Inhibitory receptor profiles for joint prediction of benefit and toxicity in immune checkpoint blockade

Summary

Immune checkpoint blockade has transformed cancer therapy, but durable benefit is limited to a subset of patients and can be accompanied by immune-related adverse events. This Review evaluates inhibitory receptor profiles as candidate biomarkers for predicting both antitumor benefit and immune toxicity. We examine how receptor abundance, co-expression, cellular lineage, clonality, spatial context, and treatment-induced dynamics shape interpretation across tissue imaging, cytometry, bulk transcriptomics, single-cell multi-omics, and spatial profiling, with particular attention to organ-specific toxicity and time-dependent analytical bias. Current evidence supports context-dependent associations, especially for multi-receptor T-cell states and early on-treatment changes, but not universal thresholds or stand-alone clinical use. Because benefit and toxicity share pathways yet arise from partly distinct biological determinants, they should remain separate model outputs. We propose a staged path from mechanistic discovery to a locked minimal assay, external validation, and prospective evaluation of clinical utility.

Graphical abstract

Immunology; Immune response

Introduction

Immune checkpoint inhibitors (ICIs) targeting cytotoxic T-lymphocyte-associated protein 4 (CTLA-4), programmed cell death protein 1 (PD-1), or programmed death-ligand 1 (PD-L1) have reshaped the clinical course of malignancies.^1,2,3 Durable tumor control and survival are now achievable in subsets of patients with melanoma, non-small-cell lung cancer (NSCLC), renal cell carcinoma, and other cancers.^4,5 These gains, however, remain uneven. Many tumors show primary resistance, whereas others recur after an initial response through adaptive or acquired mechanisms. CTLA-4 and PD-1 are required for peripheral self-tolerance, and genetic disruption of either pathway can cause systemic or organ-specific autoimmunity.^6,7,8 Immune-related adverse events (irAEs) can affect virtually any organ system and range from dermatitis and thyroid dysfunction to colitis, pneumonitis, myocarditis, and neurological syndromes.^9,10 Combination regimens can improve response rates, but often at the cost of more frequent and severe toxicity.^11,12 The clinically relevant question is therefore not simply whether a patient is likely to respond, but whether a specific regimen offers a favorable balance between antitumor benefit and immune-mediated harm.

Biomarkers capture only parts of this balance. PD-L1 immunohistochemistry, mismatch-repair deficiency or microsatellite instability, and tumor mutational burden can enrich for benefit from ICI therapy in selected clinical settings.^13,14 Their predictive performance, however, varies across tumor types and is sensitive to assay platform, specimen heterogeneity, and the threshold used for classification.^15,16 These biomarkers index selected components of antitumor immunity but do not, individually, capture tumor antigenicity, immune-cell recruitment, effector state, and dependence on checkpoint-mediated inhibition.^17,18,19 These biomarkers were validated principally against efficacy endpoints and are not established for predicting the onset, severity, or organ distribution of irAEs.^13,14 Candidate toxicity biomarkers—including circulating immune-cell states, germline variants, and gut-microbiome features—have generally been evaluated in separate cohorts using non-uniform toxicity definitions and sampling schedules.^20,21 Although efficacy and toxicity biomarkers have usually been developed separately, checkpoint pathways regulate both tumor-directed immunity and peripheral self-tolerance.^6,22

Inhibitory receptors (IRs) offer a mechanistically plausible link between these outcomes. Inhibitory receptors are distributed across T-cell, NK-cell, and myeloid compartments, where their expression reflects activation history, differentiation state, and local microenvironmental signals rather than a uniform suppressive program.^23,24,25 Their expression may identify immune populations capable of reinvigoration while also reflecting the restraints required to maintain tissue tolerance.^26,27 This dual relevance makes IR profiles attractive candidates for joint prediction.

Interpretation nevertheless requires caution. High inhibitory-receptor expression is not synonymous with irreversible exhaustion, and receptor abundance alone does not define susceptibility to immune-mediated toxicity.^23,26 Checkpoint receptors are expressed by functionally distinct lymphocyte populations, including tumor-reactive effector cells, antigen-unrelated bystander T cells, and suppressive tumor-infiltrating Tregs.^28,29 The biological meaning of receptor expression depends on cellular identity and antigen specificity, as well as the anatomical compartment and the timing of sampling relative to treatment.^17,29 The available biomarker literature remains dominated by retrospective or exploratory cohorts and is heterogeneous with respect to assay platform, treatment regimen, sampling time, and endpoint definition.^30,31 Accordingly, IR features should be evaluated as components of externally validated multivariable models rather than as stand-alone clinical tests.^19,32

This Review treats efficacy and toxicity as related but distinct prediction tasks. We examine the biological information encoded by IR profiles, the technologies that preserve it, the strength of current associations with benefit and irAEs, and the evidence required for clinical implementation. We also consider whether toxicity can be mitigated without compromising antitumor activity and propose a multicenter strategy for reproducible clinical prediction.

Interpreting inhibitory receptor expression profiles

Receptor expression reflects immune state rather than a binary inhibitory signal

PD-1 and CTLA-4 are prototypical inhibitory receptors that constrain T-cell responses through distinct cellular and molecular mechanisms.^22,33 PD-1 is induced after T-cell activation, and engagement by PD-L1 or PD-L2 suppresses T-cell proliferation and cytokine production.^34,35 Sustained PD-1 expression is characteristic of chronic antigen exposure, but PD-1 is not specific for terminal dysfunction because activated, tumor-reactive, and proliferating CD8 T cells can also express the receptor.^26,28,36 CTLA-4 limits CD28 costimulation by binding the shared ligands CD80 and CD86 with high avidity and by removing these ligands from antigen-presenting cells through trans-endocytosis.^37,38 CTLA-4 is constitutively expressed by regulatory T cells and is strongly upregulated after activation of conventional T cells.^39,40 Accordingly, the frequency of receptor-positive cells is not biologically interpretable unless the expressing lineage, differentiation state, antigen specificity, and functional context are defined.^29,41

The same principle applies to newer checkpoints. LAG-3 is expressed by activated conventional and regulatory T cells and recognizes stable peptide-MHC class II complexes; FGL1 has also been proposed as an MHC-II-independent ligand, although its functional contribution appears context-dependent.^42,43,44 TIM-3 engages galectin-9 and CEACAM1 and is enriched in chronically stimulated or dysfunctional T-cell populations.^45,46 TIGIT binds PVR-family ligands, with particularly high affinity for CD155, and restrains CD226 costimulation through ligand competition and interference with CD226 signaling.^47,48 VISTA, also known as PD-1H or VSIR, is expressed across hematopoietic cells but is particularly abundant in monocytes, macrophages, and other myeloid populations.^49,50 CD94–NKG2A recognizes HLA-E, whereas inhibitory KIRs detect specific classical HLA class I allotypes; engagement of self-MHC class I by these receptors calibrates NK-cell responsiveness during education.^51,52

Inhibitory-receptor co-expression improves state resolution and often tracks cumulative antigenic stimulation, but it does not by itself establish terminal or irreversible exhaustion.^41,53 Progenitor-like TCF1-positive PD-1-positive CD8 T cells retain self-renewal and proliferative capacity and supply the proliferative burst elicited by PD-1 pathway blockade.^26,54 More differentiated exhausted subsets generally display broader inhibitory-receptor co-expression and reduced proliferative competence, although intermediate and terminal subsets may retain cytotoxic effector programs.^53,55,56 TOX is a central transcriptional and epigenetic regulator of exhausted T-cell differentiation, underscoring that inhibitory-receptor abundance is only one component of a broader adaptation to persistent antigenic stimulation.^57,58,59 Accordingly, receptor measurements should be integrated with differentiation state, proliferative capacity, clonal architecture and, where feasible, transcriptional or functional readouts.^55,60

Cellular lineage, antigen specificity, and spatial context determine biological meaning

The tumor microenvironment contains tumor-reactive lymphocytes, bystander cells, regulatory T cells, natural killer cells, macrophages, dendritic cells, and other immune populations with distinct functions.^61,62 PD-1-positive CD8 tumor-infiltrating lymphocytes can be enriched for clonally expanded tumor-reactive cells, but tumor infiltrates also contain abundant tumor-unrelated bystander CD8 T cells, making receptor expression alone insufficient to establish antigen specificity.^28,29 On regulatory T cells, CTLA-4, TIGIT, and LAG-3 can mark or directly support suppressive programs and should not be interpreted as evidence of effector-cell exhaustion.^63,64 NKG2A contributes to NK-cell education, whereas both NKG2A and TIGIT can restrain antitumor NK-cell effector function.^24,65 VISTA-positive macrophages and PD-L1-expressing dendritic or other host myeloid cells can establish local suppressive circuits that attenuate T-cell activation and cytotoxic function.^25,66,67 Bulk-tissue measurements average signals across malignant, stromal, and immune lineages and therefore cannot assign inhibitory-receptor expression to a specific cellular source or functional state.^61,62

Anatomical location adds another layer of meaning. The functional significance of receptor-positive lymphocytes depends on their location within tumor nests, invasive margins, stromal regions, or organized lymphoid aggregates and on their proximity to tumor cells and antigen-presenting cells.^17,68,69 B-cell-rich tertiary lymphoid structures have been associated with favorable responses to checkpoint blockade, illustrating that productive immune architecture can coexist with abundant checkpoint expression.^70,71 Tissue-resolved profiling can distinguish tumor-localized immune responses from organ-specific inflammatory programs in injured tissues, as demonstrated in checkpoint colitis and myocarditis.^72,73 Figure 1 summarizes the principal cellular compartments in tumor tissue and peripheral blood.

Figure 1.

Cellular, spatial, and temporal determinants of inhibitory receptor profiles

(A) Inhibitory receptor expression has distinct biological implications across immune-cell lineages within the tumor microenvironment. Tumor-reactive CD8⁺ T cells may co-express PD-1, LAG-3, TIM-3, and TIGIT, whereas phenotypically similar bystander CD8⁺ T cells may also express PD-1 without contributing directly to tumor recognition. CTLA-4, LAG-3, and TIGIT expression on regulatory T cells may reflect suppressive activity and maintenance of tissue tolerance; NKG2A and TIGIT regulate the activation threshold of natural killer cells; and VISTA and PD-L1 expressed by myeloid cells may contribute to a locally suppressive niche.

(B) Anatomical localization further modifies the interpretation of receptor-positive populations. Cells situated within tumor nests, stroma, tertiary lymphoid structures, or vascular interfaces may differ in antigen exposure, cellular interactions, and functional relevance. Tumor tissue provides direct spatial and cellular context, whereas serial peripheral-blood sampling captures systemic and treatment-induced changes; the two specimen types are therefore complementary rather than interchangeable.

(C) An interpretable inhibitory receptor profile integrates receptor abundance, cell lineage, co-expression, clonality, spatial location, and sampling time. Because each dimension influences biological meaning, receptor positivity cannot be reduced to a universal cutoff applicable across cell populations, anatomical compartments, treatment regimens, or time points. IR, inhibitory receptor; NK, natural killer.

Tumor biopsies provide direct access to the tumor-associated immune compartment, but a single biopsy samples only a limited region of a spatially heterogeneous lesion.^74,75,76 Peripheral blood permits serial assessment of systemic immune dynamics and can capture tumor-associated clonotypes, but circulating immune states overlap only partly with those in tumor tissue; the two compartments are therefore complementary rather than interchangeable.^77,78

These data do not support a universal receptor-positivity cutoff. Biomarker definitions should prespecify the expressing lineage and analytical denominator, specimen and anatomical compartment, and sampling time relative to therapy; a phenotype measured among all CD8 T cells is not equivalent to the same phenotype measured among activated or clonally expanded CD8 T cells.

Benefit and toxicity share pathways but may arise from distinct cellular sources

PD-1 pathway blockade can reinvigorate and expand tumor-reactive CD8 T cells, while the post-treatment repertoire may also be reshaped through the recruitment of clonotypes not detected in pretreatment tumor samples.^17,22,60 CTLA-4 blockade can alter the intratumoral effector-to-regulatory T-cell balance; however, Fc-dependent regulatory T-cell depletion is context-dependent and has not been uniformly observed in human tumors.^79,80 CTLA-4 and PD-1 are required for peripheral self-tolerance, and checkpoint release can permit immune attack on normal tissues, including the expansion of T-cell clonotypes shared between tumors and injured organs.^8,10 Heterozygous loss-of-function variants in CTLA4 cause a variably penetrant multisystem immune dysregulation syndrome, illustrating the consequences of chronically impaired checkpoint restraint.^6,7 In advanced melanoma, combined PD-1 and CTLA-4 blockade causes substantially more grade 3–4 treatment-related toxicity than either monotherapy.^12,81 Together, these findings suggest that broader checkpoint release can increase both antitumor activity and toxicity, whereas the clinical phenotype of toxicity is shaped by organ-specific immune programs.^12,82

The relationship between antitumor benefit and immune toxicity is nevertheless not deterministic. Antitumor activity is concentrated within restricted tumor-reactive T-cell clonotypes rather than being uniformly distributed across receptor-positive lymphocytes.^17,60 By contrast, immune-related adverse events (irAEs) can involve cross-reactive or tissue-resident T cells, early B-cell perturbations, and organ-specific inflammatory programs.^10,20 At the cohort level, irAE occurrence can be associated with treatment outcome, but the direction and magnitude of this association vary according to toxicity grade, organ system, treatment regimen, and analytical approach.^83,84 Associations between irAEs and survival should therefore be estimated using landmark or time-dependent methods because irAEs arise after treatment initiation and naive exposure classification is vulnerable to guarantee-time or immortal-time bias.^83,85

A single global measure of immune activation is unlikely to represent antitumor benefit and immune toxicity adequately. Because efficacy and toxicity are associated with partly distinct biological features, a joint prediction framework should estimate them as separate, calibrated outcomes while allowing shared predictors and interactions where supported by the data.^19,21 Checkpoint colitis, pneumonitis, and myocarditis arise in distinct tissue compartments, involve different immune effector programs, and carry markedly different clinical hazards.^10,86 Accordingly, any-grade irAEs, severe irAEs, and organ-specific irAEs should be treated as distinct prediction endpoints; pooling these events may increase event counts but sacrifices mechanistic and clinical specificity.^21,87

Measuring inhibitory receptor profiles

Tissue protein assays preserve context but remain sensitive to assay design and sampling

Immunohistochemistry (IHC) and multiplex immunofluorescence (mIF) can be applied to formalin-fixed, paraffin-embedded tissue sections while retaining tissue morphology and spatial relationships. Multiplex panels can assign checkpoint expression to malignant, lymphoid, and myeloid compartments and quantify selected co-expression states at single-cell resolution.^88,89,90 Because these assays use routinely archived FFPE material and retain cell-assigned spatial information, they are well suited to pathology-based translational studies; clinical deployment, however, requires reproducible staining, imaging, and analysis across laboratories.^90,91

PD-L1 IHC provides a clear example of why analytical standardization must precede clinical interpretation: antibody clone, staining platform, scoring algorithm, cellular denominator, and positivity threshold can alter assay concordance and the resulting classification of individual specimens.^16,92,93 Even when nominally equivalent cutoffs are used, assay performance depends on the antibody-platform combination and on whether expression is scored in tumor cells, immune cells, or both.^16,94 Specimen age, tumor-cell content, and reader scoring introduce additional variability, particularly for samples close to clinically relevant cutoffs.^16,95 Sampling remains a major source of uncertainty: small biopsies can underestimate spatially heterogeneous PD-L1 expression, and clinically relevant discordance can occur between primary tumors and matched metastatic lesions.^96,97 Biological relevance does not establish analytical interchangeability. Multi-marker inhibitory-receptor panels should therefore enter validation only after antibody clones, staining conditions, image-segmentation and phenotyping algorithms, quality-control criteria, and scoring rules have been locked.^88,90,91

Transcript-based measurements face parallel limitations. Bulk RNA measurements represent composite signals from malignant, stromal, and immune populations and are therefore strongly influenced by tumor purity and cell-type composition.^98,99 Sequencing- and imaging-based spatial transcriptomic platforms differ in molecular sensitivity, effective spatial resolution, cell-segmentation error, and downstream analytical behavior; measurements generated by different platforms should therefore not be assumed to be directly equivalent.^100,101 Cross-platform concordance should be demonstrated empirically using matched specimens, predefined metrics, and common reference standards. Transcript-derived signals intended to represent receptor abundance or pathway activity should, where feasible, be confirmed by an orthogonal protein assay because mRNA abundance does not reliably predict protein abundance or phosphorylation state across tumors.^102,103,104

Peripheral blood supports longitudinal profiling but requires rigorous pre-analytical control

Multiparameter flow cytometry and mass cytometry can quantify lineage markers, inhibitory receptors, activation and proliferation markers, and intracellular effector proteins at single-cell resolution in peripheral blood mononuclear cells and dissociated tumor specimens.^105,106,107 Peripheral blood-based cytometry is particularly amenable to prespecified serial sampling, allowing baseline immune states to be compared with early and later on-treatment profiles.^108,109 Baseline profiles can be compared with prespecified early pharmacodynamic changes, including Ki-67 induction within PD-1-positive CD8 T cells and expansion of treatment-responsive phenotypic or clonal populations. In several cohorts, these early changes provided information not captured by pretreatment frequencies alone.^109,110,111

Longitudinal profiling nevertheless requires rigorous analytical control. Pre-analytical handling and cryopreservation, instrument configuration, fluorescence compensation, and analysis strategy can alter measured subset frequencies and marker intensities.^112,113,114 Low-abundance populations and subsets defined by dim or weakly separated markers show greater analytical variability and require adequate event counts, explicit gating definitions, and reproducibility assessment.^113,115 Biological interpretation is also time-dependent. Immune changes detected within the first one to four weeks of PD-1 pathway blockade may reflect an acute pharmacodynamic response, whereas profiles obtained after multiple cycles integrate cumulative treatment effects, disease evolution, and intervening clinical exposures.^110,111,116 Sampling windows and analytical denominators should therefore be prespecified. Relative frequencies should be interpreted alongside absolute counts or another explicit reference because cell-subset proportions are compositional: depletion of one population can produce an apparent relative increase in others without absolute expansion.^117,118

Clinical translation should reduce a high-dimensional discovery signature to a prespecified, lower-complexity marker panel that preserves the relevant biological state and meets predefined inter-laboratory precision criteria. Assay reduction should follow biological definition and independent confirmation of the target state rather than precede them.^108,112,113

Single-cell and spatial technologies are discovery tools rather than routine diagnostics

Single-cell RNA sequencing, coupled with T cell receptor sequencing or single-cell chromatin-accessibility profiling, resolves transcriptionally and epigenetically distinct activated, progenitor-exhausted, and more terminally differentiated lymphocyte states.^23,60,119 Integrated transcriptome-TCR and chromatin profiling can relate inhibitory-receptor expression to clonal expansion, differentiation state, and regulatory circuitry; antigen specificity should be described as inferred unless cognate TCR reactivity is experimentally demonstrated.^60,119 Spatial transcriptomic and highly multiplexed protein-imaging methods localize receptor-defined populations to tumor-immune interfaces, fibrovascular or stromal niches, and organized multicellular immune hubs. Their distinctive contribution is the recovery of cellular neighborhoods and interface-specific programs that are lost after tissue dissociation or bulk averaging.^68,120,121

These technologies also introduce substantial analytical uncertainty. Limited molecular capture can yield false-negative transcript measurements, and absence of a receptor transcript in an individual cell should not be equated with absence of the corresponding surface protein.^122,123,124 Tissue dissociation can alter recovered cell-type proportions and induce stress-response programs; protocol-dependent loss or under-recovery of susceptible populations should therefore be treated as a source of compositional and transcriptional bias.^125,126 The patient, rather than the individual cell, is the unit of clinical inference. Cells obtained from the same patient are correlated subsamples, and analyses that ignore this hierarchy can markedly inflate type I error and apparent precision.^127,128 Batch effects and analysis-pipeline choices can create non-biological structure or erase genuine biological variation; integration, feature selection, and downstream modeling should therefore be prespecified and confined to the development data.^129,130 Sequencing-based spatial transcriptomic spots may contain mixtures of multiple cell types, whereas imaging-based methods remain sensitive to cell-boundary assignment; uncertainty from deconvolution and segmentation should therefore be propagated into downstream analyses.^131,132

High-dimensional platforms are best positioned for discovery in clinically annotated cohorts with patient-level linkage and, where feasible, paired tumor-blood or serial specimens.^23,77 Candidate features should subsequently undergo orthogonal confirmation, reduction to a prespecified lower-complexity assay, and evaluation in independent or multicenter cohorts before they are proposed for clinical use.^90,108 Table 1 compares the principal platforms. The key distinction is between an exploratory measurement that illuminates mechanism and a locked assay capable of supporting a clinical decision.

Table 1.

Analytical platforms for inhibitory receptor profiling and their translational roles

Platform	Principal output and specimen	Main advantages	Main limitations	Recommended translational role	Representative primary studies
IHC and mIF	spatially resolved protein expression in FFPE or fresh tissue; assignment of IRs to tumor, lymphoid, and myeloid compartments	preserves morphology and cellular neighborhoods; compatible with routine pathology workflows; permits direct assessment of selected IR co-expression in situ	sensitive to antibody clone, fixation, staining platform, scoring method, segmentation, and cell denominator; limited multiplexing and dynamic range; small biopsies may not capture intratumoral or interlesional heterogeneity	tissue-based validation of spatially defined IR states; development of compact panels for trial stratification and, after analytical validation, clinical testing	Anagnostou et al.133; Hirsch et al.92; Sanchez et al.134
Flow cytometry and CyTOF	multiparametric single-cell protein measurements in peripheral blood or dissociated tumor tissue	quantitative assessment of lineage, activation, proliferation, and IR co-expression; resolves rare immune subsets; well suited to serial sampling and pharmacodynamic analysis	requires viable, well-processed cells; sensitive to collection delay, cryopreservation, antibody lot, instrument settings, compensation, batch effects, and gating strategy; lacks spatial information	longitudinal monitoring of treatment-induced immune changes; validation and reduction of blood-based signatures to standardized minimal panels	Gadalla et al.135 ; Kamphorst et al.110; Simon et al.136
Bulk RNA sequencing and targeted gene-expression profiling	aggregate transcript abundance in tumor tissue, whole blood, or purified cell populations	broad pathway coverage; scalable across relatively large cohorts; supports predefined IR, inflammation, and T-cell dysfunction scores; compatible with integrated genomic analyses	confounded by tumor purity and cellular composition; limited resolution of rare populations; transcript abundance may not reflect surface protein or signaling activity; susceptible to batch and normalization effects	screening and validation of composite efficacy or toxicity signatures; integration with protein measurements and clinical variables	Shahabi et al.137 ; Jiang et al.138; Zhang et al.139
scRNA-seq and single-cell multi-omics	cell-resolved transcriptomes, optionally integrated with TCR-seq, surface-protein measurements, or ATAC-seq, from blood or dissociated tissue	resolves heterogeneous immune states; links IR expression to differentiation, clonality, regulatory programs, and inferred antigen specificity; supports discovery across cellular lineages	transcript dropout, dissociation bias, mRNA-protein discordance, batch effects, high cost, and limited patient numbers; cells are nested within patients and cannot be treated as independent clinical observations	mechanistic discovery and nomination of candidate cellular states; derivation of features for subsequent orthogonal validation and assay reduction	Sade-Feldman et al.23; Komuro et al.140; Tooley et al.141
Spatial transcriptomics and spatial proteomics	spatially indexed RNA or protein measurements in FFPE or frozen tissue sections	links IR-defined states to tumor borders, stromal and vascular niches, immune aggregates, and cell-cell interactions; preserves tissue architecture while extending molecular depth	variable capture efficiency, resolution, sensitivity, and segmentation accuracy; spot mixing and platform-specific normalization complicate comparison; costly, low throughput, and usually restricted to one sampling time	discovery and validation of spatial biomarkers; identification of tissue features that can be translated into lower-complexity imaging panels	Sanchez et al.134

ATAC-seq, assay for transposase-accessible chromatin using sequencing; CyTOF, cytometry by time of flight; FFPE, formalin-fixed, paraffin-embedded; IHC, immunohistochemistry; IR, inhibitory receptor; mIF, multiplex immunofluorescence; scRNA-seq, single-cell RNA sequencing; TCR-seq, T cell receptor sequencing. References indicate representative primary research articles included in the manuscript bibliography; reviews, consensus statements, guidelines, editorials, and methods chapters were excluded.

Inhibitory receptor profiles and antitumor benefit

Single-receptor measurements provide partial and context-dependent information

A straightforward biomarker model posits that expression of a therapeutic target should enrich for benefit from its blockade. PD-L1 immunohistochemistry can enrich for response to PD-1 or PD-L1 blockade in several malignancies, but it is not a binary classifier: objective responses occur in PD-L1-negative disease, whereas many PD-L1-positive tumors do not respond.^142,143,144 PD-1 expression on CD8 tumor-infiltrating lymphocytes can enrich for clonally expanded tumor-reactive cells, but PD-1 alone does not establish antigen specificity because tumor-unrelated bystander populations can share overlapping activation and differentiation phenotypes.^28,29 The biological meaning of a high PD-1-positive frequency therefore depends on the expressing state and compartment: it may reflect enrichment of tumor-reactive clones, a dysfunctional PD-1-high population, or an admixture of tumor-specific and bystander cells.^28,29

Early on-treatment changes can provide more direct pharmacodynamic evidence than pretreatment frequencies. In non-small-cell lung cancer and melanoma, a transient expansion of circulating PD-1-positive Ki-67-positive CD8 T cells during the first treatment cycles has been associated with clinical benefit; in melanoma, response was more closely associated with the magnitude of T-cell reinvigoration relative to pretreatment tumor burden than with the unscaled immune change alone.^109,110,111 These early immune changes remain candidate pharmacodynamic biomarkers and require prespecified sampling, locked analytical procedures, and validation in independent cohorts before clinical use.^90,108

Among newer checkpoints, clinical activity has been established for LAG-3 blockade in advanced melanoma, whereas TIGIT and TIM-3 studies have yielded regimen-specific or preliminary signals rather than validated expression-based companion diagnostics.^145,146,147 For NKG2A, early clinical studies have demonstrated pharmacodynamic activation of NK- and T-cell compartments with variable, tumor-dependent activity; evidence for VISTA remains predominantly mechanistic and tissue-based.^24,25 No expression threshold derived from these studies has yet demonstrated reproducible predictive performance across tumor types and treatment regimens. PD-1 and LAG-3 can mediate cooperative, non-redundant inhibition of chronically stimulated T cells, and dual blockade can improve antitumor activity; however, LAG-3 abundance alone has not been validated as a companion diagnostic for relatlimab-nivolumab.^53,148 Target abundance should therefore be interpreted as one component of a lineage- and state-defined immune phenotype rather than as an independent companion diagnostic.^23,41

Co-expression, differentiation state, and clonality improve biological resolution

High-dimensional studies indicate that inhibitory-receptor co-expression resolves functionally distinct CD8 T-cell states more effectively than any single receptor measurement.^41,53,55 Broad inhibitory-receptor co-expression is enriched in chronically stimulated and clonally expanded CD8 T-cell populations, including tumor-reactive cells, but these populations remain heterogeneous in antigen specificity, differentiation state, and functional competence.^53,60,149 Progenitor-like TCF1-positive PD-1-positive CD8 T cells retain self-renewal and proliferative capacity and generate differentiated progeny during PD-1 pathway blockade.^26,41 More differentiated exhausted CD8 T-cell subsets generally display broader inhibitory-receptor co-expression, reduced proliferative potential, and limited long-term persistence, although some retain immediate cytotoxic effector programs.^41,55,56 Figure 2 illustrates these heterogeneous, context-dependent states. Accordingly, the balance between progenitor-like and more differentiated exhausted states may provide greater biological resolution than the aggregate frequency of receptor-positive CD8 T cells, although its incremental predictive value requires prospective validation.^23,27,41

Figure 2.

Inhibitory receptor co-expression defines heterogeneous CD8⁺ T-cell states and clonal responses to checkpoint blockade

(A) Schematic continuum of CD8⁺ T-cell states associated with persistent antigen exposure. Activated effector cells retain strong effector function and proliferative capacity despite PD-1 expression. Progenitor-exhausted TCF1+PD-1+ cells maintain self-renewal and substantial proliferative reserve and can respond to PD-1 blockade. Intermediate exhausted cells co-express PD-1, TIGIT, and LAG-3 and exhibit mixed effector activity with partial proliferative competence. Terminally differentiated cells are characterized by high PD-1 expression, broader co-expression of TIM-3, LAG-3, and TIGIT, and high TOX expression; these cells generally show limited persistence, although immediate cytotoxic activity may be retained. Proliferative reserve tends to decrease as inhibitory receptor co-expression increases, but the illustrated states are heterogeneous, context-dependent, and not necessarily components of a fixed unidirectional trajectory.

(B) Clonal information complements receptor abundance and differentiation state. PD-1 blockade may expand pre-existing progenitor-like clonotypes in the tumor and peripheral blood, whereas tumor-blood shared clonotypes provide evidence linking circulating populations to the intratumoral response. Serial sampling may also reveal clonal replacement, in which newly recruited clonotypes emerge as pre-existing populations contract, including during PD-1-based combination therapy. Colors denote representative clonotypes and are not intended to imply universal lineage relationships. IR, inhibitory receptor; TCF1, T-cell factor 1; TOX, thymocyte selection-associated high-mobility-group-box protein.

Clonal information improves the interpretation of receptor phenotypes, but clonal expansion alone does not establish tumor specificity. Stronger evidence is obtained when expanded clonotypes are linked to tumor enrichment, shared across tumor and blood, or validated for cognate tumor-antigen recognition.^60,77,149 Longitudinal tracking adds evidence of treatment responsiveness: expansion or phenotypic transition after checkpoint blockade identifies therapy-responsive clonotypes, whereas a static receptor-high phenotype is insufficient to distinguish persistent dysfunction from stable immune surveillance.^60,109,150 Paired tumor profiling after PD-1 blockade has also revealed clonal replacement, in which expanded post-treatment CD8 T-cell clonotypes were not detected in matched pretreatment samples, consistent with recruitment or local expansion of newly sampled tumor-reactive populations rather than exclusive reinvigoration of pre-existing intratumoral clones.^60,77

T-cell states should not be interpreted in isolation, because B-cell-rich tertiary lymphoid structures define an organized immune contexture associated with checkpoint-blockade response and survival in several tumor settings.^70,151 NK-cell and myeloid checkpoint pathways further shape the local consequences of T-cell activation by regulating cytotoxicity, antigen presentation, and suppressive myeloid circuits.^24,67 A biologically credible efficacy signature should integrate the abundance of a candidate population with its differentiation or functional state and evidence that it participates in the tumor-directed response, such as tumor enrichment, clonal expansion, treatment-associated dynamics, or experimentally confirmed antigen recognition.^23,77,149 Receptor co-expression improves phenotypic resolution but does not, by itself, establish cognate antigen specificity, functional competence, or dependence of the clinical response on the targeted pathway.^29,41,149

Predictive value depends on regimen, tumor context, and endpoint

Baseline inhibitory-receptor profiles should be interpreted within a defined regimen and disease setting. In untreated advanced melanoma, nivolumab-relatlimab improved progression-free survival over nivolumab, whereas the same combination did not improve recurrence-free survival in the adjuvant setting; adjuvant nivolumab-ipilimumab likewise failed to improve recurrence-free survival over nivolumab and caused more grade 3–4 toxicity.^145,152,153 The efficacy of nivolumab-relatlimab in untreated advanced melanoma establishes the therapeutic relevance of LAG-3, but neither the 1% LAG-3 cutoff nor another expression threshold has been validated for selecting patients for the combination.^145,154

Biomarker associations vary across tumor types and may be altered by prior immunotherapy and by the anatomical site from which the specimen is obtained.^19,155 Biomarker development should therefore begin within a prespecified tumor-regimen setting and use an endpoint aligned with the intended clinical decision. Pathological response, radiographic response, progression- or recurrence-free survival, and overall survival should be analyzed as distinct outcomes rather than assumed to be interchangeable.^111,145,153 Cutoffs should not be transferred across tumor types, treatment settings, or anatomical compartments without explicit analytical and clinical bridging.

Several influential inhibitory-receptor signatures have been derived from single-arm or non-randomized treatment cohorts and therefore establish associations within a treated population rather than differential benefit between competing regimens.^23,108 To establish treatment-predictive rather than solely prognostic value, a biomarker must identify heterogeneity in the relative treatment effect, ideally through a prespecified biomarker-by-treatment interaction in a randomized comparison. An association with outcome in a single treatment arm cannot determine whether another regimen would provide greater or lesser benefit.^156,157,158 This distinction is essential when a biomarker is proposed to guide monotherapy versus combination treatment.

Inhibitory-receptor profiles should be evaluated for incremental value over established and partly non-redundant biomarker dimensions, including tumor mutational burden and T-cell-inflamed gene expression, organized B-cell and tertiary-lymphoid architecture, and host microbiome features.^19,70,159 The purpose of adding inhibitory-receptor features is not to displace established biomarkers, but to determine whether they provide reproducible out-of-sample information and improve a predefined treatment decision beyond transparent clinical and molecular reference model.

Inhibitory receptor profiles and immune-related adverse events

Baseline profiles may reflect immune restraint, but direct evidence remains limited

Pretreatment toxicity studies have evaluated circulating immune-cell states, cytokine profiles, and host germline variation, generally in separate cohorts and against non-uniform toxicity endpoints.^21,160,161 Organ-focused studies have additionally examined thyroid autoantibodies and HLA alleles, whereas baseline gut-microbiome composition has been investigated primarily in relation to gastrointestinal toxicity.¹⁶² In ICI-associated myocarditis, routine hematologic signals remain exploratory: a treatment-associated decline in absolute lymphocyte count and rise in the neutrophil-to-lymphocyte ratio have been observed before or at clinical presentation, but baseline values have not been validated as pretreatment predictors.^163,164,165 Baseline TgAb or TPOAb positivity has been associated with an increased risk of thyroid irAEs, whereas reported HLA associations remain organ- and population-specific and require independent replication.^166,167 Together, these findings suggest that pretreatment autoimmunity and host immune set points can modify irAE risk, although no single feature defines a general susceptibility state across organs and treatment regimens.^21,160 In the principal pretreatment studies with independent validation, toxicity-associated signals arose from activated CD4 memory-cell abundance, TCR diversity, cytokine profiles, or germline variation rather than from a reproducible single-receptor expression cutoff.^21,160

A simple model might predict that weak baseline checkpoint restraint increases toxicity risk whereas high receptor expression is protective; however, receptor abundance does not map monotonically onto immune restraint. High expression can mark activated, tumor-reactive, or progenitor-like lymphocytes with residual proliferative capacity.^26,28,168 Conversely, low aggregate receptor expression may reflect lineage composition or a low abundance of antigen-experienced cells rather than weak checkpoint regulation. Regulatory T cells add further complexity because CTLA-4, TIGIT, and LAG-3 can mark or support suppressive programs that preserve self-tolerance while limiting antitumor immunity.^42,63,64 Baseline receptor profiles should therefore be interpreted alongside cellular lineage and activation state and, for organ-specific outcomes, relevant autoantibodies or other tissue-linked susceptibility markers.^160,167 Clinical models should additionally account for pre-existing autoimmune disease, baseline co-medications, and the intended checkpoint regimen, because each can alter the observed toxicity phenotype.^11,169

Available pretreatment receptor profiles should therefore be described as exploratory correlates of immune state rather than validated toxicity tests. The major studies cited above did not establish a locked receptor-based assay or test whether receptor features improve clinical decisions beyond established clinical and laboratory variables.^21,160 Their incremental predictive value should be evaluated prospectively within defined tumor-regimen settings and against prespecified, clinically consequential toxicity endpoints.

Early on-treatment dynamics may be more informative than baseline values

Treatment-induced immune changes can precede clinically apparent immune-related adverse events (irAEs) in selected cohorts. Longitudinal studies have linked subsequent toxicity to early B-cell remodeling, proinflammatory cytokine and chemokine programs, and expansion of activated or clonally expanded T-cell compartments.^20,161,170 The reported changes are heterogeneous and regimen-dependent rather than components of a uniform toxicity phenotype.²⁰ Early irAE development has also been associated with activated and cytotoxic T-cell programs, but receptor-defined CD8 T-cell subsets remain exploratory and require independent validation.^20,160 Blood transcriptomic and multi-omic studies have identified baseline and on-treatment inflammatory, neutrophil, and T-cell activation programs associated with subsequent irAEs, although organ-specific predictive performance remains limited.^171,172,173

These signals remain methodologically vulnerable. Samples obtained after irAE onset characterize active tissue injury rather than prospective risk, and immune profiles measured at this stage may reflect the inflammatory event and its evolving tissue response.^72,73,82 Dynamic biomarker studies should prespecify sampling landmarks, retain only measurements obtained before the relevant irAE, and use time-aware analyses that prevent post-event information from entering prediction.^20,87 Repeated sampling can capture transient signals, but high-dimensional analyses with few severe events require multiplicity control, patient-level inference, and independent validation.^127,128

An early-warning biomarker need not capture the complete pathogenesis of toxicity, but it should be reproducible, measurable before clinically consequential injury, and sufficiently specific to trigger a prespecified monitoring or diagnostic action.^21,90 The dynamic immune-cell signals reported in existing studies should remain exploratory because they have not been evaluated as locked assays across independent centers or tested in biomarker-guided surveillance studies.^20,160 Multi-omic studies indicate that toxicity risk is distributed across cellular, transcriptomic, and germline features, supporting integrated models rather than reliance on a receptor-defined subset alone.^21,160,172

Organ-specific toxicities require organ-specific prediction

Immune-related adverse events (irAEs) comprise organ-specific syndromes with distinct tissue immune programs and clinical consequences rather than a single biological phenotype.^82,86,174 Checkpoint regimens also produce different toxicity spectra: gastrointestinal and high-grade toxicities are more prominent with CTLA-4-containing therapy than with PD-1 monotherapy, whereas PD-1 pathway inhibition has a different distribution of organ-specific adverse events.^12,153,175 The relevant tolerance circuits likewise differ across tissues. Checkpoint colitis involves mucosal conventional and regulatory T cells, tissue-resident and cytotoxic lymphocyte programs, myeloid-cell activation, and local cytokine networks; colitis induced by CTLA-4 and PD-1 blockade can also exhibit distinct histological and immune-cell patterns.^72,82 Baseline thyroid autoantibodies are associated with an increased risk of thyroid irAEs, whereas reported HLA associations with hypothyroidism, hypophysitis, and other endocrine events remain exploratory and population-dependent.^166,167 Neurological irAEs are clinically heterogeneous, and neural autoantibodies are detectable in a subset of patients—particularly those with central nervous system or paraneoplastic-like syndromes—but antibody negativity does not exclude neurological toxicity.^176,177

Available tissue-resolved studies have largely analyzed clinically affected organs at or after irAE presentation rather than serially mapping vulnerable tissues before symptom onset.^72,86,174 Blood-based profiling can capture systemic susceptibility and treatment-induced immune activation, but influential models have generally been developed for aggregate all-grade or severe irAE outcomes rather than prospectively validated organ-specific risk.^21,160,172 Toxicity models should therefore distinguish any clinically significant irAE, severe or treatment-limiting toxicity, and prespecified organ-specific events because event grade, reversibility, and clinical hazard are not interchangeable.^21,87

Published biomarker cohorts have used different toxicity definitions, grade thresholds, surveillance schedules, and rules for assigning multisystem or recurrent events, limiting direct comparison and transportability.^21,160 Treatment-limiting toxicity can alter subsequent ICI exposure, whereas progression or death can preclude observation of later irAEs; longitudinal analyses should therefore preserve event timing and address informative treatment discontinuation and competing events.^12,87 Prospective studies will require harmonized outcome definitions, prespecified sampling schedules, patient-level sample-size calculations based on the anticipated number of events, and independent adjudication of clinically consequential toxicities.

Separating antitumor benefit from immune toxicity

Associations between irAEs and response do not imply that toxicity is required

Secondary analyses of randomized trials and observational cohorts have reported longer recurrence-free, progression-free, or overall survival among patients who develop immune-related adverse events (irAEs) during immune checkpoint inhibitor therapy.^83,84,85 The magnitude and direction of this association vary with toxicity grade, treatment setting, and analytical approach. Low-grade events may accompany favorable outcomes, whereas high-grade toxicity can be neutral or associated with shorter survival; these associations do not establish that toxicity is required for tumor control.^83,84,87 Because irAEs arise after treatment initiation, patients must remain alive and under observation long enough to be classified as exposed. Analyses that assign future irAE status as a fixed baseline variable are therefore vulnerable to guarantee-time or immortal-time bias and should use prespecified landmark or time-dependent methods.^3,83 Severe irAEs can force treatment discontinuation and systemic immunosuppression and may cause permanent organ injury or death—outcomes that are not captured by objective response alone.^10,12

The biology of benefit and toxicity overlaps but is not identical. Checkpoint blockade can mobilize antitumor and tissue-injuring T-cell responses in parallel, but paired-tissue analyses indicate that dominant T cell receptor repertoires may overlap in some patients and remain largely distinct in others.^10,60,174 Antitumor activity can be concentrated within tumor-reactive T-cell clonotypes, whereas irAEs may involve tissue-resident T-cell programs, early B-cell perturbations, and organ-localized inflammatory circuits.^20,82,149 These observations support pathway-selective mitigation rather than indiscriminate immune suppression: the aim is to interrupt tissue-injury pathways while preserving tumor-directed immunity. This remains a translational objective, because clinical evidence is preliminary and does not establish efficacy preservation across organs or regimens.^178,179,180

Selective mitigation is biologically plausible, but clinical evidence remains preliminary

Associations between immunosuppressive treatment and ICI outcomes vary according to the indication for therapy, its timing relative to ICI initiation, corticosteroid peak dose, and escalation to second-line immunosuppression.^181,182,183 Baseline corticosteroid exposure—particularly when prescribed for cancer-related palliative symptoms—has been associated with inferior outcomes in non-small-cell lung cancer. These observations should not be extrapolated directly to corticosteroids initiated after an established immune-related adverse event (irAE), for which outcome associations vary with timing and dose.^181,184 Accordingly, the available clinical data do not support the assumption that all forms of immunosuppression uniformly compromise checkpoint efficacy.

Preclinical studies provide proof of principle that selected inflammatory pathways can be inhibited without eliminating antitumor immunity. TNF blockade reduced checkpoint-related tissue injury while preserving or enhancing tumor control, and IL-23 blockade produced a similar separation in murine models.^178,185,186 Supported by this rationale, the phase Ib TICIMEL study showed that infliximab or certolizumab could be combined with nivolumab-ipilimumab with measurable clinical and immunological activity. Clinical evidence nevertheless remains preliminary: TICIMEL was small and non-randomized, and neither it nor subsequent observational analyses establish comparative efficacy or support routine prophylactic cytokine blockade.^180,183

In ICI-associated colitis, retrospective cohorts have associated early initiation of infliximab or vedolizumab with shorter symptom duration, fewer hospitalizations or steroid-taper failures, and reduced corticosteroid exposure. Vedolizumab and infliximab achieve similar clinical remission rates but differ in time to response, recurrence, and systemic steroid exposure.^179,187 These findings support further evaluation of organ- and pathway-directed treatment, but they do not establish preservation of antitumor efficacy across cancers, irAE phenotypes, or immunosuppressive regimens.^178,180,187 Current clinical evidence most directly supports early selective treatment of checkpoint colitis; local or cytokine-directed strategies for other organ toxicities remain investigational and are supported mainly by small clinical series or preclinical models.^178,179,186

ICI rechallenge should be individualized according to the initial organ toxicity, its severity and resolution, residual organ injury, prior antitumor benefit, available alternatives, and the regimen being resumed.^188,189,190 Because recurrence risk differs by organ and rechallenge regimen, the occurrence of any irAE should not be treated as an automatic permanent contraindication to further ICI therapy. Selective mitigation and rechallenge strategies require prospective assessment of both toxicity control and subsequent tumor outcome. The absence of an observed efficacy decrement in a small, non-randomized cohort should not be interpreted as evidence of equivalence.^180,182,187

Benefit and toxicity should remain separate, interpretable model outputs

Because antitumor benefit and immune toxicity are associated with partly distinct biological determinants, a dual-prediction framework should estimate separate, calibrated probabilities of clinically meaningful benefit and clinically important toxicity over prespecified time horizons.^19,21,160 Where event numbers support stable estimation, prespecified organ-specific risks should be modeled separately from aggregate immune-related adverse event risk.^72,174,176 The clinical meaning of a toxicity probability depends on event severity, organ involvement, reversibility, and the likelihood of treatment interruption or permanent injury.^10,87 Predicted benefit and toxicity should therefore be reported as separate outputs and interpreted in the context of competing treatment options, comorbidity, and patient preferences. Collapsing these dimensions into a single universal score would impose a fixed benefit-harm trade-off that may not reflect an individual patient’s priorities.

Figure 3 is a qualitative conceptual model rather than an empirical dose-response relation. Its axes are uncalibrated, and the regimen bands are illustrative; their overlap and ordering should not be interpreted as estimates from a randomized head-to-head comparison. Clinical trials show that benefit-toxicity profiles depend on treatment setting, prior therapy, regimen composition, and dose allocation. The incremental activity of nivolumab-relatlimab is treatment-setting dependent: the combination improved progression-free survival over nivolumab in untreated advanced melanoma, showed more modest activity after progression on prior PD-(L)1 therapy, and did not improve recurrence-free survival over nivolumab in completely resected stage III or IV melanoma.^145,152,154 Regimen composition and dose allocation also alter the benefit-toxicity profile. Nivolumab-ipilimumab produces substantially more high-grade toxicity than nivolumab alone, nivolumab-relatlimab increases high-grade toxicity more modestly relative to nivolumab, and lowering the ipilimumab dose within the nivolumab-ipilimumab regimen reduces severe toxicity, although that comparison was not designed to establish efficacy noninferiority.^11,145,191 Taken together, these trials support a regimen- and context-specific benefit-risk balance rather than a universal position along a single immune-activation axis. Quantifying that balance at the individual-patient level remains a prospective modeling objective.

Figure 3.

Conceptual separation of antitumor benefit and immune-related toxicity during checkpoint blockade

(A) Qualitative relationship between the extent of checkpoint release or immune activation, antitumor benefit, and the risk of immune-related adverse events. Antitumor benefit may increase and subsequently approach a plateau, whereas toxicity risk may rise disproportionately at higher levels of immune activation. The shaded region denotes an individualized therapeutic window in which clinically meaningful tumor control may be achieved with an acceptable risk of toxicity. The indicated ranges for PD-1 monotherapy, combined PD-1 and LAG-3 blockade, and combined PD-1 and CTLA-4 blockade are illustrative rather than quantitatively calibrated; their relative positions vary according to tumor type, dose, schedule, prior treatment, and patient characteristics.

(B) Selective mitigation may alter the toxicity trajectory without proportionally suppressing antitumor activity. Early detection, organ-directed treatment, and pathway-selective therapy could reduce the severity or duration of tissue injury while preserving tumor-directed immunity. The degree to which benefit and toxicity can be uncoupled remains regimen- and context-dependent and requires prospective clinical validation. irAE, immune-related adverse event.

From biomarker association to clinically credible dual prediction

Define the clinical decision before selecting biomarkers

Biomarker development should be anchored to a prespecified clinical decision and intended use rather than to post hoc selection of the strongest association in an available dataset.^157,158 Pretreatment regimen selection, early on-treatment monitoring, and post-irAE rechallenge are distinct clinical use cases because they differ in index time, predictor availability, target endpoint, and the action to be informed.^19,111,188 For pretreatment regimen selection, the prediction time should be fixed at ICI initiation; only variables available by the first dose should enter the model, and the target population, candidate regimens, efficacy and toxicity endpoints, and prediction horizons should be prespecified.^157,158 An early-monitoring model should prespecify the landmark time, use only information collected by that landmark, define the subsequent prediction window, and specify the clinical action that the result is intended to trigger.^109,110 A post-irAE rechallenge model should condition on the organ involved, severity and resolution of the initial event, residual organ injury, immunosuppressive treatment, prior antitumor benefit, and the ICI regimen being resumed; available treatment alternatives should be considered when the prediction is translated into a decision.^188,189,190

Outcome definitions should be aligned with the intended use and prespecified separately. Efficacy endpoints may include durable clinical benefit over a fixed horizon, pathological response in the neoadjuvant setting, or progression-free survival in advanced disease.^111,145,192 Toxicity endpoints may include severe or treatment-limiting irAEs, events requiring systemic immunosuppression, or selected organ-specific complications with distinct clinical hazards.^10,160,190 These outcomes are not interchangeable. A model developed for aggregate any-grade toxicity should not be interpreted as an organ-specific myocarditis model, and a model developed for objective response should not be assumed to predict progression-free or overall survival without separate validation.^10,87,193

The distinction between prognostic and treatment-predictive modeling is therefore essential. To support treatment selection, a biomarker must identify heterogeneity in the relative effect of competing regimens, ideally through a prespecified treatment-by-biomarker interaction in randomized data. A model that predicts outcome only within one treatment arm is a regimen-specific risk model, not a model of differential treatment benefit.^157,158 Applying this standard prevents associative immune signatures from being presented as validated tests for choosing monotherapy over combination treatment.

Prospective multicenter benchmarks should replace fragmented retrospective cohorts

A practical next step is a prospective, multicenter benchmark embedded within ICI trials or registries and anchored to a prespecified disease-regimen setting. Any subsequent pan-cancer extension should be treated as a separate transportability question and evaluated explicitly across independent tumor types, treatment regimens, and clinical cohorts.^172,192,194 Participating centers should collect harmonized baseline clinical data, pretreatment blood, and early on-treatment blood within protocol-defined windows relative to the first ICI dose. Pretreatment tumor tissue should be obtained when feasible, whereas an on-treatment biopsy may be reserved for a prespecified mechanistic subset in which the additional procedure is scientifically and clinically justified.^20,110,111 Core outcomes should be prespecified separately and should include a clinically defined measure of durable benefit or a time-to-event efficacy endpoint, severe or treatment-limiting irAEs, and selected adjudicated organ-specific complications. Aggregate toxicity, severe toxicity, and rare high-consequence organ events should not be treated as interchangeable labels.^10,160

The assay strategy should separate a high-dimensional discovery tier from a lower-complexity clinical tier. Discovery platforms may nominate candidate cellular, clonal, transcriptional, or spatial states, but these features should undergo orthogonal confirmation, marker reduction, and independent validation before incorporation into a locked clinical assay.^90,108 All participants should undergo a common clinical-tier assay selected for the intended specimen and use case—for example, a fixed mIF panel in tissue and a compact flow cytometry panel in blood—while higher-dimensional assays may be restricted to nested discovery subsets.^90,108 Multicenter deployment requires common reference materials, locked standard operating procedures, predefined analytical acceptance criteria, and centralized or independently validated analysis rules. Inter-laboratory proficiency testing should be completed before clinical validation begins.^90,113

Development and validation sets should be separated at the patient level. Cells, image regions, serial aliquots, and technical replicates derived from the same patient are correlated observations and should not be distributed across training and validation sets, because such splitting permits information leakage and yields optimistic performance estimates.^127,195 In multicenter studies, center or cohort membership should also define validation strata. After model development, all feature definitions, preprocessing steps, coefficients, and decision thresholds should be locked and evaluated in centers or cohorts that contributed no development data. Internal cross-validation remains useful for model development but does not replace evaluation of transportability across laboratories, populations, and clinical settings.^192,194,196

The benchmark should capture prespecified clinical, pathological, genomic, and laboratory variables that are available at the intended prediction time. Inhibitory-receptor features should be evaluated for reproducible out-of-sample incremental value over a transparent, clinically credible reference model rather than compared only with a single weak biomarker or an intentionally under-specified baseline.^19,172,196

Statistical development must respect time, hierarchy, and calibration

High-dimensional immune studies may generate many cell- or region-level measurements, but their independent clinical sample size is determined by the number of patients and outcome events. The patient, rather than the individual cell or image region, is the unit of clinical inference; observations obtained from the same patient are correlated subsamples and require patient-level aggregation or hierarchical modeling.^127,128,197 All data-dependent preprocessing—including normalization, imputation, feature selection, batch adjustment and hyperparameter tuning—should be estimated within the training resamples and then applied unchanged to held-out data.^195,198 Study design should avoid structural confounding of center, assay batch, treatment and outcome. Relative cell frequencies should be treated as compositional measurements and interpreted against an explicit reference; whenever feasible, proportions should be accompanied by absolute counts or another measure of total cellular abundance.^118,199,200

The statistical method should match the endpoint and prediction time. Models for immune-related adverse events should preserve event timing. When death, progression, or permanent treatment discontinuation precludes observation of the toxicity of interest, the estimand and handling of competing events should be prespecified rather than treating these events as non-informative censoring.^83,87 Dynamic biomarkers require landmark, joint, or other time-dependent models that use only measurements available by the prediction time.^201,202 Sample-size calculations should be based on the number of independent patients, anticipated outcome events, candidate predictor parameters and the precision required for model-performance estimates—not on the number of cells, transcripts or image regions measured.^127,197

Discrimination should be accompanied by calibration plots, calibration intercept and slope, and an overall accuracy measure such as the Brier score. When predictions are intended to trigger an intervention, clinical utility should also be evaluated across clinically relevant thresholds.^203,204 Performance should be evaluated across centers and relevant subgroups rather than reported only as a pooled average. Limited, prespecified recalibration of the intercept or baseline hazard may be appropriate when baseline risk differs, but predictor definitions and the structural form of the model should remain locked during external evaluation.^205,206 Reporting should permit independent reconstruction and appraisal of the complete modeling pipeline, including data sources, predictor and outcome definitions, preprocessing, feature selection, model fitting, validation, calibration, subgroup performance and clinical utility, irrespective of whether regression or machine-learning methods are used.^207,208 Algorithmic complexity should be justified empirically rather than assumed to confer superior performance; model class should be selected according to out-of-sample performance, calibration, stability, data structure and interpretability.^209,210

Translate discovery signals into a locked minimal assay and test clinical utility

A four-stage pathway offers a route from discovery to clinical use. Stage 1 is mechanistic discovery in paired tumor and blood samples with serial blood collection. The aim is to identify reproducible cell states and treatment-induced changes that add information beyond routine variables.^23,60 Stage 2 is assay reduction and orthogonal verification. Candidate signals should be converted into a lower-complexity platform, such as multiplex immunofluorescence in tissue and standardized flow cytometry in blood.^90,108 Marker definitions, cell denominators, reagents, quality-control thresholds, and analysis code should be fixed before validation.

• Stage 3 is external validation across centers and bridging across platforms. The minimal panel should be tested in an independent cohort receiving the same regimen, with prespecified endpoints and no feature reselection. Bridging studies should establish whether measurements from different instruments or laboratories are equivalent within predefined limits. The output should provide calibrated estimates of benefit and toxicity, uncertainty intervals, and an indication of whether the patient resembles the validation population.^203,211

• Stage 4 is a prospective clinical-utility study in which the prediction changes a predefined action. Examples include choosing monotherapy rather than combination therapy, intensifying surveillance for a high-risk organ, or informing rechallenge after toxicity has resolved. The study must test whether model-guided care improves outcomes, safety, or decision quality, not merely whether the score remains associated with events.^212,213 Figure 4 summarizes the progression from sampling and high-dimensional profiling to clinical action.

Figure 4.

Staged development and validation of a clinically deployable dual-prediction model

Stage 1 comprises mechanistic discovery in a defined disease-regimen setting using paired tumor and blood specimens, pretreatment and early on-treatment sampling, and standardized clinical annotation. Multiplex immunofluorescence, flow cytometry or CyTOF, bulk RNA profiling, single-cell RNA and T cell receptor sequencing, and spatial profiling are used to identify candidate inhibitory receptor states and treatment-induced dynamics. Stage 2 converts discovery signals into a clinically tractable assay through orthogonal verification, marker reduction, explicit cell denominators, locked reagents and analytical procedures, and predefined quality-control criteria. Stage 3 evaluates a locked model in independent multicenter cohorts, with patient-level separation of development and validation sets, time-aware analysis, prevention of information leakage, and assessment of calibration, uncertainty, external quality control, and incremental value over established clinical variables. Model outputs remain separate and include the probability of antitumor benefit, severe immune-related toxicity, and, where event numbers permit, organ-specific toxicity. Stage 4 tests clinical utility by incorporating these estimates into regimen selection, organ-specific surveillance, and decisions regarding treatment rechallenge, with the objective of improving decision quality and patient safety. Standardization, calibration, version control, and prespecified model updating are required throughout the translational pathway. CyTOF, cytometry by time of flight; irAE, immune-related adverse event; mIF, multiplex immunofluorescence; QC, quality control; scRNA-seq, single-cell RNA sequencing; TCR-seq, T cell receptor sequencing.

Implementation will require standard operating procedures, external quality assessment, version control, and predefined rules for model updating. Models should be reassessed when regimens, assay platforms, or event-ascertainment practices change. Without this infrastructure, high-resolution immune profiling will remain biologically informative but clinically unreliable.^90,205,214

Limitations and unresolved questions

Inhibitory receptors are informative but inherently non-specific biomarkers

Inhibitory receptor expression integrates activation history, differentiation, and local environmental cues. This property is both its strength and its limitation. Similar phenotypes may arise from tumor recognition, infection, bystander inflammation, chronic antigen exposure, or autoimmunity. Surface abundance does not necessarily reflect signaling competence, ligand availability, receptor occupancy, or downstream pathway activity. Receptor profiles are also dynamic, treatment-sensitive, and heterogeneous across lesions within the same patient. These features make universal, stand-alone cutoffs biologically implausible and require interpretation within a defined cellular, spatial, and clinical context.

The evidence base remains heterogeneous and largely exploratory

Most studies are retrospective, single-center, or based on biospecimens collected in trials not designed for biomarker validation. Cohorts differ in tumor type, regimen, sampling schedule, assay platform, endpoint definition, and ascertainment of immune-related adverse events. Rare organ-specific toxicities seldom provide enough events for stable multivariable modeling, while publication bias may favor positive findings. Prospective validation remains limited, and no study has shown that an inhibitory receptor-based model improves treatment selection, surveillance, or patient outcomes.

Limitations of the study

This article is a narrative rather than a systematic review. Study selection was not governed by a preregistered protocol, and heterogeneity precluded quantitative synthesis. Evidence outside the English-language literature may be underrepresented. The discussion is T-cell-centered because T-cell studies dominate the field, although B cells, myeloid cells, natural killer cells, tissue-resident populations, and non-immune determinants may be equally important in selected settings. The proposed translational pathway is an expert framework requiring prospective evaluation, not a validated clinical algorithm.

Conclusion

Inhibitory receptor profiles provide a biologically coherent framework for linking two central outcomes of checkpoint blockade: tumor control and immune toxicity. Their value does not depend on the assumption that higher expression is uniformly favorable or unfavorable. Rather, these profiles describe the abundance, state, location, and treatment responsiveness of immune populations involved in antitumor immunity and tissue tolerance.

Current evidence is strongest for context-dependent associations involving multi-receptor states, clonality, and early on-treatment dynamics. It does not support universal thresholds or stand-alone clinical use. Progress will require regimen-specific questions, harmonized prospective sampling, rigorous analysis of time-dependent outcomes, external validation, and reduction of high-dimensional signals to locked, reproducible assays. Benefit and toxicity should remain separate outputs because they share pathways without being biologically or clinically inseparable.

A credible dual-prediction strategy must be developed incrementally. It should retain sufficient biological resolution, perform consistently across centers, and show that its use improves a defined clinical decision. Until then, inhibitory receptor profiles should be regarded as promising components of integrated biomarker models rather than ready-to-use clinical tests.

Data and code availability

This paper does not report data generation or analysis.

Acknowledgments

The authors thank Professor Bingwen Zou, Department of Radiation Oncology, Cancer Center & Institution of Stress Medicine, West China Hospital, Sichuan University, for valuable advice on the conception and organization of this Review, guidance during revision, and critical comments on the manuscript. This work was supported by the National Science and Technology Major Program of China under grant nos. 2026ZD0555902 and 2023ZD0506105, the Research and Application Demonstration of Key Technologies for Computing Power and Data Sharing Services in Large AI Models, and the Department of Science and Technology of Hunan Province (2025JK2004). 3) Declaration of interests: The authors declare no competing interests. 4) Declaration of generative AI and AI-assisted technologies in the manuscript preparation process: During revision, ChatGPT (OpenAI) was used only for English-language editing and preliminary figure ideation. All scientific content, interpretations, references, and final wording were checked and approved by the authors. The final figures were prepared manually by the authors in Adobe Photoshop; no AI-generated image is included in the manuscript. 5) In the manuscript please remove hyphen from “T cell” and “B cell”.

Author contributions

X.W. and Q.W. conceived and designed the review. X.W. conducted the literature review and drafted the original manuscript. D.X. and B.D. contributed to literature screening, evidence interpretation, and preparation of the figures and tables. G.D. and Y.H. contributed to the interpretation of the oncological and immunological evidence and critically revised the manuscript. Q.W. supervised the work, critically revised the manuscript, and acquired funding. All authors read and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

The authors thank Professor Bingwen Zou, Department of Radiation Oncology, Cancer Center & Institution of Stress Medicine, West China Hospital, Sichuan University, for valuable advice on the conception and organization of this Review, guidance during revision, and critical comments on the manuscript. During revision, ChatGPT (OpenAI) was used only for English-language editing and preliminary figure ideation. All scientific content, interpretations, references, and final wording were checked and approved by the authors. The final figures were prepared manually by the authors in Adobe Photoshop; no AI-generated image is included in the manuscript.