Aiming at the World's Most Common Cancer, Chinese Scholars Established the Breast Cancer Prognostic Scoring System MIRS
The latest global cancer burden data for 2020 released by the World Health Organization's International Agency for Research on Cancer (IARC) show that the number of new breast cancer cases in the world increased rapidly in 2020, reaching 2.26 million, officially replacing lung cancer as the world's number one cancer for the first time. Among them, the number of new breast cancer cases in Chinese women was 420,000, ranking first, far exceeding other types of cancer in women.
Due to its high complication rate and high mortality rate, breast cancer seriously threatens the health of women around the world. However, if it can be detected early and treated according to best practices, the survival rate is expected to be greatly improved. According to the American Cancer Society, the mortality rate of breast cancer has decreased by 40% between 1989 and 2016.
In recent years, artificial intelligence has made great progress in medical imaging, pathology, and decision-making support systems. It has formed three major research directions in the field of breast cancer pathology: nuclear division image detection at the cellular level, tumor area detection and segmentation at the regional level, and quantitative analysis of immunohistochemistry.
Tumor-infiltrating immune cells (TIICs) and tumor metastasis are important features of tumors in the human body.Researchers from the University of Kentucky, Macau University of Science and Technology, University of Macau, and the First Affiliated Hospital of Guangzhou Medical University used a neural network model to establish an accurate prognostic scoring system, MIRS (metastasis and immunogenomic risk score), for tumor metastasis and immunogenomic risk scoring.This provides a predictive tool that is almost universally applicable to breast cancer patients, providing new directions for treatment selection in the breast cancer population.
Research highlights:
* MIRS, a scoring system for predicting breast cancer prognosis and treatment, can be used to guide the formulation of treatment strategies for breast cancer patients
* This study analyzed the impact of metastasis and immune infiltration on breast cancer prognosis
* MIRS can provide guidance for different BRCA subtypes, and IVL has the highest expression level in TNBC subtype
The corresponding author of this article, Xiaohua Douglas Zhang, is a professor of biostatistics at the University of Kentucky. He holds a Ph.D. in statistics from Carnegie Mellon University, an EMBA in management from Guanghua School of Management at Peking University, a master's degree in genetics from Peking University, and a bachelor's degree in biology from Beijing Normal University. It is worth mentioning that he also worked at Merck for 14 years as a senior chief scientist.
Personal homepage:
https://cph.uky.edu/directory/xiaohua-zhang
Get the paper:
https://doi.org/10.1016/j.isci.2023.108322
Follow the official account and reply "MIRS" to get the complete PDF
Dataset: Selecting the Differentially Expressed Genes
This study first applied single sample gene set analysis (ssGSEA) to screen out the enrichment scores of 45 immune features related to the tumor microenvironment in breast cancer patients from the TCGA (The Cancer Genome Atlas) database.
Based on these data, the study further used hierarchical clustering to divide patients into high and low immune cell infiltration groups. Subsequently, the researchers identified 1,222 differentially expressed genes (DEGs) between the high and low immune cell infiltration groups through the Wilcoxon rank-sum test.
At the same time, in order to detect genes related to metastasis, the study also performed a Wilcoxon rank sum test between breast cancer metastasis patients and primary patients based on two GEO (Gene Expression Omnibus) cohorts, GSE10893 (n=18) and GSE3521 (n=75), and screened out 2159 differentially expressed genes (DEGs). After overlapping with the previous 1222 genes, it was found that there were 52 overlapping genes.
Based on these 52 overlapping genes, the researchers further screened 12 genes related to the patient's overall survival (OS) and established a neural network framework prognostic model for these 12 genes, namely MIRS. According to the MIRS score, the study ultimately divided the patients into MIRS-high and MIRS-low subtypes, and used the MIRS score to guide chemotherapy or immunotherapy.
Model Architecture: The neural network model showed the best predictive performance
The researchers further divided the 12 candidate genes screened from the 52 genes into protective genes and risk genes. For protective genes, a value of "0" is assigned when the gene expression status is higher than the sample average, and a value of "1" is assigned when it is lower than the average; for risk genes, a value of "1" is assigned when the gene expression status is higher than the sample average, and a value of "0" is assigned when it is lower than the average.
Subsequently, the researchers randomly divided the TCGA cohort (N = 1100) into training data and test data in a ratio of 7:3, and used four machine learning methods, including neural network (NN), statistical regression (LR), random forest (RF) and support vector machine (SVM), to establish a prognostic scoring system based on 12 candidate genes.
As shown in the figure below, ROC curve analysis shows that the neural network model shows the best prediction performance regardless of how the training set or test set changes.
On this basis, the study also used a classic neural network with a hidden layer to establish a prognostic model as shown in the figure below, defining the netn1 = W1,1i1 + W2,1i2 +…+ W12,1i12 + b1, where W is the weight of each input node and ij (j = 1,2…12) is the “0-1” state of the gene.
In the output layer, the study used Tensorflow and Keras to build a neural network, used ReLU as the activation function in the hidden layer, applied the Softmax function to the "survival" and "death" nodes of the output layer, used the cross entropy error as the loss function, and used the Adam algorithm to optimize the prognostic gene weights. After training, the coefficient of each prognostic gene is determined by the maximum weight of the hidden layer.
Here are two examples of MIRS calculations in action:
Research conclusion: MIRS-low subtype is more likely to metastasize, and chemotherapy is more beneficial to MIRS-high subtype
This study further investigated the correlation between the immune genome profiles of breast cancer patients and MIRS using the ESTIMATE algorithm. The results showed that the MIRS-low subtype had a higher proportion of immune cells and stromal cells, but lower tumor purity. This indicates that there are higher levels of tumor-infiltrating immune cells in the MIRS-low subtype.
In addition, the study also performed ssGSEA scores on 17 immune-related pathways between the two MIRS subtypes. The results showed that the immune infiltration level of the MIRS-low subtype was significantly higher than that of the MIRS-high subtype in almost all pathways.
In addition, the study also explored the correlation between MIRS scores and cancer metastasis. In the MIRS-low subtype, there were signs of upregulation of the activity of metastatic signaling factors, including hypoxia, TGF-β signaling, angiogenesis, and epithelial-mesenchymal transition (EMT) scores (Figures F and H below). At the same time, the researchers observed increased activity of macrophage M1 (Figure D below) and angiogenesis (Figure F below) in the MIRS-low subtype, which is consistent with previous views. In addition, MIRS was negatively correlated with angiogenesis marker genes (Figure G below).
In conclusion, the MIRS-low subtype with high immune infiltration may be more prone to metastasis.
Since the MIRS-low subtype has high infiltration of TIICs in the TME, theoretically, patients with the MIRS-low subtype should be sensitive to immune checkpoint blockade (ICB) therapy, and MIRS ultimately also showed that the MIRS-low subtype may be more sensitive to ICB therapy than the MIRS-high subtype.
and immunophenotypic score (IPS)
The study also analyzed the correlation between MIRS and chemotherapy. Survival analysis showed that the survival rate of MIRS-high subtype patients after chemotherapy was higher than that of MIRS-low subtype patients. In addition, the study also found that the survival rate of MIRS-high patients who received chemotherapy was significantly higher than that of MIRS-high patients who did not receive chemotherapy. These results indicate thatChemotherapy may be more beneficial for the MIRS-high subtype.
Triple-negative breast cancer subtypes: MIRS reveals potential prognostic gene targets
Among breast cancer subtypes, the treatment progress of triple-negative breast cancer (TNBC) still faces many challenges, and there is an urgent need to find more biomarkers that can improve prognosis. Compared with the MIRS-low subtype, the MIRS-high subtype has a lower survival rate, so it can be used as a candidate target gene for TNBC progression. There are 58 genes that are significantly overexpressed in the MIRS-high subtype.
Subsequently, the researchers used three machine learning algorithms, XGboost, Borota RF, and Elastic net lasso regression (ElasticNet), to select the most critical genes. Finally, 9 dominant genes were found in the Venn diagram of the four groups of genes. The survival curve showed that the expression level of IVL (Involucrin) affects the survival outcomes of different breast cancer subtypes.
Association with tumor subtype
In addition, IVL had the highest expression level in the TNBC subtype, while the expression level was lower in the BRCA subtype. Through the analysis of cell migration-related pathway markers, the cell migration pathways in the high IVL group were significantly enriched. These results suggest that IVL may be a potential target for exploring the prognosis of TNBC subtypes.
There is a significant METASCAPE pathway
Breast cancer diagnosis and treatment is still at the stage of weak artificial intelligence
Doctors play a vital role in the diagnosis and treatment of breast cancer. However, it takes a lot of time and effort to train a qualified breast surgeon. Today, the application of artificial intelligence provides more medical power for breast cancer, which is expected to reduce the high incidence and mortality of breast cancer.
However, current AI research on breast disease diagnosis and treatment is mainly focused on breast mammography, ultrasound, deep learning technology for pathological image data, as well as breast cancer diagnosis and treatment decision-making, hospital management and other branches and fields.There is a lack of top-level design for the integration of multiple deep learning models, that is, there is a lack of deep learning models that can update, store and analyze real information in real time, and integrate artificial intelligence image reading diagnosis, individualized treatment and risk prediction.
In order to accelerate the innovation of AI-related technologies and industrial development, my country proposed in the Outline of the Healthy China 2030 Plan issued in 2016 that the diagnosis and treatment of breast diseases should be moved forward, and AI will make great progress in the diagnosis and treatment of breast diseases in the future. The CSCO Breast Cancer Diagnosis and Treatment Guidelines 2021 even proposed that the expert group encourages the development of AI-related clinical research and the development of AI systems with independent intellectual property rights in my country.
In China, there are about 400,000 new breast cancer patients each year, of which more than 70% are from third- and fourth-tier cities, and the trend is younger and more urbanized. In recent years, local governments have continuously increased their investment in breast cancer screening for women, and more and more women have benefited from it. They also hope that artificial intelligence will be able to explore new solutions for cancer in the future.
References: