Artificial intelligence: opportunities and challenges in the clinical applications of triple-negative breast cancer

Abstract

Triple-negative breast cancer (TNBC) accounts for 15–20% of all invasive breast cancer subtypes. Owing to its clinical characteristics, such as the lack of effective therapeutic targets, high invasiveness, and high recurrence rate, TNBC is difficult to treat and has a poor prognosis. Currently, with the accumulation of large amounts of medical data and the development of computing technology, artificial intelligence (AI), particularly machine learning, has been applied to various aspects of TNBC research, including early screening, diagnosis, identification of molecular subtypes, personalised treatment, and prediction of prognosis and treatment response. In this review, we discussed the general principles of artificial intelligence, summarised its main applications in the diagnosis and treatment of TNBC, and provided new ideas and theoretical basis for the clinical diagnosis and treatment of TNBC.

Introduction

Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies, and application systems to simulate, extend, and expand human intelligence [1]. AI has become a popular branch of computer science in recent years. Currently, there are four main technical branches of AI: (1) Pattern recognition: through the analysis and processing of various forms of information presented by various objective things and phenomena, classify, judge, and describe things and phenomena [2]; (2) Machine learning (ML): simulation to achieve the learning behaviour of the human brain, reorganising its knowledge framework to improve the performance of the program [3]; (3) Data mining: relevant algorithms from numerous databases are used to eliminate redundant data and extract key pieces of information [4]; (4) Intelligent algorithms: specialised algorithms for solving specific problems. ML and deep learning (DL) are the main methods used to implement AI, and are sometimes used as synonyms. In computer science, ML is a subfield of AI and DL is a specific subset of ML characterised by its ability to perform automatic feature extraction and its power in the assimilation and evaluation of large amounts of complex data [5]. For example, convolutional neural networks (CNN) are widely used DL algorithms (Fig. 1). In recent years, owing to the development of large amounts of medical data, AI, especially DL, has been an important tool in the accurate diagnosis and design of better treatment schemes for diseases. Particularly in breast cancer, Decision Support and Information Management System for Breast Cancer (DESIREE), which is a web-based software ecosystem derived from AI, is committed to the individualised, collaborative, and multidisciplinary management of primary breast cancer, from diagnosis to treatment and follow-up [6–8].

Fig. 1. Branches of artificial intelligence technology and the relationship between AI and ML and DL, with examples of commonly used algorithms.

CART classification and regression trees, CNN convolutional neural networks, LASSO least absolute shrinkage and selection operator, SMV support vector machines.

TNBC is the most aggressive subtype of breast cancer [9]. These cells lack the expression of the oestrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER-2); therefore, they are not suitable for endocrine therapy or anti-HER2 treatment. Surgery, chemotherapy, and radiotherapy are the main methods of TNBC treatment. Recently, with the increased speed and scale of biological data collection and new advances in computational technology, AI-based approaches can provide powerful analytical frameworks for TNBC molecular and clinical data. These parallel advances will constitute breakthroughs in the accurate diagnosis, classification, treatment, and prognosis prediction of TNBC (Fig. 2). This study reviewed the aspects of these research advances.

Fig. 2. A framework of AI in TNBC.

AI artificial intelligence, TNBC triple-negative breast cancer, DEG differentially expressed genes, TIL tumour-infiltrating lymphocytes, CAF cancer-associated fibroblast.

AI improves the accuracy of TNBC diagnosis

Early screening and accurate diagnosis are important for improving treatment. Current screening and diagnostic methods for TNBC have some limitations, such as mammography, which typically provides 10–14% false negatives in screening; breast ultrasound, which is highly operator skill dependent and difficult to analyse retrospectively; and magnetic resonance imaging (MRI) methods, which are expensive and cannot discriminate with other subtypes of breast cancer [10]. AI has been increasingly researched in recent years for TNBC screening and diagnosis, numerous AI algorithms have been proposed and applied to many of these diagnostic tools, which could improve diagnostic accuracy. However, a central AI-based algorithm that can be broadly implemented into screening programmes to provide a consensus approach is the ideal situation.

Interpretation of pathological markers

It is difficult to make accurate pathological morphological judgements owing to poor differentiation, especially when distant metastases occur, and it can be easily confused with primary carcinoma and metastatic carcinoma of other site origins. The current pathological diagnosis depends on immunohistochemical (IHC) staining [11]. However, conventional pathological diagnosis varies significantly between pathologists. Therefore, ML-based systems can be of great help in interpreting pathological diagnostic markers. Castaldo et al. showed that quantitative radiomic analysis had the potential to be used as a method for high-throughput image-based phenotyping to detect breast cancer receptor status using ML methods. They evaluated the predictive ability of imageological examination of tumour features for four clinical phenotypes (ER-positive, PR-positive, HER-2 positive, and triple-negative) using three ML techniques: support vector machines (SMV), random forest (RF), and naive Bayes. The results showed that the tumour-enhanced texture features selected by the random forest algorithm were good predictors of ER and HER-2 protein expression and triple-negative status in breast cancer, with the area under the receiver operating characteristic (ROC) curve (AUC) values of 86, 91, and 91%, respectively. Thus, ML algorithms combined with radiological features can automatically identify clinical variables associated with tumour genomic status, such as ER status and HER-2 status. It would facilitate clinical diagnosis and treatment planning for different breast cancer subtypes. Radiological features could be potential imaging biomarkers to advance precision medicine [12].

Combined with mammography

Mammographic density is associated with the sensitivity and specificity of breast cancer screening and is considered an independent risk factor for breast cancer [13]. Women with high-density breasts are at a higher risk of developing cancer and high-density breasts are less sensitive to imageological examination; therefore, this type of breast cancer is difficult to detect early. This challenge is now being addressed by using a mammography-aided diagnostic system. This system consists of three main components, which are lesion detection, lesion registration, and malignancy prediction modules. All components are made up of CNNs. The framework and output of CNN are capable of performing diverse tasks. The system’s first component may carry out three different types of tasks to detect lesions precisely, and the latter two modules comprise the classification network, which provides the lesion types and BI-RADS rating for each lesion. Automated mammography diagnostic systems offer significant improvements in both detection and diagnostic performance over conventional imaging examinations, even in very dense breasts, and have been shown to detect 75% of TNBC lesions (usually irregular, non-calcified masses with indistinct or burr-like margins) when used in TNBC cases, thereby improving the detection rate and reducing missed diagnoses [14]. Moreover, another study developed and established ML models based on clinical and imaging signs to differentiate molecular subtypes of breast cancer and reproduce image-based disease reasoning and diagnosis. Of the five models, the decision tree (DT) model performed best in distinguishing TNBC from other breast cancer subtypes, with an AUC of 0.971; precision, 0.947; sensitivity, 0.905; and specificity, 0.941. In the DT model, the two most important features in distinguishing TNBC from other subtypes were blurred margins and calcification on mammography. Therefore, the application of this model to assist diagnosis in clinical practice may allow radiologists to better identify the molecular subtypes of breast cancer patients [15].

Combined with multiparametric MRI

Leithner et al. evaluated the performance of radionics and AI using multiparametric MRI for the assessment of breast cancer molecular subtypes. The results showed that the overall median area under the ROC curve (AUC) was 0.86 (0.77-0.92) when using a multi-layer perceptron feed-forward artificial neural network (MLP-ANN) to differentiate TNBC from other subtypes, the classification of luminal A and TNBC generated an overall median AUC of 0.8 (0.75–0.83) [16]. Another study showed similar results in that some specific radiomic features derived from multiparametric MRI could predict the molecular subtype of breast cancer. The researchers used six classification algorithms to construct the model, with the MLP classifier showing the best diagnostic performance in the task of differentiating TNBC from non-TNBC on the test dataset, with the AUC of 0.965 and the accuracy of 92.6% [17]. The combination of AI and multiparametric MRI radionics can non-invasively and accurately distinguish TNBC from other breast cancer subtypes.

Combined with spectroscopy

Combining AI and ML with Raman spectroscopy (RS) on breast cancer biopsy tissue to explore biochemical variations within and between datasets related to lipid, collagen, and nucleic acid content showed that RS combined with linear difference analysis predicted TNBC with an accuracy of 96.7%. Thus, the combination of RS with AI and ML may provide an accurate real-time approach for TNBC diagnosis and monitoring [10].

Combined with breast ultrasound

It has been found that the analysis of ultrasound images of breast masses using ML enables accurate differentiation between TNBC and non-TNBC subtypes. Ultrasound image features were measured using greyscale and colour Doppler images, followed by logistic regression and classification using ML. The results showed statistically significant grey-scale (GS) and colour Doppler (CD) features with an AUC of 0.85 and 0.65 respectively. When GS and CD features were used together, the AUC increased to 0.88 with a sensitivity of 86.96% and a specificity of 82.91%. This significantly exceeds the diagnostic performance of standard vision for image assessment, and this diagnostic technique could effectively assist clinicians in the future to improve the accuracy of diagnosis [18].

Combined with characteristic cellular uptake responses

Alafeef et al. proposed a method for using AI to predict the cellular internalisation of nanoparticles against different cancer stages. They demonstrated for the first time that a combination of ML algorithms and characteristic cellular uptake responses for individual cancer cell types could be successfully used to classify various cancer cell types. Building on this, this study designed a diagnostic platform consisting of eight carbon nanoparticles with multiple surface chemistries that can be more accurately sub-classified between TNBC and non-TNBC cells in the TNBC group with the aid of ML, offering potential applications in TNBC diagnosis [19].

AI assists to classify TNBC subtypes

There is growing evidence that TNBC is highly heterogeneous and many studies have focused on identifying TNBC subtypes based on genomic analyses [20]. Continued exploration of molecular typing can help guide precise treatment of TNBC. In 2011, Lehmann et al. initiated the first typing analysis of TNBC based on transcriptome data, classifying TNBC into six categories: basal-like 1 (BL1), basal-like 2 (BL2), immunomodulatory (IM), mesenchymal (M), mesenchymal stem-like (MSL), and luminal androgen receptor (LAR) [21]. Another important study on transcriptome-based TNBC subtypes was recently conducted by Jiang et al. [21], who analysed clinical, genomic, and transcriptomic data from 465 primary TNBC cases and classified TNBC into four subtypes: LAR, IM, basal-like immune-suppressed (BLIS), and mesenchymal-like (MES). Notably, these four tumour-specific subtypes have different genetic alterations and should be treated using different strategies. However, the diversity of molecular data, lack of uniformity between approaches, and lack of cost-effective systematic classification methods have made it difficult to widely implement these advances. A recent study revealed the constraints of formalin-fixed, paraffin-embedded samples and batch effect removal across microarray platforms, as well as refined the use of computational tools for TNBC subtyping. To address the lack of consensus on the molecular subtypes of TNBC due to the choice of different workflows, sampling methods, computational tools and algorithms, they described a protocol for molecular subtyping of TNBC (Fig. 3), which will aid in the rational design of future studies [22]. The application of AI in translational oncology promises to bring new ideas for identifying TNBC subtypes (Table 1) [21].

Fig. 3. A patterned pipeline for TNBC subtyping.

Each step introduces alternative methods and data types. Following subtype discovery, several optional modules are listed for downstream analysis. This pipeline can also be used for similar research in other types of cancer.

Table 1.

Machine learning algorithms applied to TNBC classification.

Classification method	Classification model	Subtypes	Model performance	Prognosis	Characteristics	Potential treatments
DEGs	SMV algorithm	BLIA	Accuracy: 95.7% AUC: 0.99	Better	Upregulated DEGs were related to the immune system	Immune checkpoint inhibitor
		BLIS	Accuracy: 95.6% AUC: 0.99	Worse	Downregulated DEGs in immune regulation pathways	–
		MES	Accuracy: 95.0% AUC: 0.99	Neutral	Upregulated DEGs were related to mesenchymal differentiation	Anti-angiogenic therapy
		LAR	Accuracy: 98.8% AUC: 1.00	Neutral	Upregulated DEGs enriched in oestrogen-dependent gene expression	Androgen receptors
Immunogenomic profiling	RF model	Immunity-H	Accuracy: 89%	Better	Immunity-H contains the highest number of immune cells and stromal cells Highest PD-L1 expression levels High expression of HLA genes	Anti-PD-1/PD-L1 therapy
		Immunity-M	Accuracy: 89%	Neutral	Between Immunity-H and Immunity-L	–
		Immunity-L	Accuracy: 89%	Worse	Immunity-L contains the highest number of tumour cells Lowest PD-L1 expression levels Low expression of HLA genes	–
	RF model	S1	AUC: 0.76	Better	Higher immune scores Higher levels of immune cells Better prognosis for immunotherapy	Immunotherapy
		S2	AUC: 0.76	Worse	Lower immune scores Lower levels of immune cells	–
Cancer-associated fibroblasts	RF model	CAF-	AUC: 0.921	Better	Higher immune cells Higher immune-related pathways	Immune checkpoint blockade
		CAF+	AUC: 0.921	Worse	Lower immune cells Lower immune-related pathways	Targeting cancer-associated fibroblasts
Transcriptomics data	GB model	C1	Sensitivity: 97.39% Specificity: 99.57% AUC: 0.98	Better	Apocrine	An association of PI3K inhibitors and antiandrogens
		C2	Sensitivity: 91.59% Specificity: 96.47% AUC: 0.94	Neutral	Highest pro-tumourigenic response High neurogenesis activity High biological aggressiveness	Immune checkpoint inhibitor
		C3	Sensitivity: 94.67% Specificity: 94.72% AUC: 0.95	Neutral	Highest anti-tumourigenic responses Complete B cell differentiation Immune checkpoint upregulation	Immune checkpoint inhibitor

DEG differentially expressed gene, AUC area under the curve of the receiver operating characteristic, SMV support vector machines, BLIA basal-like immune-activated, BLIS basal-like immune-suppressed, MES mesenchymal, LAR luminal androgen receptor, RF Random Forest, Immuntiy-H Immunity High, Immunity-L Immunity Low, Immunity-M Immunity Medium, CAF cancer-associated fibroblast, GB gradient boosting, C1 molecular apocrine, C2 immunosuppressive response, C3 adaptive immune response.

Subtyping according to differential genes

To date, many researchers have studied the genetic signatures of TNBC subtypes using different methods, and with the advancement of gene chip technology, a large number of gene (signature) datasets have been generated. ML is gradually being explored for its use in this field. Combining previous studies and based on gene expression characteristics, Bissanum et al. identified unique differentially expressed genes (DEGs) and used them as a training gene set in seven different classification models [23], and the results showed that this training gene set was suitable for the development of TNBC classification models, and the SVM algorithm was the most accurate compared to other algorithms in ML (accuracy 95–98.8%; AUC 0.99-1.00) into four subtypes of TNBC: basal-like immune-activated (BLIA), basal-like immune-suppressed (BLIS), mesenchymal (MES), and luminal androgen receptor (LAR) [24]. These factors are crucial for the individualised treatment and prognosis of TNBC.

Subtyping according to immunogenomic profiling

One study used immunogenomic profiling to classify TNBC into three distinct subtypes, Immunity High (Immunity-H), Immunity Medium (Immunity-M), and Immunity Low (Immunity-L), and demonstrated the stability and reproducibility of this classification in four separate datasets using the RF approach [25]. They first used tenfold cross-validation (CV) to evaluate the classification performance of the RF algorithm in METABRIC, and then used the METABRI dataset as a training set to predict the TNBC subtypes in the other three datasets. Their study showed that, the tenfold CV accuracy for the classification of the METABRIC dataset was 89%, and the classification accuracies for TCGA, GSE75688 and GSE103091 were 70, 84 and 63%, respectively. The weighted F scores for METABRIC, TCGA, GSE75688 and GSE103091 were 89, 71, 83 and 63%, respectively. Thus, TNBC classification based on immunogenomic analysis was stable and predictable. In addition, compared to other subtypes, Immunity-H was distinguished by higher immune cell infiltration and anti-tumour immune activity, as well as a better survival prognosis. Along with immunological markers, some cancer-related pathways, such as apoptosis, calcium signalling, MAPK signalling, PI3K-Akt signalling, and RAS signalling, were hyperactivated in immunity-H mice. Immunity-L, on the other hand, displayed weakened immune signatures and elevated activation of the cell cycle, Hippo signalling, DNA replication, mismatch repair, binding of cell adhesion molecules, spliceosome, adherens junction function, pyrimidine metabolism, glycosylphosphatidylinositol (GPI)-anchor biosynthesis, and RNA polymerase pathways [25]. Another study conducted an immunophenotypic analysis of a large number of TNBC samples by unsupervised clustering, stratifying TNBC patients into two new immune subtypes (called S1 and S2). Compared to S2, S1 was associated with higher immune function scores, higher levels of immune cells and a better prognosis for immunotherapy, both in terms of overall survival (OS) and relapse-free survival (RFS). Bioinformatic analysis identified 11 central genes associated with immune subtypes (LCK, IL2RG, CD3G, STAT1, CD247, IL2RB, CD3D, IRF1, OAS2, IRF4 and IFNG). A robust ML model based on the RF algorithm was created by 11 hub genes, and it performed reasonably well with AUC of 0.76 [26]. These studies identified subtype-specific molecular features of TNBC, including genes, gene ontologies, pathways, and networks, which may help optimally stratify TNBC patients who respond to immunotherapy [25].

Subtyping according to cancer-associated fibroblasts

A recent study concluded that cancer-associated fibroblasts (CAFs) are one of the independent prognostic factors in TNBC. The extent of CAF in the tumour microenvironment (TME) is higher in those with poorer prognoses. CAF may have an impact on prognosis by promoting tumour cells and suppressing immune cells such as CD8 + T cells. The investigators classified TNBC patients into two CAF subtypes (CAF+ and CAF−) by cluster-free analysis of gene expression profiles. The CAF− subtype of TNBC was associated with longer OS and more immune cells compared to the CAF+ subtype. Bioinformatics analysis identified nine CAF subtype-associated markers (ADAMTS12, AEBP1, COL10A1, COL11A1, CXCL11, CXCR6, EDNRA, EPPK1, and WNT7B). They constructed a robust random forest model using these nine genes with an area under the curve (AUC) value of 0.921 for the model. Thus, this stratification could identify CAF− subtype patients as ideal candidates to receive Immune checkpoint blockade. In addition, targeting CAF may be a promising therapeutic approach that complements conventional treatments and immunotherapy [27].

Building an absolute assignment predictor

One study classified TNBC into three potential therapeutic subtypes: molecular apocrine response (C1), immunosuppressive response (C2), and adaptive immune response (C3) [28]. Ben et al. used the above typing results as a basis for clustering the genomic dataset of the TNBC cohort and used the transformed optimal variables in three prediction models, RF, gradient boosting (GB), and extreme gradient boosting, to predict the TNBC subtypes. Fifty strongly correlated indices highlighted the biological characteristics of each TNBC subtype. The GB performance based on this subset of metrics was superior to that of other models. This prediction model provides the probability that each patient belongs to one of the three TNBC subtypes mentioned above, which is key to the transition to precision medicine and the selection of targeted therapies for TNBC [29].

AI applies to the treatment of TNBC

AI is uniquely positioned to enable the personalised treatment of TNBC [30], where it can explore potential therapeutic targets for TNBC based on large amounts of genetic and molecular data. At the same time, AI is often used to predict TNBC therapeutic response, virtual screening of potentially effective therapeutic agents, and activity prediction [31]. In addition, recent studies have shown that AI can also successfully interface with nanocarriers that carry all the necessary weapons (drugs, tracking probes, and ligands) to target TNBC cellular sites and that the combination of targeted delivery systems and multifunctional molecules will provide a new nanotherapeutic paradigm for TNBC [32].

Identify selective therapeutic targets

In TNBC, several targets are currently being studied in clinical trials and include growth factor receptors (EGFR, cMET, VEGFR), downstream signalling (PI3K/mTOR pathway, SRC, Wnt-Catenin signalling), cell cycle checkpoints (CHK1/2), PARP inhibitors, CSPG4 inhibitors, the androgen receptor, and others [33, 34]. Most are only effective for a subset of breast cancer patients and have not been very promising because of a variety of other underlying factors [35, 36]. Thus, ML algorithms have demonstrated significant advantages. Gautam et al. developed an ML algorithm (idTRAX) to identify selective therapeutic targets in TNBC cell lines. The results showed that AKT inhibition selectively killed MFM-223 and CAL148 cells, whereas FGFR2 inhibition killed only MFM-223 cells. This ML algorithm is essential for identifying specific targets for TNBC small-molecule drug development and personalised therapy [37]. Kothari et al. used ML algorithms to identify two potential genes, TBC1D9 and MFGE8, TBC1D9 was underexpressed in TNBC patients, whereas MFGE8 was overexpressed. TBC1D9 overexpression is associated with a good prognosis, whereas MFGE8 overexpression is associated with a poor prognosis. Furthermore, it has been suggested that TBC1D9 plays a role in maintaining cell integrity, whereas MFGE8 may be involved in various tumour survival processes. Therefore, MFGE8 could be targeted, it might be downregulated by using MFGE8-specific inhibitors. Additionally, MFGE8 effector molecules could also be specifically targeted. As for TBC1D9, the dilemma of how to effectively target a gene that is downregulated in a disease emerges. This could be accomplished in a number of ways, including by focusing on its regulators, using gene therapy, a vaccination strategy, or blocking the active pathways. These two genes may be potential targets for TNBC therapy [38].

Predict treatment response

Predicting the sensitivity of TNBC to chemotherapeutic drugs is an important research topic [37]. Although advances in cellular high-throughput drug screening technologies have led to the generation of large amounts of relevant data, the analysis of such data is a challenging task; therefore, advanced ML algorithms are urgently required to accurately predict clinical drug responses [39]. A recent study described the use of an AI platform in breast hollow needle biopsies to predict treatment response to neoadjuvant chemotherapy for TNBC. The study built multiple deep convolutional neural networks to automate tumour detection and nuclear segmentation and showed that the pathological complete response (pCR) group had fewer multifocal/multicentric tumours, higher nuclear intensity and lower grey-level co-occurrence matrix (GLCM-COR) compared to the pathological partial response (pPR) group, thus using AI to analyse tumour cell nuclear features could be a good way to build a predictive model for TNBC neoadjuvant model for pathological response to chemotherapy [40]. Applying AI to synthesise metabolic profiles can reliably identify the patients who were non-response to neoadjuvant chemotherapy (NACT). Previous studies had found that when compared to TNBC patients who had achieved pCR/residual cancer burden-0 (RCB-0) or had minimal residual disease (RCB-I), those with moderate to extensive tumour burden (RCB-II/III) after NACT had higher pre-treatment plasma levels of acetylated polyamines. Based on this conclusion, using a deep learning model (DLM), a metabolite panel containing two polyamines and nine additional metabolites was established for improved prediction of RCB-II/III. The DLM had high predictive performance, with an AUC of 0.97 (95% CI: 0.93-1.00), 85% sensitivity, and 95% specificity for identifying RCB-II/III, it could be used in the clinical setting to identify TNBC patients who were at high risk of not responding to NACT and might benefit from alternative treatment options. On the other hand, TNBC patients who were likely to respond to NACT, may benefit from dose reduction, allowing for the management of treatment-related toxicity [41]. In addition, Kim et al. performed a computational study of TNBC drug response scores by applying chemical perturbation gene signatures and cross-validation [42]. The development of these treatment response prediction algorithms and models could accelerate the transfer of AI to clinical settings [43] and drive precision and personalised treatment in TNBC clinics.

AI applies to the prognosis prediction of TNBC

Several clinicopathological factors have been shown to have prognostic value in TNBC patients. For example, the presence of high tumour-infiltrating lymphocytes (TILs) is associated with disease-free survival and overall survival in early-stage TNBC treated with standard adjuvant/neoadjuvant therapy [44], an elevated neutrophil-to-lymphocyte ratio (NLR) suggests an association with poor outcomes in TNBC [45], and studies have suggested that upregulation or downregulation of specific miRNAs is associated with prognosis in TNBC [46]. Many studies have been devoted to AI for the prognosis of TNBC and have found that AI can achieve accurate disease prognosis in early-stage TNBC in a variety of ways, identifying gene signatures associated with TNBC prognosis and excluding factors unrelated to TNBC prognosis.

Assessment of TNBC prognosis by TILs

TILs are considered to have predictive value for prognosis [47]; however, they suffer from inter- and intra-observer variability, which prevents their widespread use [48]. Therefore, to overcome the low reproducibility of TILs assessments, ML algorithms may contribute to the standardisation of TILs assessments in the future [49]. A study that constructed an ML-based approach to breast cancer TILs scoring and validated its prognostic potential in several TNBC cohorts showed that in several independent validation cohorts of TNBC patients, machine read-measured TILs variables stratified TNBC patients into good and poor prognosis cohorts, where higher TILs scores were significantly associated with better overall survival; thus, TILs volumes are considered robust and independent prognostic factors [50]. In addition, Balkenhol et al. looked at three TIL markers (CD3, CD8, and FOXP3) in different parts of the tumour and its surrounding environment. To objectively assess TILs, they used advanced full-slide image analysis algorithms, including CNN, which can detect each individual positive lymphocyte. These results are consistent with those of previous studies, in which TILs abundance was negatively correlated with relapse-free survival (RFS) and overall survival (OS), providing direction for the optimisation of the assessment [48].

Assess the risk of TNBC metastases

Patients with TNBC often show an increased risk of brain metastasis as the initial site of metastasis, early brain metastasis and the shortest survival associated with brain metastasis [51]. Therefore, it is crucial to determine when TNBC cells are likely to metastasise. Currently, no intervention can elucidate the possibility of primary TNBC tumour metastasis to the brain. Oliver et al. constructed a model that could predict the potential of cancer cells to metastasise. The model combined advanced living cell imaging algorithms and AI to describe cancer extravasation, detected subtle phenotypic differences between cells with and without brain metastases potential based on their behaviour in a brain-like tumour micro-environment, and analysed when cancer cells were likely to metastasise to the brain, with the potential to be translated into a clinical predictive diagnosis of brain metastases with additional studies [52]. Another study developed a brain metastasis risk-prediction model based on the radiomics features taken from pre-operative MRI. When brain metastasis was the target event in 10-fold verification mode, the model that used the Nave Bayes algorithm performed best (AUC = 0.878, accuracy = 0.786, sensitivity = 76.2%, and specificity = 81.0%). This application may aid in improving early stratification, brain metastasis screening, and overall prognosis in TNBC [53]. TNBC development is strongly linked to the physiological state of TME. Different components of TME contribute to the progression and metastasis of TNBC through various molecular mechanisms, such as induction of angiogenesis, proliferation, inhibition of apoptosis and suppression of the immune system [33]. Recursive Feature Elimination with Random Forest (RFE-RF), a feature selection algorithm, was used in the gene expression dataset, which helped to identify the core representative and functional genes in TNBC. The Kaplan–Meier (KM) analysis for Distant metastasis-free survival (DMFS) of the 34 differentially regulated genes identified by RFE-RF revealed two potential prognostic genes (POU2AF1 and S100B). POU2AF1 and S100B genes may inhibit distant metastasis of TNBC through immunomodulatory and signalling effects in the TME, giving them the possibility to become novel prognostic markers and therapeutic targets in TNBC metastasis [54].

Identify other TNBC prognosis-related variables

Traditional prognostic factors such as age, tumour size, tumour grade, and lymph node status have restricted risk-predictive value in TNBC because these tumours are mostly of higher grade, with greater odds of recurrence and metastasis [55]. As a result, using advanced techniques to identify TNBC prognosis-related variables is an unfulfilled need. A group of cases from a large, well-characterised cohort of patients with primary TNBC was subjected to RNA sequencing. An artificial neural network identified two gene panels that strongly predicted survival without distant metastases and breast cancer-specific survival. Correcting for clinicopathological factors and using multivariate Cox regression analysis yielded a two-gene prognostic signature (ACSM4 and SPDYC) that was significantly associated with poor prognosis in TNBC, independent of other prognostic variables [56]. This study identified genes that predicted the prognosis of patients with TNBC and divided TNBC patients into high- and low-risk groups for developing distant metastasis, which could potentially be used to guide the clinical management of patients with TNBC. Another study screened out the top ten hub genes (PBK, TOP2A, CDCA8, ASPM, CCNA2, KIF20A, BUB1, AURKB, CDK1, and CCNB2) related to TNBC using the CytoHubba plugin in Cytoscape, and six significant TNBC genes (CDK1, CENPF, MCM7, PACC1, TUBB, and UBE2C) were screened using LASSO-based ML. They further evaluated the prognostic value of the 10 highest scoring genes and 6 significant genes in the KM Plotter database. The OS analysis showed that BUB1 and CENPF had a significantly poor prognosis in TNBC patients, while BUB1, CCNA2, CDCA8 and PACC1 showed a significantly poor prognosis in TNBC patients in the DFS analysis [57]. Millar et al. used digital image analysis and ML algorithms to analyse automated images of the tumour stromal ratio, thus validating the finding that a high stroma is an independent poor prognostic factor for TNBC [58]. In addition, another study using deep learning and manual assessment to assess the prognostic value of absolute mitotic counts in TNBC showed that nuclear split counts to reflect proliferation levels are not a prognostic factor in TNBC and that management of patients with TNBC based on mitotic counts should not be encouraged [59].

Discussion and future perspectives

AI, especially ML and DL, is being used for diagnosis, staging, treatment, and prognosis of TNBC, with its advantages of high efficiency and accuracy, providing new theoretical guidance and clinical ideas for TNBC, a medical problem with poor differentiation, early metastatic tendency, and high recurrence rate [60]. For example, the use of AI systems for mammography can significantly improve the detection rate of TNBC, which can efficiently screen potential therapeutic targets for TNBC and validate TNBC prognosis-related variables, such as TILs and tumour stroma ratio. In addition, rapid advances in AI subfields, such as ML and DL, will enable the construction of models using different molecular data types, which will provide important tools for integrating different molecular data from multiple current TNBC subtype systems and hold the promise of providing a standardised stratification system to guide the treatment of patients with TNBC.

However, there are still many unsolved problems that hinder the application of AI in the clinical work of TNBC diagnosis and treatment: (I) from the current clinical applications, AI has been better applied in TNBC in the areas of screening and auxiliary diagnosis, and is still weak in drug development; (II) a large amount of TNBC clinical data lacks systematic collation and management. Meanwhile, data access is difficult, data standards are not uniform and data theory is controversial [61]; (III) AI systems have high development costs, low promotion rates, and a lack of professional operators; (IV) poor stability is a fatal weakness of modern AI models. Uncertainty in a model can arise from the construction of algorithms, the selection of data, the accuracy and completeness of the data, inherent biases in the data and misspecification of the model. Hence, stability testing is crucial for the widespread use of AI models; (V) DL is now a black box that does not clearly explain the decision-making process. Exploring the details of how DL analyses data and makes decisions may improve the interpretability and stability of AI models [62]. Thus, limitations in data integrity, large volumes of data, translation of these data into knowledge, ethical considerations, and regulatory approval gradually add to form a huge barrier to the clinical translation of AI. To maximise the potential of ML for TNBC research, sufficiently annotated medical data needs to be deposited in large-scale databases. With the development of new technologies like single-cell sequencing, spatial transcriptomics, and multiplex imaging, the quantity and quality of labels for TNBC data will be enhanced. As AI continues to evolve, the use of multimodal learning to integrate medical images and holographic data holds great promise for TNBC research. Improvements in algorithms, accumulation of big data, increased computational power and breakthroughs in interpretability bottlenecks will enable widespread clinical application of AI in TNBC.

Author contributions

JM and SZ designed and prepared the manuscript. JG, JH, YZ and JM drafted the manuscript.

Funding

This research was supported by the Key Research and Development Project of Sichuan Province (2023YFS0171 to JM).

Competing interests

The authors declare no competing interests.