Harnessing multi-omics and machine learning for predicting immune checkpoint blockade responses: Advances, challenges, and future directions

Abstract

Immune checkpoint blockade (ICB) therapies have revolutionized cancer treatment, showing success across various cancer types. However, there is variability in response rates among different cancers and individual patients. This highlights the critical need for precise patient stratification. Machine Learning and Deep Learning models are increasingly utilized to predict ICB responses by integrating multi-omics data, such as clinical, genomic, radiomic, and transcriptomic information. This review outlines the key methodologies of these predictive models. It underscores their role in enhancing response prediction. We delve into the advanced mechanisms of ICB response and discuss the biological foundations that inform these models. This demonstrates how basic research informs clinical application. We aim to offer comprehensive insights into how artificial intelligence can optimize patient stratification for ICB therapy.

1. Introduction

Immune checkpoint blockade (ICB) therapies have revolutionized oncology by targeting immune checkpoints. They target immune checkpoints, which are key regulators of immune homeostasis that tumors exploit to evade immune detection. ICB restores cytotoxic T-cell activity and amplifies systemic immune responses using immune checkpoint inhibitors (ICIs). The primary targets of ICIs include programmed cell death protein 1 (PD-1), its ligands PD-L1/2, and cytotoxic T lymphocyte-associated antigen 4 (CTLA-4). ICIs like pembrolizumab and nivolumab disrupt PD-1/PD-L1 interactions, restoring T cell function to attack cancer cells. Similarly, ipilimumab, an anti-CTLA-4 antibody, blocks CTLA-4’s competition with CD28 for B7 molecules (CD80/CD86), enhancing T cell activation and proliferation. Beyond PD-1/PD-L1 and CTLA-4, emerging targets such as LAG-3, TIM-3, and TIGIT are under investigation. These targets aim to expand the therapeutic landscape and address resistance to current ICB therapies [1].

Since the first Food and Drug Administration (FDA) approval of ipilimumab in 2011 for metastatic melanoma, ICIs have been approved for various cancers, including non-small cell lung cancer (NSCLC) and renal cell carcinoma (RCC). However, response rates vary significantly due to differences in tumor microenvironments and patient-specific factors. For instance, melanoma shows high overall response rates (ORRs) of 40%–50%, while NSCLC and RCC exhibit ORRs of 20%–50% and 25%–30%, respectively. Other cancers, such as breast cancer, often demonstrate even lower responses [2]. Additionally, some patients experience hyperprogressive disease (HPD), characterized by accelerated tumor progression post-treatment, observed in ∼20% of metastatic NSCLC patients treated with PD-1/PD-L1 inhibitors [3]. Furthermore, ICB can exacerbate autoimmune conditions, particularly in patients with pre-existing disorders, underscoring the need for precise patient stratification to optimize treatment outcomes [4].

As a result, precise patient stratification is critical for optimizing the outcomes of ICB therapy. Accurate stratification not only identifies patients most likely to benefit from ICB therapy but also reduces risks for non-responders. It enhances cost-effectiveness and facilitates biomarker discovery for predicting responsiveness or resistance. Machine learning (ML) models are increasingly employed to achieve this by analyzing diverse data modalities, including clinical data, digital pathology, radiomics, and multi-omics (genomics, transcriptomics). Clinical data provide insights into patient demographics, comorbidities, and treatment histories. Digital pathology and radiomics uncover histopathological and imaging features linked to tumor heterogeneity. Genomic and transcriptomic data reveal mutations, gene expression profiles, and immune pathway activities. Conventional ML models such as logistic regression (LR), random forest (RF), support vector machines (SVM), and Extreme Gradient Boosting (XGBoost) have shown promise in integrating these modalities to improve patient stratification and treatment outcomes [5].

Deep learning models further enhance ICB response prediction by automating feature extraction and integrating complex, high-dimensional datasets. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) excel in uncovering hierarchical patterns. These include spatial relationships between tumor-infiltrating lymphocytes and tumor cells in digital pathology or subtle variations in tumor texture and shape in radiomics [6]. Natural language processing (NLP) techniques expand DL’s scope by analyzing clinical text data from electronic health records, including patient histories and genomic profiles [7].

This review delineates the biological mechanisms underlying ICB responses. It also explores how artificial intelligence (AI) models integrate prior biological knowledge to optimize clinical outcome prediction (Table 1). We emphasize how these approaches harness critical biological insights derived from multi-omics data. This simultaneously improves predictive performance and mechanistic understanding (Table 2). We also introduce model details for further disease and modality adaptation (Table 3). This work presents a conceptual framework for advancing ICB therapy, with implications for both research and clinical practice.

Table 1.

Representative AI models developed for predicting ICB response based on diverse feature types.

Features	Data source	Cancer type & therapy context	Model design	Performance	Ref.
Tumor mutation burden (55 SNVs)	Training: POPLAR/OAK cohort (n = 429); Validation: UCMC cohort (n = 137); MSKCC cohort (n = 349)	All: NSCLC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	DL regression: 1D-CNN; cML: SVM, RF, and LR	PFS: HR = 3.70 [3.16-4.33] , P < 0.001 OS: HR = 3.20 [2.52–4.06], P < 0.001	Peng et al. [16]
mTMB, SCNA, and MSI	Training: TCGA (n = 8646); Validation: Van Allen's study (n = 108) and Snyder's study (n = 64)	Training: 29 tumor types; validation Melanoma (Anti-CTLA-4)	DL classification: DBN for feature extraction, DAE for stratification analysis	GC1/GC2 vs GC3/GC4 OS: P < 0.001	Xie et al. [18]
Neoantigen burden and antigen presentation features	Training: ISCI cohort (n = 51); Validation: dbGaP (n = 110)	All: Melanoma (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML regression: XGBoost models (neoantigen prediction & HLA LOH detection); LR to calculate NEOPS score	AUC = 0.76	Abbott et al. [21]
TCR CDR3β sequences, antigen peptide sequences, class I MHC alleles.	Training: peer-reviewed publications (n = 32,607 TCR-pMHCs pairs); four ICB cohorts (n = 187)	Validation: Melanoma; ccRCC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	DL classification: pMTnet: transfer learning-based model; LSTM	NIES-high vs NIES-low: OS: HR = 2.04 [1.37–3.05], P < 0.001	Lu et al. [23]
11 tumor immunogenic and immune response-related features	Training: CPI1000+ cohort (n = 1008); Validation: KEYNOTE-028 study (n = 76), UHE study(n = 121), and Samsung MC study (n = 144)	Training: Melanoma, HNC, UC, RCC, NSCLC, and CRC (Anti-PD-1/PD-L1/CTLA-4); Validation: Pan-cancer; Melanoma; NSCLC (Anti-PD-1)	ML classification: XGBoost	AUC = 0.86	Litchfield et al. [24]
16 tumor immunogenic features and demographic features	MSK-IMPACT ICB cohort (n = 1479; Training: n = 1184; Validation: n = 295)	All: 16 cancer types (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: RF	AUC = 0.79	Chowell et al. [25]
The abundance of 22 immune cell types	Training: three GEO datasets (n = 91); Validation: two dbGAP datasets (n = 85)	Training: Melanoma; Validation: Melanoma and mGC (Anti-PD-1/CTLA-4)	ML classification: SVM	AUC = 0.84	Miao et al. [30]
The abundance of 51 immune cell types; PD-L1 expression; TMB	ArrayExpress and GEO databases (n = 2166)	Training: 24 cancer types; Validation: Melanoma, HCC, GC, RCC, and NSCLC (Anti-PD-1/PD-L1/CTLA-4)	ML classification: RF	Average AUC = 0.76	Zaitsev et al. [31]
75 genes related to anti-tumor immunity core pathways	Training: TCGA- SKCM (n = 453); Validation: ICBatlas datasets (n = 505)	Training: Melanoma; Validation: 15 cancer types (Anti-PD-1/PD-L1/CTLA-4)	ML classification: RSF	Average AUC = 0.76	Zhang et al. [32]
29 functional gene expression signatures of TME	Training: TCGA-SKCM (n = 470); Validation: GEO datasets (n = 407)	Training: Melanoma; Validation: UC, GC, and LUSC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: LR	AUC = 0.78	Bagaev et al. [36]
Communication probabilities of L-R pairs within specific pathways	GEO datasets (n = 685)	All: Melanoma, NSCLC, BC, GC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: LR	Average AUC = 0.79	Lee et al. [37]
80 ICGs genes associated with immune response	Training: FANTOM5 project and TCGA datasets; Validation: NCBI SRA database (n = 256)	Training: 33 cancer types; validation: GC and melanoma (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: LR	Average AUC = 0.75	Hu et al. [39]
Spatial features: cell density, proliferative fractions, cell-cell interactions	NeoTRIP cohort (n = 279)	All: TNBC (Anti-PD-1)	ML classification: LR, LASSO, RF	AUC = 0.82	Wang et al. [41]
TCR repertoire diversity features	Training: TCGA (n = 667); Validation: Riaz cohorts (n = 333)	Validation: Melanoma (Anti-PD-1)	ML classification: elastic net regression models	TCR entropy high vs low OS: P = 0.0049	Bortone et al. [45]
CD4 T memory cell abundance and TCR clonotype diversity	Lozano’s cohort (n = 93): training (n = 26); validation (n = 27)	All: Melanoma (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: LR	AUC = 0.90	Lozano et al. [46]
TCR and HLA sequences	Training: Checkmate-038 cohorts (n = 83); Validation: Yost’s study (n = 11) and sade's study (n = 19).	Training: Melanoma Validation: BCC and SCC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	DL classification: VAE, MIL	AUC = 0.86	Sidhom et al. [47,48]
Cancer stemness gene signatures	Discovery:TISCH portal (n = 345) 10 ICI cohorts from GEO datasets: training(n = 620); validation (n = 154); test (n = 149)	Melanoma, UC, GBM, GC, and RCC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: Naïve Bayes model	Average AUC = 0.71	Zhang et al. [53]
T cell exhaustion gene signatures	Discovery: Zhang cohort (n = 26); Training: Maria cohort (n = 298) validation: GHR cohort (116) and Nathanson cohort (n = 9)	Discovery: TNBC; Training: UC and BC; Validation: Melanoma, UC and BC (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	DL & ML: DeepAKR (deep autoencoder), XGBoost	Average AUC = 0.778	Zhang et al. [54]
Gene signatures related to ICB response of tumor cells, lymphocytes, myeloid cells, and stromal cells	Discovery: Sun’s cohort (n = 11) Training and validation: four SKCM cohorts (training, n = 195; validation, n = 49)	All: Melanoma (Anti-PD-1)	ML classification: Adaboost	AUC = 0.847	Sun et al. [55]
Tumor intrinsic, immune activation, and immune evasion features	Discovery and training: TransNEO study (n = 168); Validation: PBCP study and ARTemis trial (n = 75)	All: Breast cancer (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML ensemble model: SVM, RF, and the LR model	AUC = 0.87	Sammut et al. [57]
Gut microbial signatures and clinical features	Discovery and training: CA209–538 cohort (n = 108); Validation: 6 other comparable cohorts (n = 383)	Discovery and training: UGB, NEN, and GYN (Anti-CTLA-4 + anti-PD-1); Validation: Melanoma (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: RF	Average AUC = 0.75	Gunjur et al. [58]
Routine blood tests, features and clinical variables	Training: MSKCC cohorts (n = 1628); Validation: MSKCC cohort (n = 407), MSHS cohort (n = 1159), 10 global ICI phase 3 clinical trials (n = 4447)	All: Pan-cancer (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	ML classification: ensemble model: ridge Cox regression, SVM, RSF	Clinical benefit: AUC = 0.714; OS: AUC = 0.736	Yoo et al. [56]
Radiolomics, patholomics, and genomics features	Training: MSK IMPACT cohort (n = 247); Validation: radiology cohort (n = 50) and pathology cohort (n = 52)	All: NSCLC (Anti-PD-1)	DL & ML classification: DyAM model: LR, logistic regression, Multimodal attention	AUC = 0.8	Vanguri et al. [60]
H&E, CT, clinical variables, and longitudinal CT sequences	Training and validation: ESCC PD-L1 cohort (n = 220), ESCC immunotherapy cohort (n = 75)	All: ESCC (Anti-PD-1/PD-L1)	DL classification: ResNet50, LASSO, multi-head self-attention fusion, RNN	Survival prediction: 1-/3-year AUC = 0.802 High vs low risk HR = 2.46; Treatment response: AUC = 0.80; AUC = 0.937 (with longitudinal CT)	Liu et al. [61]
Blood markers, imaging features, and medication ATC codes;	Training and validation: NKI cohorts (n = 694)	All: pan-cancer (Anti-PD-1/PD-L1 or anti-CTLA-4 + anti-PD-1)	DL classification: transformer, temporal attention, MLP classifier	Survival prediction: AUC = 0.84 (3mo), AUC = 0.83 (6mo), AUC = 0.82 (9mo)	Yeghaian et al. [63]
Longitudinal multiparametric MRI and blood-based CEA levels	The SAH and SYUCC cohort; Training cohort (n = 321); internal validation cohort (n = 160); external validation cohort (n = 141)	All: rectal cancer (Neoadjuvant chemoradiotherapy)	DL classification: CNN, Siamese subnetworks	Imaging-based model: AUC = 0.92; combined model with CEA level: AUC = 0.97	Jin et al. [62]
44 biologically grounded immune concepts in TME	Training cohorts: TCGA (n = 10,184) Validation cohorts: 16 independent clinical cohorts (n = 1133)	BLCA, GBM, KIRC, LUAD, LUSC, SKCM, and STAD (Anti-PD1, anti-PD-L1, anti-CTLA4, and combination therapies)	DL classification: Transformer-based foundation model with concept bottleneck	To unseen cancer types (stomach adenocarcinoma): AUC = 0.837	Shen et al. [66]
Germline SNVs weighted by PPI network	Private cohort and public datasets. (n = 737)	Melanoma, bladder, breast, lung, colon or rectum, head and neck, and renal cancer. (Anti-PD1, anti-PD-L1, anti-CTLA4, and combination therapies)	DL classification: Double-channel attention neural network	Average AUC = 0.95	Yan et al. [67]

TMB, Tumor Mutation Burden; SNV, Single Nucleotide Variant; UCMC, University of Chicago Medical Center cohort; MSKCC, Memorial Sloan Kettering Cancer Center cohort; NSCLC, Non-Small Cell Lung Cancer; PD-1, Programmed Cell Death Protein 1; PD-L1, Programmed Death-Ligand 1; CTLA-4, Cytotoxic T-Lymphocyte-Associated Protein 4; DL, Deep Learning; CNN, Convolutional Neural Network; ML, Machine Learning; SVM, Support Vector Machine; RF, Random Forest; LR, Logistic Regression; PFS, Progression-Free Survival; HR, Hazard Ratio; OS, Overall Survival; mTMB, modified Tumor Mutation Burden; SCNA, Somatic Copy Number Alteration; MSI, Microsatellite Instability; TCGA, The Cancer Genome Atlas; dbGaP, Database of Genotypes and Phenotypes; DBN, Deep Belief Network; DAE, Deep Autoencoder; XGBoost, Extreme Gradient Boosting; NEOPS, Neoantigen Presentation Score; AUC, Area Under Curve; TCR, T-Cell Receptor; pMHCs, Peptide-Major Histocompatibility Complexes; ccRCC, Clear Cell Renal Cell Carcinoma; pMTnet, Peptide-MHC Transfer Network; LSTM, Long Short-Term Memory; NIES, Neoantigen Immunogenicity Effectiveness Score; CPI, Checkpoint Inhibitor; HNC, Head and Neck Cancer; UC, Urothelial Carcinoma; CRC, Colorectal Cancer; GEO, Gene Expression Omnibus; mGC, Metastatic Gastric Cancer; HCC, Hepatocellular Carcinoma; ICBatlas, Immune Checkpoint Blockade Atlas; RSF, Random Survival Forest; SKCM, Skin Cutaneous Melanoma; SRA, Sequence Read Archive; TME, Tumor Microenvironment; GC, Gastric Cancer; LUSC, Lung Squamous Cell Carcinoma; ICGs, Immune Checkpoint Genes; FANTOM5, Functional Annotation of the Mammalian Genome 5; NCBI, National Center for Biotechnology Information; TNBC, Triple-Negative Breast Cancer; LASSO, Least Absolute Shrinkage and Selection Operator; VAE, Variational Autoencoder; MIL, Multiple Instance Learning; BCC, Basal Cell Carcinoma; SCC, Squamous Cell Carcinoma; scRNA-seq, Single-cell RNA sequencing; GBM, Glioblastoma Multiforme; DeepAKR, Deep Autoencoder-based Kruskal–Wallis and RFE model; RFE, Recursive Feature Elimination; SAGA, Stochastic Average Gradient Descent; ICI, Immune Checkpoint Inhibitor; MSK IMPACT, Memorial Sloan Kettering Integrated Mutation Profiling of Actionable Cancer Targets; DyAM, Dynamic Deep Attention-based Multiple-instance Learning Model with Masking; H&E, Hematoxylin and Eosin; ESCC, Esophageal Squamous Cell Carcinoma; CT, Computed Tomography; ResNet50, Residual Network 50 Layers; RNN, Recurrent Neural Network; NKI, Netherlands Cancer Institute; ATC, Anatomical Therapeutic Chemical Classification System; MLP, Multilayer Perceptron. CEA, Carcinoembryonic Antigen; SAH, Sixth Affiliated Hospital; SYUCC, Sun Yat-sen University Cancer Center; BLCA, Bladder Cancer; KIRC, Kidney Renal Clear Cell Carcinoma; LUAD, Lung Adenocarcinoma; STAD, stomach adenocarcinoma; PPI, Protein-Protein Interaction.

Table 2.

Representative biomarkers for predicting ICB response.

Biomarkers	Association with favourable clinical outcome (Biological rationale)	Cancer type applicability	Sample type for assessment	Representative assay methodologies	Ref.
Tumor genomic mutation features: High TMB, dMMR/MSI-H status, low SCNAs (positive).	Correlates with increased neoantigen load, enhanced tumor immunogenicity, and improved T-cell recognition.	Multiple cancer types (e.g., Melanoma, NSCLC, CRC, GC)	Blood or tumor tissue	NGS WES, IHC or targeted gene panel sequencing	[18]
Tumor-specific mutated genes: BRAF, PBRM1, SERPINB3, and SERPINB4 mutations (positive); KRAS and LKB1, PTEN mutations (negative).	Enhance tumor immunogenicity or disrupt immunosuppressive pathways (positive predictors); impair critical components of the anti-tumor immune response (negative predictors).	Melanoma (PTEN, SERPINB3 and SERPINB4); NSCLC (KRAS and LKB1); RCC (PBRM1); UC (BRAF)	Tumor tissue	NGS WES or targeted gene panel sequencing	[13]
Host genetics features: high level of impact mutations on HLA-I and β2M genes, LOH at HLA class I alleles (negative).	Impairs neoantigen presentation, reducing immune recognition by T cells.	Melanoma, NSCLC	Blood or tumor tissue	NGS or PCR-based HLA typing	[21]
Neoantigen burden features: low neoantigen burden, high level of ITH (negative).	High ITH can diminish effective immunogenicity, as neoantigens restricted to subclones may elicit weaker or less T cell responses.	Multiple cancer types especially in NSCLC and Melanoma	Tumor tissue	WES + RNA-Seq	[23]
Neoantigen immunogenicity features: high TCR-pMHC binding specificity, low neoantigen fitness (positive).	Neoantigens possessing high immunogenicity are more effectively recognized by T cells, driving a potent anti-tumor immune response.	Multiple cancer types	Blood or tumor tissue	tetramer analysis, T-scan or TCR sequencing + HLA typing + peptide prediction (NetMHCpan)	[23]
Immune cell population composition features: high TIL density, high PD1+TIL, high CD8+TCF7+/ CD8+TCF7− TILs ratio (positive).	This profile represents an immunologically active TME, conducive to an effective anti-tumor response.	Melanoma, NSCLC	Tumor tissue	RNA-Seq or IHC	[26]
T cell response signatures: IFNγ signatures, T cell cytotoxic signatures (positive); TIDE (negative).	High expression of positive prognostic gene signatures correlates with effective T cell responses, and elevated expression of genes in negative signatures suppresses T cell function.	Multiple cancer types	Tumor tissue or blood	RNA-Seq	[34]
Immune checkpoint genes expression features: high PD-L1 expression, high IMPRES (positive).	High expression of immune checkpoint genes reflects critical immune cell interactions and effective anti-tumor response.	NSCLC, melanoma, urothelial (PD-L1 expression); Melanoma (IMPRES)	Tumor tissue	RNA-Seq	[38]
Spatial cell-cell interaction features: TLS, macrophage and T cell interactions (positive).	Specific patterns of immune cell-cell interactions are indicative of an effective anti-tumor response.	Multiple cancer types	Tumor tissue	IF, IHC, IMC, and so on	[40]
TCR repertoire features: high clonality, low gini-simpson index, or low Shannon entropy (positive).	Effective T cell responses induce the expansion of tumor-specific clones, thereby decreasing TCR repertoire evenness and resulting in high clonality.	Multiple cancer types	Tumor tissue or blood	RNA-Seq, TCR-seq	[46]
Commensal microbiota features: high gut microbial diversity (positive); Specific gut microbial species (Akkermansia muciniphila (positive), Bacteroidales spp. (negative).	Diverse microbiota enhances systemic immunity and primes T-cell responses. Akkermansia muciniphila promotes antitumor immunity via TLR4 activation and DC maturation; Bacteroidales spp. associated with immunosuppressive metabolites.	Melanoma, NSCLC, RCC (high gut microbial diversity and Akkermansia muciniphila); Melanoma (bacteroidales spp.)	Stool	qPCR, metagenomics	[13]
Radiology and pathology features: tumor volume reduction (CT), high baseline metabolic activity (SUVmax on PET/CT), texture, shape, intensity, and heterogeneity information of tumor and peritumoral (positive)	These biomarkers may indicate higher baseline inflammation/proliferation, suggesting a more immunologically active TME susceptible to ICB	Multiple solid tumors	Tumor tissue	CT/PET/MRI imaging (Radiology); H&E slides, PD-L1 IHC slides, WSI, mIF (Pathology)	[6]
Demographic features: age (conditional); sex-female, history of smoking (positive);	Age-related immunosenescence may impair ICB response, female hormonesmay enhance anti-tumor immunity, while males exhibit higher baseline inflammation inducing resistance. Smoking history is associated with increasing TMB and upregulates PD-L1 expression.	Melanoma, NSCLC	EHR	Clinical reports	[13]
Clinical features: poor ECOG performance status, history of autoimmune disease, High NLR, Low-ALB (negative); Prior lines of therapy, BMI (conditional).	Poor ECOG status, high NLR, low ALB, and autoimmune history often reflect systemic inflammation and immunosuppressive environments. Prior treatments enhance immunogenicity or exacerbate T-cell dysfunction. BMI ≥30 may enhance anti-tumor immunity or promote T cell exhaustion.	Melanoma, NSCLC	HER or blood	Routine blood tests and clinical reports	[13]

TMB, Tumor Mutational Burden; dMMR, Deficient Mismatch Repair; MSI-H, Microsatellite Instability-High; SCNAs, Somatic Copy Number Alterations; NGS, Next-Generation Sequencing; WES, Whole Exome Sequencing; IHC, Immunohistochemistry; BRAF, B-Raf Proto-Oncogene; PBRM1, Polybromo 1; SERPINB3, Serpin Family B Member 3; SERPINB4, Serpin Family B Member 4; KRAS, Kirsten Rat Sarcoma Viral Oncogene; LKB1, Liver Kinase B1; PTEN, Phosphatase and Tensin Homolog; HLA, Human Leukocyte Antigen; β2M, Beta-2 Microglobulin; LOH, Loss of Heterozygosity; ITH, Intratumoral Heterogeneity; RNA-Seq, RNA Sequencing; TCR, T Cell Receptor; TME, Tumor Microenvironment; IFNγ, Interferon Gamma; TIDE, Tumor Immune Dysfunction and Exclusion; IMPRES, Immune-Related Prognostic Score; TLS, Tertiary Lymphoid Structures; IMC, Imaging Mass Cytometry; TLR4, Toll-Like Receptor 4; DC, Dendritic Cell; SUVmax, Maximum Standardized Uptake Value; PET, Positron Emission Tomography; MRI, Magnetic Resonance Imaging; H&E, Hematoxylin and Eosin; WSI, Whole Slide Imaging; mIF, Multiplex Immunofluorescence; EHR, Electronic Health Records; ECOG, Eastern Cooperative Oncology Group; NLR, Neutrophil-to-Lymphocyte Ratio; ALB, Albumin; BMI, Body Mass Index.

Table 3.

Training strategies of representative deep learning models for predicting ICB response.

Model name	Model architecture	Datasets	Validation strategy	Performance metrics	Model interpretation	Ref.
3D RP-Net	Key Concepts: multi-task learning, Siamese subnetworks for longitudinal comparison, multi-scale feature integration, and depth-wise convolution Modalities: longitudinal multiparametric MRI and blood-based CEA Integration: late integration with RF	Source: the Sixth Affiliated Hospital, Sun Yat-sen University, and the Sun Yat-sen University Cancer Center. Cancer types: locally advanced rectal cancer. Therapy: Neoadjuvant chemoradiotherapy followed by total mesorectal excision. Training/test cohorts: Training cohort:n = 321 Internal validation cohort:n = 160 External validation cohort:n = 141	Independent testing on both an internal cohort from the same institution and an external cohort from a second institution	AUC score for response prediction	Feature map visualization: key channels were visualized to identify their relationship to pathophysiologic characteristics like tumor invasion, mesorectum invasion, and extramural vascular invasion	Jin et al. [62]
DyAM	Key concepts: attention mechanism, MIL, masking for missing modalities. Modalities: radiology, pathology, genomics. Integration: attention-based fusion	Source: memorial sloan kettering cancer center Cancer types: NSCLC (79% adenocarcinoma, 15% squamous) Therapy: anti-PD-(L)1 (n = 235), anti- PD-(L)1, anti-CTLA-4 combination (n = 12) Training/test: 10-fold cross-validation Training cohorts: multimodal (n = 247) Validation cohorts: radiology (n = 50), pathology (n = 52)	10-fold cross-validation, independent validation cohorts, permutation testing, subsampling analysis.	AUC score for response prediction, hazard ratio for prognosis prediction.	Attention weight	Vanguri et al. [60]
DeepTCR	Key concepts: MIL, multi-head attention, variational autoencoder Modalities: TCR sequence data, HLA genotype information Integration: late integration	Key concepts: attention mechanism, MIL, masking for missing modalities. Modalities: radiology, pathology, genomics. Integration: attention-based fusion	Independent external cohorts. Unsupervised embedding validation.	AUC score for response prediction. log-rank test for PFS stratification.	TCR motif analysis	Sidhom et al. [47]
MMTSimTA	Key concepts: simple temporal attention, transformers, longitudinal data analysis, multimodal dropout. Modalities: non-invasive, longitudinal data of blood test, CT-based organ volumes, prescribed medications. Integration: intermediate fusion	Source: Netherlands Cancer Institute. Cancer types: 11 different tumor types. Therapy: anti-PD1, anti-PD-L1, anti-CTLA4, and combination therapies. Training/test: three-fold stratified cross-validation Training cohorts: 694 patients Validation cohorts: no independent validation cohort	Early (up to 3 months) and late (up to 6 months) on-treatment data for evaluation.	AUC for evaluating the predictive performance of the models for survival at 3, 6, 9, and 12 months.	Not available.	Yeghaian et al. [63]
DANN	Key concepts: separate horizontal and vertical attention neural networks, random walk on a PPI network Modalities: germline WES data. Integration: the model processes a gene-expression matrix through both attention modules and is combined via a hadamard product	Source: a combination of a private cohort from the authors' lab and several public datasets. Cancer types: Melanoma, bladder, breast, lung, colon or rectum, head and neck, and renal cancer. Therapy: anti-PD1, anti-PD-L1, anti-CTLA4, and combination therapies. Training/test cohorts: a total of 737 patients were included in the study.	Five-fold stratified cross-validation was used to train and test the model. Ablation studies were also performed to validate the contribution of the attention modules	AUC score for response prediction	Gene importance analysis using importance scores generated by the horizontal attention module to rank PPI network genes.	Yan et al. [67]
COMPASS	Key concepts: self-supervised contrastive learning, transformer-based gene language model, hierarchical projector. Modalities: tumor transcriptomic data. Integration: hierarchical integration.	Source: TCGA dataset, 16 independent clinical cohorts. Cancer types: BLCA, GBM, KIRC, LUAD, LUSC, SKCM, and STAD. Therapy: anti-PD1, anti-PD-L1, anti-CTLA4, and combination therapies. Training/test: pretraining, fine-tuning and evaluation Training cohorts: TCGA database (n = 10,184) Validation cohorts: 16 independent clinical cohorts (n = 1133)	Leave-one-cohort-out evaluation, intra-cohort validation, cohort-to-cohort transfer analyses, multi-stage fine-tuning.	Precision and AUPRC for response prediction. log-rank test for overall survival stratification.	SHAP analysis to measure the impact of the 44 high-level concepts, concept score analysis.	Shen et al. [66]

CEA, Carcinoembryonic Antigen; AUC, Area Under the Curve; MIL, Multiple Instance Learning; NSCLC, Non-Small Cell Lung Cancer; PD-(L)1, Programmed Death-(Ligand) 1; CTLA-4, Cytotoxic T-Lymphocyte Antigen 4; TCR, T-Cell Receptor; HLA, Human Leukocyte Antigen; PFS, Progression-Free Survival; CT, Computed Tomography; PPI, Protein-Protein Interaction; WES, Whole Exome Sequencing; TCGA, The Cancer Genome Atlas; BLCA, Bladder Urothelial Carcinoma; GBM, Glioblastoma Multiforme; KIRC, Kidney Renal Clear Cell Carcinoma; LUAD, Lung Adenocarcinoma; LUSC, Lung Squamous Cell Carcinoma; SKCM, Skin Cutaneous Melanoma; STAD, Stomach Adenocarcinoma; AUPRC, Area Under Precision-Recall Curve; SHAP, SHapley Additive exPlanations.

2. The role of the cancer-immunity cycle in ICB therapy

The efficacy of ICB therapy is intricately linked to the cancer-immunity cycle, a series of stepwise and iterative events (Fig. 1). This cycle initiates with the release of cancer antigens, particularly neoantigens. These antigens are tumor-specific, formed by somatic mutations within the tumor, and are captured and presented by dendritic cells (DCs). They then migrate to secondary lymphoid organs to prime and activate effector T cells. Once activated, T cells circulate via the bloodstream and infiltrate tumors, where they recognize and bind to cancer cells through interactions between T cell receptors (TCRs) and antigens presented by major histocompatibility complex (MHC) molecules. This recognition triggers T cells to destroy cancer cells, thereby releasing more tumor antigens and further amplifying the immune response.

Fig. 1.

ICB response biomarkers within the cancer-immune cycle. Researchers have identified diverse biomarkers of ICB response, capturing both cancer-intrinsic and cancer-extrinsic factors that influence the cancer-immune cycle. Cancer-intrinsic factors primarily determine tumor immunogenicity, which is critical for initiating the immune response. On the other hand, cancer-extrinsic factors reflect the overall immune activity across lymph nodes, peripheral blood, and the tumor microenvironment. These involve multiple steps, from DCs presenting neoantigens to T cells executing tumor cell killing. This figure provides an overview of key biomarkers associated with each step of the cancer-immune cycle, highlighting their role in the ICB response. ICB, immune checkpoint blockade; DC, dendritic cell; TCR, T cell receptor; TME, tumor microenvironment; TMB, Tumor Mutational Burden; dMMR, Deficient Mismatch Repair; MSI-H, Microsatellite Instability-High; SCNAs, Somatic Copy Number Alterations; HLA, Human Leukocyte Antigen; LOH, Loss of Heterozygosity; ITH, Intertumoral Heterogeneity; IFN, Interferon; PD-L1, Programmed Death-Ligand 1.

The cancer-immunity cycle within the host is governed by both positive and negative feedback mechanisms. Positive feedback mechanisms enhance the immune system’s capacity to target and eliminate cancer cells. When DCs present tumor antigens to T cells, this initiates a cascade of immune activation. The activated T cells proliferate and differentiate, boosting the number of effector T cells that target the tumor. These cells secrete cytokines such as IL-2, IFN-γ, and TNF-α, which enhance the immune response by increasing antigen presentation, upregulating MHC on tumor cells, and attracting additional immune cells like macrophages and NK cells. This forms a positive feedback loop that continuously targets tumor cells [8].

Conversely, negative feedback mechanisms are crucial for modulating immune activity and preventing autoimmunity but can be exploited by tumors to dodge immune surveillance. Key inhibitory pathways include PD-1/PD-L1 and CTLA-4. PD-1 on activated T cells interacts with PD-L1, which is often upregulated on tumors, to dampen T cell activity. CTLA-4 competes with CD28 for binding to B7 on antigen-presenting cells (APCs), reducing T cell activation. Tumors also secrete immunosuppressive cytokines like IL-10, TGF-β, and VEGF, and recruit regulatory cells like Regulatory T Cells (Tregs) and myeloid-derived suppressor cells (MDSCs). All of these foster an immunosuppressive tumor microenvironment (TME). Additionally, tumors can alter metabolic pathways, increasing lactic acid levels, which impair T cell function and reduce MHC expression, thereby decreasing tumor immunogenicity [9].

ICB therapy aims to reinitiate a self-sustaining cancer immunity cycle that enhances immune responses while minimizing autoimmune side effects. Recent studies have demonstrated that ICB treatment can reverse the cancer-immunity cycle by addressing T-cell exhaustion [10]. The exhaustion of T cells is a condition where continuous tumor stimulation leads T cells to lose proliferation capacity and effector function while overexpressing inhibitory receptors like PD-1, CTLA-4, TIM-3, LAG-3, LAYN, and TIGIT. These exhausted T cells often become terminally differentiated within the tumor, disrupting the cancer-immunity cycle. Meanwhile, emerging research underscores the roles of precursor exhausted T cells (Tpex) in cancer immunotherapy, which sustain CD8⁺ T cell responses. Notably, the efficacy of ICB largely depends on the expansion of these Tpex cells. It leads to the influx of rejuvenated cytotoxic T cells within tumors rather than merely reactivating terminally exhausted T cells. This “clone revival” involves the induction of new Tpex clones from peripheral blood and lymph nodes that migrate into the TME, as well as the expansion of both new and existing Tpex cells within tumors [11]. Nonetheless, under persistent tumor-induced immunosuppression, T cells develop unique epigenetic landscapes with distinct DNA methylation and chromatin accessibility, differentiating into dysfunctional exhausted T cells (Tex). Although ICB therapy can reinvigorate exhausted CD8⁺ T cells, it does not alter the epigenetic landscape of dysfunctional Tex cells, thus potentially limiting the long-term efficacy of ICB treatments [12].

In conclusion, while ICB therapy can partially restore antitumor immune activity, it might not completely reverse tumor-induced immunosuppression. It could explain some patients’ nonresponsiveness. Thus, predicting the response to ICB requires an effective evaluation of both the positive and negative feedback loops in the patient’s cancer-immunity cycle.

3. AI-powered ICB response prediction with diverse data modalities

With the rapid advancements in high-throughput sequencing technologies and computational tools, researchers can now effectively monitor the cancer immune cycle in patients. Alongside these technological strides, a range of features has been identified for predicting responses to ICB therapies [13]. These features can be broadly classified into two categories: cancer-intrinsic and cancer-extrinsic features (Fig. 1).

Cancer-intrinsic features are typically identified through genomic alterations within cancer cells, whereas cancer-extrinsic features are often derived from transcriptomic analyses or other modalities within the TME [14]. Given the inherent complexity of the TME, relying on a single biomarker is insufficient for achieving robust prognostic and predictive insights. Instead, AI-based approaches have emerged as a promising solution, capable of integrating multiple features to provide a comprehensive understanding of the TME. Numerous AI models have been developed to predict ICB responses by utilizing diverse data modalities. Most of these models employ conventional supervised machine learning techniques such as LR, RF, SVM, and XGBoost (Fig. 2). In addition, some models leverage deep learning approaches, particularly for complex modalities like radiomics, digital pathology, and clinical data (Fig. 3). While previous literature has reviewed these AI models and the fundamental principles underlying AI technologies [6], we will not delve into those details here. Instead, this section emphasizes how researchers are utilizing AI models to characterize the biological processes of the cancer-immune cycle and predict patient responses to ICB therapies.

Fig. 2.

Conventional machine learning pipeline for predicting ICB response. Traditional machine learning models for predicting ICB responses typically utilize features derived from multi-omics datasets. The feature selection process primarily relies on expert-driven prior knowledge, where biomarkers associated with ICB responses are identified based on existing research. These biomarkers are then validated using query datasets to pinpoint prognostic features specific to the conditions under study. The validated features are subsequently input into machine learning algorithms to generate predictive outcomes. ICB, immune checkpoint blockade; TCR, T cell receptor; BCR, B cell receptor; LR, logistic regression; RF, random forest; SVM, support vector machines; XGBoost, extreme gradient boosting.

Fig. 3.

Deep learning pipeline for predicting ICB response. Current deep learning models, driven by advancements in vision-related technologies, leverage diverse information from real-world data, pathomics, radiomics, and genomics. Unlike traditional machine learning methods, deep learning networks automatically extract features from raw data and integrate them into a unified latent space. Fully connected layers within these architectures utilize the multi-omics latent representations to directly generate classification outcomes, streamlining and enhancing the predictive process.

3.1. Cancer-intrinsic features for ICB response prediction

With the rapid advancement of genomic sequencing technologies, genomic data from patients’ tumor biopsies have become increasingly accessible for clinical applications. This progress enables researchers to delve into the cancer-intrinsic factors that influence the response to immune checkpoint blockade.

3.1.1. Models based on tumor mutation profiles

Tumor antigens, including non-mutated self-antigens and neoantigens, enable immune recognition. While non-mutated self-antigens are overexpressed in tumors, their correlation with ICB response remains weak. In contrast, neoantigens derived from somatic mutations are key targets of tumor-specific immunity, supported by evidence from mouse models and humans. Tumor mutation burden (TMB), defined as non-synonymous mutations, reflects immunogenic potential and correlates with improved ICB response and survival in melanoma and NSCLC. However, TMB’s predictive power is modest, with a score of Area Under the Curve (AUC) around 0.6, limited by threshold variability and mutation heterogeneity [13].

Predicting response to immune ICB is critically dependent on the selection of relevant tumor mutations. Miao et al. demonstrated that gene mutations across distinct biological pathways exert differential effects on ICB response. By refining TMB through the selective inclusion of genes within specific pathways, predictive robustness was significantly improved, achieving AUC values of 0.74–0.82 in melanoma and NSCLC cohorts [15]. Building on this, Peng et al. combined random forest and convolutional neural networks to identify 55 somatic mutations predictive of durable clinical benefit in NSCLC, achieving an AUC > 0.9, outperforming TMB and PD-L1 [16]. Notably, the model identified responders with low TMB, highlighting its utility.

TMB alone often fails to predict ICB response due to its weak correlation with neoantigen load and tumor-infiltrating lymphocytes (TILs). Additional mutational signatures have been identified to assess tumor immunogenicity and predict ICB efficacy, such as deficient mismatch repair (dMMR) and high microsatellite instability (MSI) in colorectal cancer [17]. Consequently, dMMR/MSI-H serves as a robust biomarker for predicting metastatic colorectal cancer (mCRC) response to ICB therapy. Somatic copy number alterations (SCNAs) also influence ICB response, with low SCNA levels associated with better outcomes, likely due to impairing the genes related to neoantigen presentation. Integrating TMB with MSI and SCNA enhances predictive accuracy. Xie et al. employed deep learning to combine these features across tumor types, revealing that high MSI, high TMB, and low SCNA levels correlate with elevated immunogenicity of the tumor environment [18]. Conversely, tumors with high MSI, low TMB, and high SCNA levels were immunologically inactive and associated with poorer survival outcomes. This integrative approach highlights the value of combining multiple genomic features to refine ICB response prediction.

3.1.2. Models based on neoantigen presentation and recognition

While highly mutated tumors are more likely to generate neoantigens, not all mutations yield immunogenic peptides. T-cell recognition requires mutant peptides to be processed, transported, and subsequently presented on the cell surface by MHC-I molecules as peptide-MHC-I (pMHC-I) complexes [19]. Computational tools predicting neoantigen-MHC binding often perform comparably to TMB in predicting ICB response, with limited functional validation [20]. HLA genes, encoding MHC molecules, are critical for immune surveillance. Tumors may evade detection by downregulating HLA-I, acquiring HLA-I mutations, or losing β2-microglobulin. HLA-I heterozygosity enhances antigen presentation and correlates with improved survival, while HLA-I loss of heterozygosity (LOH) reduces ICB response. Abbott et al. developed an XGBoost-based model integrating neoantigen-MHC binding affinity and HLA-related alterations, achieving an AUC of 0.7 in predicting ICB response and progression-free survival in melanoma [21].

Beyond MHC binding, TCR recognition of pMHC complexes is crucial. Łuksza et al. introduced a neoantigen fitness model, quantifying tumor immunogenicity based on neoantigen-MHC affinity and TCR recognition likelihood, effectively predicting survival and response in melanoma and NSCLC [22]. Lu et al. advanced this with pMTnet, a deep-learning model predicting pMHC-TCR binding by integrating neoantigen-MHC features and TCR sequence data. They considered the influence of intratumoral heterogeneity (ITH) when evaluating tumor clone immunogenicity. Under immune pressure, tumor clones generate subclones, increasing ITH to improve fitness. Neoantigens derived from truncal clones typically induce a stronger cytotoxic effect than those from subclones. Lu et al. introduced a neoantigen immunogenicity effectiveness score (NIES), incorporating variant allele frequency (VAF) and TCR clonal fraction to account for ITH and truncal clone immunogenicity [23].

3.1.3. Models based on mixed genomics markers

Tumor immunogenicity, driven by neoantigen-TCR interactions, involves complex biomarker interplay. Researchers used ML tools to integrate these biomarkers nonlinearly and construct more robust predictive models for ICB response. Litchfield et al. integrated 11 biomarkers, including clonal TMB, mutational signatures, and immune gene expression, into an XGBoost model, achieving an AUC of 0.86 for pan-cancer ICB response prediction [24]. Similarly, Chowell et al. developed a random forest model (RF16) combining 16 genomic, clinical, and demographic features, such as TMB, HLA-LOH, and neutrophil-to-lymphocyte ratio (NLR), achieving an AUC of 0.85 [25]. These studies underscore the value of integrating diverse biomarkers using ML to enhance ICB response prediction.

3.2. Cancer-extrinsic features for ICB response prediction

While tumor immunogenicity is central to ICB response, the therapy’s efficacy depends on immune cell engagement within the TME. Analyzing immune cell diversity, functional states, and signaling pathways in the TME has emerged as a critical strategy for predicting ICB outcomes.

3.2.1. Models based on immune cell population composition

The TME’s immune cell composition reflects anti-tumor activity, with TILs serving as positive prognostic markers. Conversely, immunosuppressive cells like MDSCs correlate with poorer outcomes [26]. Computational tools such as CIBERSORT [27], TIMER [28], and xCell [29] deconvolve bulk RNA-seq data to estimate TME composition but often fail to resolve immune cell subsets critical for ICB response.

To address these challenges, Miao et al. developed the ImmuCellAI [30], a method that uses a reference gene expression matrix to estimate the abundance of 24 immune cell types, including 18 T-cell subsets. They further utilized an SVM model to predict ICB response based on the abundance of 22 immune cell types, achieving an AUC exceeding 0.8 in melanoma and gastric cancer cohorts. Similarly, Zaitsev et al. created Kassandra [31], a LightGBM-based algorithm trained on synthetic RNA profiles to identify 50 cell populations, including 16 T-cell subsets. Integrating Kassandra’s predictions with TMB and PD-L1 expression yielded an AUC of 0.75 in melanoma, highlighting the value of detailed immune cell profiling for ICB response prediction.

3.2.2. Models based on immune cell functional states

Beyond cell composition, immune cell functional states in the TME, particularly T cell activity, are critical for predicting ICB responses. Gene signatures capturing T cell priming, effector functions, TCR signaling, cytolytic activity, and chemokine expression have been widely used. For example, IFN-γ signaling, a key marker of effective T cell responses, enhances antigen presentation, upregulates MHC molecules, and recruits T cells via chemokines like CXCL9 and CXCL10. T cell-driven inflammation promotes tertiary lymphoid structure (TLS) formation, which correlates with positive outcomes [13].

Zhang et al. developed ICBnetIS based on prior T cell functional signatures. They selected a 75-gene signature derived from co-expression and protein-protein interaction networks, focusing on antigen presentation, B Cell Receptor (BCR), and TCR pathways. When integrated with a random survival forest model, ICBnetIS achieved an AUC of 0.784, outperforming other immunotherapy-related signatures [32]. Conversely, immune resistance signatures, such as the Resistance to Immunotherapy (RIR) and Tumor Immune Dysfunction and Exclusion (TIDE) scores, are designed to predict T cell dysfunction and exclusion, both of which are associated with poorer ICB responses [33,34]. Furthermore, Yang et al. introduced TCellSI, a comprehensive scoring system capable of distinguishing resting, activated, and suppressed T cell states. It has demonstrated strong predictive value for both patient prognosis and responsiveness to immunotherapy [35].

The TME’s other immune cells and stromal components also influence ICB outcomes. Bagaev et al. selected 29 gene functional signatures that represent key functional components and cellular populations within tumors. Based on these signatures, they identified four conserved TME subtypes across 20 cancers: immune-enriched fibrotic (IE/F), immune-enriched non-fibrotic (IE), fibrotic (F), and immune-depleted (D). Combining TME subtypes with TMB in a logistic regression model achieved an AUC of 0.82 for predicting ICB response in melanoma, highlighting the value of integrating TME analysis [36].

3.2.3. Models based on cell-cell communication profiles

Immune checkpoint genes (ICGs) encode ligands and receptors that regulate immune cell activity, serving as biomarkers for therapeutic targeting and T-cell response insights. Lee et al. developed a predictive model using cell-cell communication networks, where cell types are nodes and communication probabilities, inferred from ligand-receptor-cofactor gene expression, are edges. Integrated into a logistic regression model, this approach achieved an AUC of 0.79 across multiple cancers [37].

Noam et al. identified 15 ICGs predictive of immune response in neuroblastoma (NB) spontaneous regression, forming the immuno-predictive score (IMPRES). IMPRES outperformed other markers, such as cytolytic activity and PD-L1 expression, in predicting ICB responses in melanoma [38]. Similarly, Hu et al. screened 340 ICGs, selecting 80 strongly associated with immune activity markers like TMB and cytotoxic T cell (CTL) levels. The fitted logistic regression model achieved AUCs of 0.64–0.82 across cancers, demonstrating broad applicability [39].

The rapid advancement of spatial transcriptomics and proteomics technologies in recent years has led to the emergence of spatial biomarkers. These biomarkers are a powerful tool for predicting ICB response, as they capture the intricate cell-cell communication dynamics within TME. These biomarkers are typically derived from the metrics that quantify cell-cell distances and the co-localization of specific target expressions. For example, PD-L1⁺ macrophages at tumor margins and CD39⁺ memory T cells proximal to tumor cells have been identified as robust spatial biomarkers in melanoma, and clusters of CD8⁺ T cells adjacent to PD-L1⁺ cells are associated with favorable outcomes in NSCLC [40]. Leveraging regularized logistic regression models, Wang et al. identified predictive spatial biomarkers from imaging mass cytometry data collected from triple-negative breast cancer cohorts. These biomarkers fall into three categories: cell phenotype densities, cell-cell interactions, and proliferative fractions. The model demonstrated strong predictive performance and achieved an AUC of 0.82, underscoring the potential of spatial biomarkers in forecasting ICB response [41].

3.2.4. Models based on TCR repertoire analysis

The TCR repertoire, representing the complete set of T-cell receptors, provides critical insights into immune surveillance and tumor recognition. During ICB therapy, dynamic changes in the TCR repertoire occur through the infiltration of novel T-cell clones and the expansion of pre-existing ones, making TCR profiling a valuable tool for assessing anti-tumor responses. Substantial large-scale TCR sequence databases and analytical tools have been developed to support comprehensive TCR repertoire analysis [[42], [43], [44]].

TCR repertoire diversity, reflecting the richness and abundance of clonotypes, is predictive of ICB response. However, accurately measuring diversity remains challenging due to incomplete sequencing coverage and a lack of consensus on optimal metrics, such as the Gini-Simpson index and Shannon entropy. Bortone et al. addressed this by developing an elastic net regression model to predict Shannon entropy from partial TCR clonotypes, improving measurement accuracy [45]. Lozano et al. used Shannon entropy and the Gini-Simpson index to analyze TCR diversity in metastatic melanoma patients, linking higher peripheral blood TCR diversity to severe immune-related adverse events (irAEs). By combining TCR diversity with circulating CD4 memory T cell abundance, their logistic regression model achieved an AUC > 0.8 in predicting severe irAEs [46].

Sidhom et al. advanced this field with DeepTCR, a deep learning framework integrating TCR and HLA data [47,48]. Using a variational autoencoder (VAE) and multiple-instance learning (MIL) with multi-head attention, DeepTCR constructs joint TCR-HLA representations to predict ICB response. The model achieved AUCs of 0.86 in melanoma and 0.71 in basal and squamous cell carcinomas, demonstrating the power of deep learning in capturing systemic TCR repertoire patterns.

3.3. Advanced TME features for ICB response prediction

Rapid advancements in single-cell sequencing and multi-omics technologies have significantly enhanced our understanding of the anti-tumor response within the TME from both cancer-intrinsic and cancer-extrinsic perspectives. These technologies provide high-resolution insights into cellular heterogeneity and functional states within the TME, offering a nuanced understanding of the intricate interactions involved in ICB responses. By integrating data across modalities, this approach overcomes the limitations of single-modality analyses, enabling more comprehensive profiling of the true biological responses occurring within the TME [49].

3.3.1. Models based on single-cell RNA sequencing analysis

The advent of scRNA-seq has enabled the development of advanced computational tools for dissecting the TME, providing unprecedented cellular resolution. Tools such as scPred, CellAssign, and SCCAF [50] facilitate the automatic annotation of cell types within the TME. In addition, recent advancements have introduced deep learning-based models, including STCAT [51] and scAtlasVAE [52], which improve the accuracy of annotation at the T cell subtype level. Complementary tools such as SCENIC enable the identification of gene regulatory networks, Monocle reconstructs developmental trajectories, and CellChat, along with CellPhoneDB, elucidate cell-cell communication networks [50]. Collectively, these technologies empower researchers to investigate the functional states of specific cell populations in greater detail, providing critical insights into the mechanisms by which intercellular interactions contribute to or inhibit cancer progression.

By leveraging advanced single-cell analysis tools, researchers can predict ICB responses by integrating both cancer-intrinsic and cancer-extrinsic features with enhanced prior knowledge. For instance, Zhang et al. quantified cancer stem cell (CSC) stemness using pan-cancer scRNA-seq data, developing a stemness signature that predicted ICB response with an AUC of 0.71 via a Naïve Bayes model [53]. Exhausted T cells are crucial contributors to the response to ICB and have emerged as one of the most extensively studied cancer-extrinsic factors influencing ICB response. Zhang et al. identified T-cell exhaustion-associated genes using single-cell data and developed DeepAKR, an ensemble deep-learning framework. Their XGBoost model, based on a 16-gene signature, achieved an AUC of 0.778 in melanoma, outperforming established markers like IFN-γ and TIDE [54].

Beyond T cells, other cell types, including B cells, DCs, and Cancer-Associated Fibroblasts (CAFs), influence ICB outcomes. Sun et al. analyzed snRNA-seq data from melanoma patients pre- and post-anti-PD-1 therapy, identifying four cell subgroups strongly linked to treatment efficacy. Using an AdaBoost model, their predictive signature achieved an AUC of 0.95, highlighting the power of single-cell approaches in refining ICB response prediction [55].

3.3.2. Models based on multi-omics data

Multi-omics models integrate diverse data modalities to capture the complexity of tumor-immune interactions. Traditional approaches combine genomic and transcriptomic features, such as TMB, mutation signatures, HLA variations, and PD-L1 expression, using conventional classification and regression models. These models enhance predictive performance by synthesizing multi-modal data, offering valuable insights for clinical decision-making.

Beyond genomics and transcriptomics, clinical and imaging data offer additional avenues for refining predictions. Yoo et al. developed an ensemble machine learning system, SCORPIO, which leverages routine blood tests and clinical characteristics from large cohorts to achieve robust performance. SCORPIO demonstrated an AUC of 0.714 across 21 cancer types for predicting clinical benefit, highlighting the prognostic value of integrating clinical features [56]. Similarly, Sammut et al. developed an ensemble machine learning model that achieved an AUC of 0.87 in predicting ICB response in breast cancer. This model integrated clinical features with digital pathology, genomic, and transcriptomic data, underscoring the importance of multimodal data integration [57]. Additionally, growing research demonstrates that the gut microbiome plays a critical role in shaping patient responses to ICB treatment, largely through microbe‑derived metabolites and immune modulation. When metagenomics data are integrated with clinical features, they form a powerful foundation for predictive modeling. Gunjur et al. showed that combining strain-resolved microbial abundances with clinical features enhances machine learning predictions of ICB response [58]. Similarly, Björk et al. applied artificial intelligence to longitudinal metagenomic profiles of melanoma patients, identifying specific gut microbes consistently associated with survival outcomes. By constructing a predictive model based on the ratio of these microbes, they achieved moderate success in forecasting progression-free survival, illustrating how AI-driven analysis of the microbiome can improve ICB response prediction [59].

Deep learning advances have significantly enhanced multimodal integration, leading to more sophisticated approaches for predicting ICB responses. To tackle data heterogeneity, these models often incorporate advanced mechanisms like cross-attention and contrastive learning for robust feature representation. For example, Rami et al. introduced DyAM, a dynamic attention-based deep-learning model that fuses medical imaging, histopathology, and genomic data. DyAM employs dynamic weighting of features across modalities and achieved an impressive AUC of 0.8 in advanced NSCLC, outperforming simpler models [60]. Furthermore, Liu et al. proposed an end-to-end multimodal framework that combines self-supervised contrastive learning for histopathology feature extraction with radiomics and clinical data fusion. Using an attention-based integration approach, their model achieved high predictive accuracy in esophageal cancer, with an AUC of 0.809 [61]. Beyond static data, deep learning’s application to time-series data offers a significant advantage by autonomously extracting intricate patterns, thereby capturing temporal dynamics and treatment-induced alterations. An early example of this is the 3D RP-Net developed by Jin et al., a convolutional Siamese network to discern longitudinal changes from pre- and post-treatment multiparametric Magnetic Resonance Imaging (MRI) scans in rectal cancer patients. This model demonstrated high predictive accuracy for treatment response, achieving an AUC of 0.97 by fusing imaging-derived features with blood test results [62]. More recently, the transformer framework has emerged as a superior method for integrating diverse longitudinal multi-omics data. Yeghaian et al. exemplified this with MMTSimTA, a transformer-based architecture that integrates longitudinal Computed Tomography (CT) radiomics, blood biomarkers, and medication records. The model's innovative use of temporal attention modules led to excellent performance in pan-cancer survival prediction (AUC of 0.84 at 3 months) [63]. In addition to longitudinal imaging data, well-curated databases like ICBatlas [64] and scICB [65] provide essential longitudinal bulk and single-cell RNA-seq data, crucial for developing comprehensive time-series ICB response prediction models.

3.3.3. Model interpretability empowered by prior biological knowledge

Integrating prior biological knowledge into model development is essential for translating AI models from opaque “black boxes” into transparent and trustworthy tools for clinical decision-making. The incorporation of this knowledge not only influences a model’s predictive performance but, more importantly, fundamentally determines its interpretability. Prior knowledge used in models for predicting ICB response can be broadly classified into three types: (1) expert-curated biomarkers, such as PD-L1 expression or TMB; (2) systemic knowledge of biological pathways and networks, like protein-protein interaction (PPI) networks; and (3) hybrid knowledge-based biomarkers derived from the first two categories. Generally, traditional ML models preferentially leverage expert-curated and hybrid knowledge through feature engineering, whereas DL models are increasingly designed to embed systemic knowledge directly into their architecture.

The conventional ML pipeline relies on manual feature engineering. Researchers carefully curate and validate prognostic features from multi-omics data, drawing upon existing literature and biological databases. These interpretable, pre-selected features then serve as inputs for algorithms such as LR, RF, or XGBoost. For instance, both the established RF model by Chowell et al. [25] and the XGBoost model by Litchfield et al. [24] effectively utilize expert-curated biomarkers. More advanced models, like ICBnetIS [32], construct features based on interactions between curated biomarkers and PPI as well as gene co-expression networks, representing a more complex, hybrid approach.

In contrast, the DL pipeline can automate feature extraction directly from raw, high-dimensional multi-omics data. However, to address the “black box” problem often associated with DL, advanced models are being developed using two primary strategies: architecturally-integrated knowledge and knowledge-guided learning. These approaches embed biological principles directly into the model’s design. For example, COMPASS integrates prior knowledge by encoding tumor gene expression into 44 biologically-grounded “immune concepts”. Its novel concept bottleneck architecture forces the model to learn and reason through these human-interpretable concepts, offering mechanistic insights via personalized response maps [66]. Similarly, DANN integrates systemic knowledge by processing germline genomic data through the lens of a human PPI network. Its dual-channel attention mechanism explicitly learns the importance of specific genes and their network interactions, providing a clear rationale for its predictions [67].

In summary, the future of clinical AI likely lies in furthering these hybrid approaches. By combining the powerful pattern-recognition capabilities of DL with structured biological knowledge, we can create models that are not only highly accurate but also fundamentally understandable and clinically trustworthy.

4. Limitations and prospects in ICB response prediction

The landscape of ICB response prediction is being reshaped by advancements in multi-omics technologies and sophisticated computational methods. Recent strides in multi-omics, specifically scRNA-seq, scATAC-seq, and mass spectrometry-based proteomics, offer a promising avenue for understanding ICB efficacy. These techniques provide unparalleled, granular insights into the tumor microenvironment, immune cell dynamics, and epigenetic landscapes, all of which are critical determinants of ICB effectiveness. The substantial scale and inherent complexity of these multi-omics datasets frequently exceed the capabilities of conventional ML models, which typically rely on manually curated feature sets. DL is uniquely positioned to address this challenge due to its inherent capacity for automated hierarchical feature learning directly from raw, high-dimensional data. While end-to-end DL pipelines are commonplace in fields such as medical imaging, their application to multi-omics data for ICB prediction remains an emerging area. By learning intricate, non-linear interactions among genes, regulatory elements, and cell populations, deep neural networks possess the potential to uncover novel biomarkers and mechanistic signatures of ICB response that are often overlooked by traditional, hypothesis-driven methods [68].

The advent of large language models (LLMs), underpinned by transformer architectures, further expands the horizon for ICB response prediction. Models such as ChatGPT and other state-of-the-art LLMs demonstrate a remarkable capacity for sequence analysis, pattern discovery, and reasoning about emergent behaviors within extensive datasets [69]. Their ability to encode and process vast amounts of text-based or sequential biological data makes them highly suitable for analyzing genomic and transcriptomic profiles, as well as unstructured clinical notes and pathology reports. A particularly promising direction lies in the multimodal integration of diverse data types, including genomics, transcriptomics, proteomics, and even medical imaging data, within a unified deep learning or LLM-based framework. By capturing and correlating systemic patterns of immune activity, tumor heterogeneity, and patient-specific factors, these models could yield more accurate and generalizable ICB response predictions. Preliminary studies suggest that multimodal fusion approaches can robustly handle varying data modalities, identify patterns missed by single-mode analyses, and ultimately generate insights that may accelerate biomarker discovery and personalized immunotherapy [70].

Despite the promise of these emerging technologies, the development of robust and generalizable predictive models faces significant hurdles. A primary concern is data availability and quality. Training sophisticated DL and LLM models necessitates large-scale, high-quality, and diverse multi-omics clinical datasets, which remain challenging and expensive to acquire. Current immunotherapy studies often suffer from limited sample sizes, incomplete follow-up data, and inconsistencies in data collection protocols, increasing the risk of overfitting to cohort-specific noise rather than generalizable biological patterns. Furthermore, biological heterogeneity across different cancer types and patient populations introduces profound complexities, meaning biomarkers identified in one context often lack universal applicability. This issue is compounded by systemic biases in existing research, which has predominantly focused on immunologically “hot” tumors, leading to models that perform well for common indications but struggle with less-studied malignancies. Finally, technical variations stemming from diverse sequencing platforms, sample handling, or biopsy sites introduce batch effects and data quality issues, significantly degrading model robustness and predictive stability across external datasets [71].

Translating advanced predictive models into routine clinical practice introduces a distinct set of challenges beyond model development. A critical impediment is the “black-box” nature inherent in many complex deep learning models. For medical practitioners and regulatory bodies to trust and adopt AI-driven recommendations, the model’s decision-making processes must be transparent and explainable. This necessitates the development of inherently interpretable DL frameworks and rigorous validation of their outputs by clinical experts. Furthermore, the computational resources required for training and inference of large models are substantial, posing a practical barrier, especially for resource-constrained clinical environments. Efficient training paradigms, model compression techniques, and distributed computing strategies are essential for clinical feasibility [72]. Finally, the path to clinical acceptance requires a comprehensive credibility assessment, as outlined in established frameworks. The European ITFoC consortium specifies a validation process based on seven key steps, including ensuring data safety (quality, privacy, and security), defining performance metrics, and guaranteeing model explainability. This aligns with guidelines by de Hond et al., which stress the importance of post-deployment monitoring and auditing to manage risks and track performance over the model’s lifecycle. Adherence to such structured criteria—emphasizing data integrity, reproducibility, explainability and post-deployment surveillance—is non-negotiable for ensuring patient safety and building trust in AI-driven clinical decision support systems [73,74].

Additionally, the application of AI in ICB response prediction faces new and underexplored challenges that are critical for advancing personalized oncology. Most current models are designed for ICB monotherapy, despite the increasing clinical use of ICB in combination with chemotherapy or targeted therapy. Developing models capable of deconstructing complex treatment effects and accurately predicting response within these multi-drug contexts remains a significant challenge [66]. Secondly, while AI shows promise in synergistic drug discovery to enhance ICB efficacy [75], this area is relatively nascent. Leveraging sophisticated DL to predict novel synergistic interactions and guide the development of next-generation immunotherapy regimens requires substantial methodological advancements. Lastly, a crucial unmet need lies in the prediction of irAE. These complications of ICB treatment can be severe, yet robust tools to predict high-risk patients are largely absent. Developing clinically validated models to address irAE prediction is paramount for improving patient safety and optimizing treatment strategies [46].

The evolution of artificial intelligence, particularly with the advent of multi-omics, deep learning, and large language models, holds transformative potential for predicting ICB response and advancing personalized cancer immunotherapy. While significant progress has been made in capturing the intricate biological complexities of anti-tumor immunity, persistent challenges in data quality, generalizability, and model interpretability demand continued innovation in model development. Furthermore, successful clinical translation hinges on overcoming practical hurdles related to computational resources and strict regulatory compliance. The frontier of AI in oncology extends to predicting outcomes for combination therapies, discovering novel synergistic drug regimens, and foreseeing adverse events, each presenting unique application-specific challenges. Realizing this profound impact on precision oncology necessitates a concerted, multidisciplinary effort that unites computational scientists, cancer biologists, clinicians, and patient advocates under a shared commitment to ethical and transparent research.

Abbreviations

ICB, Immune Checkpoint Blockade; ML, Machine Learning; DL, Deep Learning; ICI, Immune Checkpoint Inhibitor; PD-1, Programmed Cell Death Protein 1; PD-L1/2, Programmed Cell Death Ligand 1 and 2; CTLA-4, Cytotoxic T Lymphocyte-Associated Antigen 4; FDA, Food and Drug Administration; NSCLC, Non-Small Cell Lung Cancer; RCC, Renal Cell Carcinoma; ORR, Overall Response Rate; HPD, Hyperprogressive Disease; LR, Logistic Regression; RF, Random Forest; SVM, Support Vector Machines; XGBoost, Extreme Gradient Boosting; CNN, Convolutional Neural Network; RNN, Recurrent Neural Network; NLP, Natural Language Processing; AI, Artificial intelligence; DC, Dendritic Cell; TCR, T Cell Receptor; MHC, Major Histocompatibility Complex; APC, Antigen-Presenting Cell; Treg, Regulatory T Cell; MDSC, Myeloid-Derived Suppressor Cell; TME, Tumor Microenvironment; Tpex, Precursor Exhausted T Cell; Tex, Exhausted T Cell; TMB, Tumor Mutation Burden; AUC, Area Under the Curve; TIL, Tumor-Infiltrating Lymphocyte; dMMR, DNA Deficient Mismatch Repair; MSI, Microsatellite Instability; mCRC, Metastatic Colorectal Cancer; SCNA, Somatic Copy Number Alteration; MHC-I, Major Histocompatibility Complex Class I; pMHC-I, Peptide-MHC-I Complex; HLA-I, Human Leukocyte Antigen Class I; LOH, Loss of Heterozygosity; ITH, Intratumoral Heterogeneity; NIES, Neoantigen Immunogenicity Effectiveness Score; VAF, Variant Allele Frequency; NLR, Neutrophil-to-Lymphocyte Ratio; TLS, Tertiary Lymphoid Structures; BCR, B Cell Receptor; RIR, Resistance-Immune Resistance; TIDE, Tumor Immune Dysfunction and Exclusion; ICG, Immune Checkpoint Gene; NB, Neuroblastoma; IMPRES, Immuno-predictive Score; CTL, Cytotoxic T Cell; irAE, Immune-Related Adverse Event; VAE, Variational Autoencoder; MIL, Multiple-Instance Learning; scRNA-seq, single-cell RNA sequencing; CSC, Cancer Stem Cell; CAF, Cancer-Associated Fibroblast; MRI, Magnetic Resonance Imaging; CT, Computed Tomography; PPI, Protein-Protein Interaction; LLM, Large Language Model.

CRediT authorship contribution statement

Shiwei Cao: Writing – review & editing, Writing – original draft, Visualization. Junwei Liu: Writing – review & editing, Writing – original draft, Supervision, Conceptualization. Yixue Li: Supervision, Conceptualization.

Declaration of competing interest

The authors declare that they have no conflicts of interest in this work.

Biographies

Shiwei Cao graduated from Huazhong Agricultural University with a bachelor’s degree in 2019. Currently, she’s a bioinformatics PhD candidate jointly trained by ShanghaiTech University and Guangzhou National Laboratory. Her research focuses on applying single-cell multi-omics and artificial intelligence technologies to cancer immunology and precision oncology.

Yixue Li (BRID: 09731.00.26087) received his Ph.D. from the University of Heidelberg, Germany, in 1996. His research focuses on bioinformatics, systems biology, genomics, precision medicine, and complex biological systems. Currently, he is leveraging deep learning methods to unravel complex biological processes. He has published over 300 papers and has an H-index of 78.