The artificial intelligence and machine learning in lung cancer immunotherapy

The training process in ML can be categorized into three main types: supervised learning, unsupervised learning (UL), and semi-supervised learning (SSL). Supervised learning involves using labeled data to train the model. Some typical classification and regression models for supervised learning include k-nearest neighbor (KNN), linear regression (LR), support vector machine (SVM), decision trees (DT), and random forests (RF). In contrast, UL does not involve labeling the data, and SSL combines labeled and unlabeled data. The labeling process can be time-consuming and labor-intensive, but it can result in better performance of the models since they have been externally validated [6]. Clustering models are common UL algorithms, including k-means clustering, hierarchical clustering, and principal component analysis (PCA).

Overall, the choice of learning type depends on the data and the task at hand. Supervised learning is suitable for tasks that require predicting a specific output from input data with known labels, while UL is used to discover hidden patterns and structures in data. SSL is used when there is a limited amount of labeled data available.

DL is a subset of ML that utilizes neural networks as its fundamental algorithm. Unlike traditional ML, DL does not rely on domain experts to manually engineer features. Instead, it mimics the iterative transmission of information in the human brain by using neural networks to automatically learn representations of data. During training, the algorithm adjusts its parameters to optimize the model and produce the best output. DL has found applications in various fields, including computer vision (CV), natural language processing (NLP), and speech recognition.

In the medical field, NLP has been applied to various scenarios, including simple internet-based AI consultations, information extraction from electronic medical records, and automatic case writing [7]. Additionally, CV has been extensively employed for medical image recognition in areas such as computer tomography (CT), X-ray, positron emission tomography/computer tomography (PET/CT), and immunohistochemistry (IHC) [8]. Many reported models for predicting lung cancer risk utilize supervised ML, such as artificial neural network (ANN), DT, RF, SVM and Bayesian classification.

The concept of AI was officially proposed at the Dartmouth Conference in 1956 (Fig. 1). Scientists want to create machines that can mimic human intelligence [3].

During the early days of the 1960s, computers’ operation relied on the “expert system”, which refers to a large number of manual interpretation rules input by experts, forming a knowledge database [3].

In the 1970s, the limitation of the development of hardware equipment led to insufficient computing power, making it difficult to calculate large-scale data and complex missions. As a result, capital investment gradually decreased, and the evolvement of AI reached a stalemate, entering the “AI winter” period in history [3].

Until the 1980s, the concepts of ML and neural networks emerged. Canadian scholar Geoffrey Hinton improved the traditional perceptual network structure, coupled with the invention of back propagation and the extensive application of statistical principles, AI gained the ability to solve practical problems and gradually had commercial value [9]. Concurrently, AI has also developed in the fields of life sciences and medicine. Additionally, the development of the Internet promoted progress in NLP and data mining greatly.

In 2009, Li Feifei presented the ImageNet database for the first time as an academic poster at the Conference on Computer Vision and Pattern Recognition (CVPR), which expanded the types of samples that can be used for AI training, promoting the process of CV and image recognition greatly. With the advent of the Big Data era and the development of computer hardware, the concept of DL was proposed and emerged, which led to the development of convolutional neural networks (CNN) and deep neural networks (DNN). Since then, AI has entered a peak period of research and development, becoming well-known to the public [3].

Additionally, bioinformatics and semantic analysis technologies were also developed rapidly. In 2015, Canada’s DNA sequencing data enabled the identification of mutation sites and therapeutic targets, thus providing personalized treatment plans for patients. Furthermore, a speech recognition assistant developed by iFlytek and Tsinghua University was able to analyze patients’ conditions and provide auxiliary diagnoses [3].

Later on, with the development of big data, the evolution of ML algorithms, and the improvement of model prediction performance and generalization capabilities, AI is increasingly being applied in the field of biomedicine, including protein structure and function prediction, nucleotide sequencing analysis, drug characteristics, speech recognition and network consultation, auxiliary diagnosis mapping, risk prediction modeling, robot-assisted surgery and other fields [3] (Fig. 1).

However, individualized diagnoses and therapeutic strategies, such as early screening and diagnosis, functional visualization of key molecular events and targeted drugs, are still imperative for lung cancer treatment. New technologies such as tumor-assisted diagnosis combined with AI, analysis of molecular pathology information, prediction of tumor invasion and treatment resistance, and multi-omics fusion modeling to predict treatment outcomes and prognosis are providing new ideas and opportunities for clinicians.

Lung cancer is one of the most common cancers in the world with the highest mortality rate (1). Immune checkpoint inhibitors (ICIs) targeting PD-1, cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) and PD-L1 have been widely developed and have shown good efficacy in treating non-small cell lung cancer (NSCLC). However, only about 30% of patients are eligible for treatment. Immune-related adverse events (irAEs) always occurred with patients, too. Traditional inspection methods are often limited in their ability to achieve expected benefits [10, 11].

PD-L1 expression and TMB were the first clinically assessed biomarkers. TMB is the total number of non-synonymous somatic mutations per megabase in the coding region of the tumor genome, with a wide range of mutations [12]. High TMB is positively associated with more tumor-associated neoantigens and improved immunotherapeutic efficacy [12]. Whole exome sequencing (WES) is the gold standard for assessing TMB. Studies have shown that NSCLC patients with above-median WES mutations have longer overall survival (OS) [13]. TMB is commonly used as a pan-cancer biomarker to identify patients who may benefit from PD-1 therapy, as it is a surrogate for tumor neoepitope burden. In addition, the expert consensus on immunotherapy for lung cancer recommends the use of mutational landscapes to assess the efficacy of PD-L1 immunotherapy in NSCLC [11].

According to the National Comprehensive Cancer Network (NCCN) guidelines, the expression status of PD-L1 protein levels determined by IHC via biopsy is the sole clinically approved biomarker for the evaluation of ICI therapy [14]. Higher levels of PD-L1 expression are typically associated with more favorable immunotherapeutic outcomes [15], but this relationship is not necessarily positive [16].

PD-L1 values obtained through routine pathology reports are lacking in a definitive gold standard and are instead crude, subjective, and semi-quantitative [17], resulting from gene mutations and sampling site differences. These factors result in significant inter-observer variability and approximately 30% inconsistency in judgments around the cut-off point [18, 19].

Moreover, due to the highly invasive nature of puncture biopsies or surgical specimens, which are often sampled during a single visit, the results are susceptible to static tumor characteristics and intratumoral heterogeneity making them less effective in predicting the benefit of immunotherapy [20, 21]. Thus, it is crucial to develop noninvasive and robust methods for interpreting PD-L1 expression that can be reviewed multiple times during follow-up or to identify alternative biomarkers [12].

The accumulation of patient demographics, imaging, pathology images, laboratory data, medical history, sequencing data and other comprehensive information furnishes clinicians and statisticians with a substantial foundation of big data to analyze and identify the characteristics of people who benefit from ICIs. The utilization of ML to analyze multi-omics data for modeling and prediction efficiency and survival status has become the most promising development in precision medicine (Fig. 2) (Table 1).

In prior investigations about NSCLC, radiomics images have been implemented to anticipate alterations in tumor-infiltrating CD8⁺ T cell levels, with the aim of distinguishing patients who would benefit from PD-L1 therapy. Such endeavors have revealed the potential of radiomics in predicting TME [31].

Understanding the individual differences in TME can aid in the screening of populations who may respond to immunotherapy. Researchers have collected baseline PET/CT radiomics data and CD8 expression data from tumor specimens of 221 NSCLC patients. They employed ML models to predict the TME phenotype, thus ascertaining the immune status of NSCLC. This constitutes one of the several attempts to achieve noninvasive TME prediction through imaging-clinical joint models [37].

Presently, the utilization of CT scans for the early detection of lung cancer in high-risk groups is being widely advocated. Ground-glass opacity (GGO) has been identified as an imaging characteristic of early lung cancer. Although the corresponding pathological features do not meet surgical criteria, the diagnostic potential of GGO for early-stage lung cancer cannot be disregarded. Consequently, a team of researchers has constructed a 15-gene risk signature related to pure ground glass opacity (pGGO) through transcriptome analysis. They have utilized this signature to predict the prognosis of early-stage lung adenocarcinoma (LUAD) and to investigate the immune microenvironment of GGO. The predictive ability of this signature for patients with early-stage adenocarcinoma has been verified in TCGA and Gene Expression Omnibus (GEO) datasets [38].

Many studies of similar nature are currently emerging, and the algorithms and performance of the models are being continuously improved [39]. These studies substantiate the promising future of radiomics in TME prediction.

Researchers use both authentic cases and HE stained images in databases to train models. Previous research endeavors incorporated HE images of thirteen cancer types from TCGA in order to map TILs using CNN. Researchers try to elucidate the local spatial structure in TME and its association with OS [40].

By assessing the density of CD3⁺ T cells and CD8⁺ T cells in the tumor area through WSI, Lin et al. [41] established an automatic assessment model of I-score cell density that can be utilized for clinical risk prediction, demonstrated that a high immune infiltration rate of TME was related to a favorable prognosis of NSCLC.

In a separate investigation by Wang et al., researchers attempted to address the unclear boundaries in conventional pathological image recognition by refining the method for identifying cell boundaries. Then, they proceeded to segment and classify cell nuclei, while representing blood vessels and necrosis using images of red blood cells and nuclear lysis. Finally, 48 features related to the cellular spatial organization were extracted and combined with the National Lung Screening Trial (NLST) dataset to develop a prognostic model, which was validated in TCGA. It was demonstrated that the predicted survival rate of the high-risk group was significantly lower than that of the low-risk group (p = 0.001) [42].

Park et al. [43] developed the TIL spatial analysis model Lunit SCOPE IO, which can segment and quantify tissue components in WSI images and was successfully utilized to predict the benefit of ICIs in patients with advanced NSCLC. Presently, numerous studies have used pathological image analysis to demonstrate the correlation of TMB, CD8⁺ T cells, regulatory T (T-reg) cells, and TILs with PD-1 therapy [29].

Just like PD-L1 and TMB, the compositional changes of TME are influenced by various factors [44]. Therefore, several studies have focused on multi-omics to predict TME.

In a recent study, experts used unsupervised cluster analysis with tumor immune microenvironment (TIME) data and survival data of 1906 adenocarcinoma patients. The resulting TIME score scoring model is characterized by significantly mutated genes (SMG), copy number variation (CNV), and cancer stemness. This model distinguishes immune infiltration and effectively predicts the sensitivity of immunotherapy and the accuracy of prognosis [44].

As is known to us, immunotherapy resistance may relate to dysregulated lactic acid metabolism that inhibits dendritic cell (DC) maturation, thereby minimizing T cell infiltration. A study screened a gene map associated with lactate metabolism with ML and verified the effectiveness of the seven screened genes related to lactate metabolism in predicting survival and immunotherapy efficacy in a cohort of adenocarcinoma patients [45].

Costimulatory molecules play a vital role in activating immune cells. However, the characterization of the many co-stimulatory molecular genes (CMGs) in LUAD is poorly understood. Thus, some scholars estimated the composition of stroma and immune cells in malignant tumor tissues through K-means clustering. They used SVM to screen out the CMG (CD80, LTB and TNFSF8) as the final markers to predict the TIME status of patients, achieving the purpose of predicting the effect of immunotherapy [46]. IHC verification of 16 specimens revealed a significant positive correlation between the screened biomarkers and the response to immunotherapy.

Taken together, these studies underscore the potential value of genomics in predicting TME and offer a promising direction for future research.

Similar to the utilization of radiomics for the prediction of other biomarkers, when utilizing image data to predict the efficacy of immunotherapy, the majority of studies rely on feature extraction, amalgamated with immunotherapy data, and modeled with a range of ML algorithms. Although adjuvant therapy such as immunotherapy can control the progression of lung cancer patients to a large extent and prolong patients’ survival time with subsequent radiotherapy, chemotherapy, and surgery, there are still many patients who do not benefit from immunotherapy or are not suitable for surgery.

Clinical benefit such as progression-free survival (PFS) is the top predictive value for radiomics-based AI models. When utilizing durable clinical benefit (DCB, PFS ≥ 6 months) and non-DCB (NDCB, PFS < 6 months) as the endpoints to conduct predictive models, researchers found that weighted radiomics signatures of multiple intrapulmonary lesions have the potential to predict long-term PFS benefit in PD-1/PD-L1 immunotherapy candidates, which aggregates performance across all models and yields excellent results (AUCs: 0.75–0.82) [47].

Additionally, different algorithms were used to predict the immunotherapy efficiency or response rate. Tang et al. extracted image features from 422 NSCLC patients, assessing radiological parameters using DT, boosting tree (BT), RF, SVM, generalized linear model (GLM), and deep learning artificial neural network (DL-ANN). An AUC > 0.7 result was obtained for the prognostic performance of omics features. RF achieved an excellent performance of AUC = 0.938 among the models. This comparative study demonstrates the value of ML algorithms in the prediction process of immunotherapy. It is believed that with the iterative update of modeling methods, AI-assisted clinical decision-making will eventually become a reality [48]. Another study focuses on CT-based short-term follow-up radiomics features, utilizing an SVM model to predict the response to immunotherapy and PFS in patients with advanced NSCLC [10].

Moreover, Gong et al. used short-term follow-up CT images of patients for radiological histological features extraction, and performed a SVM model to predict the response of advanced NSCLC to immunotherapy and the PFS of patients [49]. Tian et al. analyzed a cohort of 939 patients with IIIB-IV NSCLC using a deep convolutional neural network (Deep CNN) algorithm to train the model "PD-L1ES". The test set results showed that PE-L1ES was able to predict high PD-L1 expression (PD-L1 ≥ 50%) with the AUC = 0.76. Meanwhile, a low PD-L1ES score was associated with the improvement of PFS (p = 0.010). When features commonly used in clinical prediction models such as age, gender, smoking history, and family history of malignant tumors are added to this prediction model, it can better predict the response of patients to immunotherapy and improve the stratification ability of the model [50]. The results showed that PD-L1 classification with DL features and quantitative radiomics features are complementary, which may be one of the directions to optimize the model performance in the future. Follow-up studies can incorporate more comprehensive immunotherapy patients’ information to improve the performance of the model in predicting the effect of immunotherapy, too.

Furthermore, the tumor and tumor organismal environment (TOE) features of pre-treatment CT images were extracted. SVM was used to make a risk stratification model. AUC = 0.869 was achieved in the validation set, which assists in predicting the differences in treatment response and survival outcomes in patients with standard locally advanced non-small cell lung cancer (LANSCLC) after radical concurrent chemoradiotherapy [51]. Moreover, Yan's team developed a detection model from the LUNA16 public database using DL. They performed detection on the Anti-PD-1_Lung dataset, comparing it with the effectiveness of immunotherapy, and ultimately demonstrating the ability of the model to predict immunotherapy in lung cancer [10]. Simultaneously, Mu et al. [4] collected PET-CT images before receiving ICI treatment, extracted features, and modeled them to predict clinical outcomes such as OS and PFS of patients.

In general, the research on radiomics to predict lung cancer immunotherapy efficacy is very popular and various algorithm types based on various characteristics have shown good model performance.

Changes in genetic material, epigenetic information, oncogenic and tumor suppressor signals, and transcription factors in tumors also influence the expression of PD-L1 in addition to TME. A reliable assessment of treatment efficacy requires accurate characterization of these compositions [30].

Immune regulation-related gene profiles in patient biopsy samples were analyzed by Wiesweg et al. [52] to develop an ML model to predict the impact of genomic information and TME on the therapeutic response for IV NSCLC.

DNA methylation (DNAm) or RNA methylation can serve as an epigenetic marker to predict cancer recurrence risk at a molecular level. ML algorithms were used to identify 4 CpG methylation markers associated with cell proliferation markers, somatic changes, TMB and DNA damage response (DDR) genes in a recent study. These markers were combined with clinical stage and survival data to construct a risk score model. And the model effectively predicted recurrence-free survival and prognosis of NSCLC patients (p = 0.0002). This 4-DNAm marker panel was useful for NSCLC prognosis, treatment decision-making and assessment of treatment response [53]. Similarly, Shang et al. developed MeImmS, a DNAm scoring system that accounts for differences in the methylation status of 8 CpG islands. They analyzed its correlation with T cell exhaustion, immune regulation, and immune cell activation. The researchers also confirmed that the combination of DNA methyltransferase inhibitors (DNMTi) and ICIs has a favorable effect on the outcome of NSCLC patients [54]. m6A-mediated immune genes were also developed in a prediction model by various ML algorithms. The model demonstrated its applicability to predict survival and distinguish patients' TME, genomic background, chemotherapy response and immunotherapy response propensity. It shows the potential of m6A modification in changing the TME of LUAD, participating in tumorigenesis as well as predicting efficacy [55].

Acetylation is also a common and reversible epigenetic alteration that plays a critical role in the initiation and progression of malignant tumors. However, the prognostic value of acetylation-related genes in early-stage LUAD remains unclear. Some scholars collected acetylation-related genes in the transcriptome of early LUAD patients in the TCGA database to try to identify the important biomarkers of early LUAD recurrence through differential analysis and protein–protein interaction network construction. Finally, they concluded that the two gene signatures of RBBP7 and YEATS2 can be used to predict the recurrence-free survival of early LUAD [56].

Currently, many studies attempt to filter gene combinations [57, 58], transcriptional profile information, blood microRNA, etc., as features for modeling. And these studies have exhibited the predictive performance not inferior to existing markers.

Proteomics is an intuitive biomarker distinct from genomics for the reason that expressed proteins undergo post-translation modifications and interact directly with the host immune system and TME. Recently, mass spectrometry (MS) analysis of serum samples has been utilized to characterize independent proteome features and develop models for predicting clinical outcomes and side effects of patients [59].

Over 1,600 autoantibody biomarkers in serum were trained through an iterative unsupervised ML algorithm. Ultimately 13 were selected as features for modeling. The results showed that high expression of these features was associated with an overall 5-year survival rate of 7.6% in lung cancer patients, further highlighting the potential of serum proteomics in predicting survival [60].

In recent years, research pertaining to the gut microbiome has garnered significant attention. Numerous studies have explored the correlation between the gut microbiome and the effectiveness of immunotherapy. However, there exists a dearth of models that predict the correlation between them.

Liu et al. conducted an analysis of gut microbiome samples from 79 NSCLC patients who underwent immunotherapy. The study utilized RF and multilayer perceptron (MLP) neural network models to predict PFS. Notably, the results indicated that two prediction models based on function rather than classification were both AUC ≥ 0.95. These outcomes suggest that the models can effectively predict the potential benefits of immunotherapy for NSCLC patients. Furthermore, it highlights the promising future of gut microbiome analysis in predicting cancer immunotherapy outcomes [52].

Due to the intricate mechanisms governing anti-tumor immune responses, conventional biomarkers are inadequate in dynamically illustrating tumor-immune system interactions and identifying patients who would benefit from ICI therapy [61]. Consequently, there is an urgent need to develop new markers. Recent research has focused on mining simple and readily available laboratory blood data. Some researchers collected medical history and laboratory examination data of NSCLC patients receiving ICI treatment and utilized various ML algorithms to model and screen predictive feature markers. Among these, the neutrophil to lymphocyte ratio (NLR) has exhibited predictive value for disease control rate (DCR) and 6-month survival in multiple studies [62, 63]. In addition to directly using routine inspection data, studies have employed techniques such as MS to obtain phenotypic characteristics of immune cells. By applying algorithms such as RF, partial least squares discriminant analysis, MLP and elastic network, scholars have discovered that B cell-related phenotypes can be used as features to distinguish healthy, responders and baseline non-responders and predict the ability to respond to immunotherapy [64].

Furthermore, another study employed 10 ML algorithms to screen data from three NSCLC immunotherapy cohorts for TMB, intratumoral heterogeneity and loss of heterozygosity for human leukocyte antigen. By combining these three genomic biomarkers using the SVM-poly method, a model was constructed to predict DCB, with the model exhibiting an AUC of 0.78 [61]. Expanding the training samples in such models is expected to enable the development of additional biomarkers, providing more personalized treatment options and promoting precision medicine.

Approximately one-third of patients who received immunotherapy experienced irAEs [65]. The progression of irAEs is the primary reason for discontinuing immunotherapy. Meanwhile, irAEs are a significant factor that affects the prognosis of patients [66, 67]. Currently, there are no biomarkers that can predict the early occurrence of irAEs, and few studies have been conducted in this regard. Therefore, creating a painless, accurate and standardized prediction method is a considerable challenge that requires further research [68].

Neural network models were utilized to predict skin irAEs caused by PD-L1 therapy considering variables such as tumor type, treatment drug, age, autoimmune history, derived NLR, lactate dehydrogenase, albumin, body mass index, Eastern Cooperative Oncology Group performance status (ECOG PS) and tumor M-stage features. The results demonstrated that ML has a high sensitivity and ability to predict cutaneous irAEs in the early stages of immunotherapy, with a positive predictive value (PPV) of 76.5% (± 10.5%), a negative predictive value (NPV) of 79.4% (± 11.9%), a sensitivity of 85.3% (± 8.8%), and a specificity of 67.6% (± 15.8%) [67].

Cardiac irAEs could be deadly. XGBoosted decision tree was used to predict the probability of cardiac adverse events in patients receiving PD-1 or PD-L1 therapy by modeling 356 basic medical history information as potential risk factors. In this study, 4,960 patients receiving PD-1/PD-L1 therapy were included, among whom 418 experienced cardiac events. The final model showed that age, corticosteroids, heart disease history drugs, extreme body weight, low lymphocyte percentage and high neutrophil percentage were associated with the occurrence of cardiac adverse events, but the model had limited predictive value with an AUC of 0.65 [69].

To predict adverse events in atezolizumab-treated advanced NSCLC patients, seven ML methods were employed to explore the role of 21 blood markers. XGBoost and LASSO methods performed the best, and the AUC of XGBboost for 10 markers was 0.692. However, even after narrowing down to the combination of C-reactive protein (CRP), platelet-to-lymphocyte ratio (PLR), and thyroid-stimulating hormone (TSH), the predictive effect was still not satisfactory, despite the high consistency between the training set and the test set [68].

It is evident that the results of irAE prediction studies have been unsatisfactory, and the biomarkers screened out by ML are just primarily related to adaptive immunity, inflammatory state, liver and thyroid function. Although these indicators correspond to clinical symptoms such as fatigue, vomiting and elevated transaminases, their abnormality already signifies that irAEs are in progress, rendering them of little predictive value. Therefore, exploring inflammatory pathway targets of different types of irAEs from a mechanistic perspective and realizing early prediction and targeted therapy of irAEs are future research directions (Fig. 4).