S4Additionalinformationonthestatisticalanalysisofeachfactorialdesign.docx
S4: Additional information on the statistical analysis of each factorial design;
TableS5Listofthegenesofthefourcommercialpanels.docx
S5: List of the genes of the four commercial panels.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of biomarkers for breast cancer early diagnosis based on the molecular classification using machine learning algorithms on transcriptomic data and factorial designs for analysis

Status:

Version 1

Abstract

Figures

Introduction

Methods

Trascriptomic data from cell lines and breast cancer patient biopsies

Data pre-processing

Gene selection approaches

Computational methodology

Machine learning algorithms

Equation 1. Calculation of the accuracy

Classification categories

Statistical analyses

Kaplan-Meier plots

Venn diagram for the genes of the combined approach and the commercial panels

Results

Initial analysis

Analysis on patient samples datasets

Validation on patient samples datasets

Analysis of the combined genes approach

Comparison with network analysis approach

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1