ICP were profiled using 17 lung cancer cohorts from differing origins, and using different methods: 1) at the protein expression level on a TMA created from a baseline NSCLC cohort (n = 81) (Additional file 1: Table S1; La Rabta Hospital of Tunis, Tunis, Tunisia); 2) at the whole-tumor RNA level using RNA-Seq datasets from two NSCLC cohorts from the TCGA, the LUAD (n = 504) and LUSC (n = 494) (http://www.cbioportal.org); and 3) at the whole-tumor RNA level using microarray datasets from 14 NSCLC cohorts from the GEO, the EGA and TCGA (n = 2435) Kaplan-Meier Plotter (http://kmplot.com). Additional breast (n = 5143), gastric (n = 2183), and ovarian (n = 1816) cohort datasets were from Kaplan-Meier Plotter. Written and informed consent procedures were approved by the ethics review committees and were obtained from patients prior to the collection of specimens. Clinical patient data was randomly numbered for complete anonymity. Censoring of cohort patient data was from time of diagnosis to last follow-up or death.

An illustration of the TMA construction is provided in Fig. 1a. Four μm cuts made using a microtome (Leica Biosystems) from all biopsies were α-CD45 stained for IHC using the Benchmark XT automated stainer with CC1 antigen retrieval buffer (Ventana Medical Systems) for 1 h. Slides were incubated with α-CD45 (1:50) at 37 °C for 1 h, followed by the ultraView DAB detection kit and counterstaining with haematoxylin and bluing reagent (Ventana Medical Systems). Slides were scanned with an Olympus BX61VS microscope equipped with a VS110 slide scanner and 20x / 0.75 NA objective with a resolution of 0.3225 mm (Olympus). Images were exported and visualized using OlyVia image viewer software ver. 2.8 (Olympus) to identify CD45⁺ IIC-rich regions. Three to five IIC-rich regions of the biopsies were selected for 0.6 mm core transfer into the receiving TMA paraffin block using a TMArrayer (Pathology Devices). Paraffin blocks were kept at 4 °C until used for TMA construction. TMA cores were press-sealed into place after incubation at 50 °C for 10 min. The TMA was cooled at RT ON, and was chilled on ice ahead of being cut into 4 μm sections. Sections were floated onto 1 mm slides (Fisher Scientific), dried ON, and stored at 4 °C until stained.Fig1Fig. 1

Creation and analysis of IIC-enriched biopsy-based NSCLC TMA. a Illustration depicting TMA creation workflow. Baseline biopsies from a NSCLC patient cohort (n = 81) were paraffin embedded, and cut sections were stained using α-CD45 to demarcate IIC-dense regions then selected for TMA construction using original blocks. Cut sections from resulting TMA were then stained using MP-IF panels targeting immune-related antigens including ICPs and IIC subsets. Slides were scanned to create super images permitting the development of algorithms computing antigens of interest and their colocalization for normalization (figure elements modified from Servier medical art). b Image representing α-CD45 IHC stained biopsies defining IIC-dense areas. c Example of MP-IF panels demonstrating α-ICP (green), α-CD3 (pink), α-CD4 (red), and α-CD8 (yellow) antibodies validated to surround DAPI-staining nuclei (blue). IIC-enriched core selection was performed by two different operators. TMA cores were randomized and TMAs were created by two operators. HRP, horseradish peroxidase; 2° ab, secondary antibody; AF, Alexa-Fluor dye; α, anti; μm, micron; mm, millimeter

TMA sections were deparaffinized by incubation at 50 °C for 1 h prior to 5 min incubations in successive baths (3x xylene, 95, 90, 70, and 50% ethanol, dH₂O). Antigen retrieval was performed using Target Retrieval Solution, Citrate pH 6 (DAKO) as recommended by the manufacturer. Protein Block (DAKO) was applied against non-specific staining for 40 min. Slides were rinsed with PBS before incubation with primary antibody mixtures diluted in Antibody Diluent (DAKO), 0.05% Tween 20 (Fisher Scientific) ON in a humidified chamber at 4 °C. Antibodies and their dilutions are in Additional file 1: Table S3. Following three 15 min PBS washes, slides were incubated with secondary antibody mixes for 1.5 h at RT (cross-adsorbed donkey α-rabbit, α-rat, or α-goat IgG (H&L) and/or goat α-mouse IgG1, IgG2a, IgG2b or IgGM specific secondary antibodies conjugated to Alexa-Fluors (405, 488, 594, 647 and 750) (ThermoFisher Scientific and Abcam) (1:250) Additional file 1: Table S3. Slides were washed with three 15 min incubations in PBS, and incubated in Sudan Black (1% in 70% ethanol) for 15 min. Slides were washed with dH₂O for 5 min, and dried for 30 min before being set with ProLong gold antifade reagent (±DAPI) (ThermoFisher Scientific) under 0.17 mm coverslips (Fisher Scientific). Primary antibodies were individually detected by donkey α-host IgG (H&L) Alexa-Fluor 594 antibodies, and images were acquired using a Zeiss Axio Observer Z1 automated microscope equipped with a Plan-Apochromat 20x / 0.8 NA objective, a Zeiss HRm Axiocam and LED pulsed light illumination (Additional file 1: Figure S1d). Fluorescence minus one controls were used for potential fluorescence bleed-through between detection channels. In other control experiments, primary antibodies: 1) were not added, 2) were detected by alternative secondary antibodies, 3) were tested on a TMA containing 14 cancer cell lines (e.g., prostate, breast, ovarian, kidney, cervical cancer cells and Jurkat), and 4) were replaced with isotype control antibodies (MOPC-31C, G155–178, MPC-11) (BD Pharminogen). MP-IF stained slides were scanned using an Olympus BX61VS microscope housing a BrightLine Sedat filter set (Semrock) optimized for DAPI, FITC, TRITC, Cy5 & Cy7, and equipped with a 20x / 0.75 NA objective with a resolution of 0.3225 mm and a VS110 slide scanner running FW-AS software (Olympus) that stitches individual images to build high resolution .vsi images.

High resolution images were imported into Visiomorph software (Visiopharm), where cores were identified and linked to patient numbers using an Array-Imager module. Using fluorescence intensity thresholding, algorithms were designed to define a region of interest (ROI) and calculate total core area, which was further trained to eliminate holes within tissues to correct for actual tissue-occupying areas (Additional file 1: Figure S1f). Two independent operators used fluorescence intensity thresholding and size exclusions to create algorithms generating labels counting cells positive for biomarkers. Single marker labeling and co-labeling of dual, triple, and quadruple colocalizing markers were performed in the same way. For co-labeling, labels created for counting cells positive for multiple biomarkers were determined using the same thresholds used to identify and count single marker labeling cells. Created co-labels were also verified as accurately staining immune cells by two independent operators. Labels identifying markers were adjusted for IIC sizes, and were centered on DAPI staining when present in panels. Baseline fluorescence thresholds assigned for minimal signal to noise ratios determining positivity were used to calculate MFIs. Counts of algorithm-determined labels on cores were validated to reflect visual operator counts. Inter-rating correlations from algorithms created by independent operators was assessed to be > 75%. Each single or multi-marker label counts (e.g., totaling up to 15 marker permutations for each individual 5-color panel in case of DAPI + 4 markers) of individual cores were automated to be reassigned to patient ID numbers, and were then log-transformed and normalized to core size, prior to being merged with the clinical data for averaging of replicate core values, resulting in data from 73 patients for further analyses from .csv data file exports. High (hi) and low (lo) values were defined as being above or below mean ± SEM. Receiver operative characteristic (ROC) curves (SPSS software v.23, IBM), were used to validate that selected cutoff values corresponded to the best sensitivity and specificity any given marker. ICPs having inter-patient variability were found from a second method of analysis applied whereby values from individual cores were not averaged.

Power analysis determined that our retrospective biomarker study based on overall patient survival required a minimal sample size of n = 62 to reach a power of 0.80 at α = 0.05 (two-tailed) (G*Power ver. 3.1.9.2; Universitat Düsseldorf, Germany). Prism 6 ver. 6.01 (GraphPad) and SPSS software packages were used for statistical analysis of biomarkers with patient data. Log-rank (Mantel-Cox) tests with log-rank HR were used for K-M. A student’s t test was used to compare two groups, and two-way ANOVA (with Tukey’s or Bonferroni’s multiple-comparisons tests) was used for multiple comparisons. Pearson correlation coefficients were calculated with two-tailed P values with 95% confidence intervals. P-values of less than 0.05 were considered to indicate a statistically significant difference. R with a collection of libraries was used for additional statistical correlation, linear regression, variance and clustering analysis, patient clinical characteristics and biomarker expression value relationships analyses. Here, expression values were log transformed towards a Gaussian distribution. Linear regression matrices were computed using the R glm function. Link functions were adapted phenotype distribution type (binomial, Gaussian, Poisson) for model compatibilities for explorations of relationships between biomarkers and clinical data. K-M calculations, cox model p-values and HR were validated using a survival model coupling survival status and months of survival post biopsy. PCA was used for coexpression analysis. Cumulative correlations for the expression of each ICP (and CD3-ICP) were calculated from their respective correlation matrices.

Kaplan Meier plotter was used to validate the prognostic value of the ICP signature, and to assess ICP gene expression modulation between tumors and normal tissues. Gene ID symbols were mapped to Affymetrix probes from GEO, EGA and TCGA datasets, and their mean expression was used to assess OS. For K-M, default settings were used with auto select best cutoff and best specific probes (JetSet probes). The 2017 version of Kaplan Meier plotter contains information on 54,675 genes for survival, including 2437 lung, 5143 breast, 1065 gastric, and 1816 ovarian cancer patients with mean follow-up times of 49, 69, 33, and 40 months, respectively. Multigene classifier function using default settings from KM-plotter was used to run the analysis on all ICPs simultaneously, where global ICP coexpression represents combined prognostic effects of all ICPs investigated in this study.

Identified biomarkers were subjected to comprehensive pathway enrichment analysis using pathDIP ver. 2.5 (http://ophid.utoronto.ca/pathDIP) (Additional files 2 and 3). Default settings were used, with extended pathway associations (combining literature curated core pathways with associations predicted using physical protein interactions with minimum confidence levels of 0.99). Lists were also used to retrieve physical protein interactions and explore biologically relevant links. IID ver. 2016–03 (http://ophid.utoronto.ca/iid) was used to map identified biomarkers to proteins and retrieve their interacting partners. Default settings were used, and interactions among partners of query proteins, source information (detection methods, PubMed IDs, reporting databases), and tissue information (presence/absence of interactions in selected tissues) were included. Corresponding networks were visualized using NAViGaTOR ver. 3 (http://ophid.utoronto.ca/navigator) (Additional file 4). Word-cloud analysis was performed using Wordle software ver. 2014 (http://www.wordle.net).

We aimed to develop a standardizable, immune-based, prognostic scoring method for biopsies. To reduce tumoral heterogeneity, a CD45-enriched TMA was constructed from baseline biopsies from a NSCLC cohort (Additional file 1: Tables S1 and S2). Figure 1a illustrates the construction of the TMA. Ahead of construction, nine random biopsy sections where stained for immunofluorescence (IF) using DAPI, α-CD45 and α-cytokeratin; verifying these for epithelial cancer and IIC-densities (Additional file 1: Figure S1a). Cut sections from all biopsies were then stained for IHC using α-CD45, defining IIC dense regions selected for TMA construction (Fig. 1b). IIC density of biopsies did not correlate with clinical parameters (P > 0.416) (Additional file 1: Figure S1b) or overall survival (OS) (P = 0.7880) (Additional file 1: Figure S1c). All antibodies were validated independently (Additional file 1: Figure S1d and e), and TMAs were stained with five-color multiplex-IF (MP-IF) panels using a two-step, semi-automated method (Fig. 1a and c). Algorithms calculated core areas to normalize labels identifying size- and fluorescence intensity-gated, colocalizing IICs and ICPs (Additional file 1: Figure S1f).

To determine whether IIC subsets and activation markers could predict OS, TMAs were stained with MP-IF panels labeling CD45⁺ leukocytes; CD3⁺, CD4⁺, and CD8⁺ T cells; CD20⁺ B cells; CD56⁺ natural killer (NK) cells; CD68⁺ macrophages; proliferating cells (Ki-67⁺); and activation and cytotoxic markers (human leukocyte antigen-DR, HLA-DR⁺; granzyme B, GZMB⁺; interferon-gamma, IFN-γ⁺). IIC densities of TMA cores had Gaussian distribution (Additional file 1: Figure S2a). Kaplan-Meier survival analyses (K-M) demonstrated that CD45 density did not correlate with OS (P = 0.4763) (Fig. 2a and Additional file 1: Figure S2b), as expected from its demarcating all IIC subsets having differential effects on prognoses. Ki-67 was associated with positive OS (P = 0.0068) (Fig. 2a and Additional file 1: Figure S2b), contrary to Ki-67 in cancer-centric studies [18], and attributable to an IIC-enriched TMA. OS was associated with CD45-Ki-67 co-labeling cells (P = 0.0040) (Fig. 2a and b). The same was observed for TILs, where association of CD3 with OS was enhanced by Ki-67 co-labeling (P = 0.0297 to P = 0.0044) (Fig. 2a and b, and Additional file 1: Figure S2b). CD4⁺ TILs were modestly associated with OS (P = 0.0453) (Fig. 2a and c), likely due to this mixed population having differential effects on prognosis [19]. CD8⁺ TILs strongly associated with OS (P = 0.0074) (Fig. 2a and c) [20].Fig2Fig. 2

Highly proliferating, effector TIL and TIL-B densities are linked to positive prognostic of NSCLC patients. a Summarizing graph of P-values generated from K-M survival analyses of markers applied to IIC-enriched biopsy TMA, where significance indicates positive associations of IIC subsets, and proliferation and effector molecules with OS. b K-M curves (top) from Ki-67 co-labeling with CD45⁺ IICs or CD3⁺ TILs on TMA, and representative close-up IF images from cores (bottom) demonstrating co-labeling on cells. c K-M curves (top) from CD4⁺ and CD8⁺ TILs on TMA, with representative close-up IF images from cores (bottom) demonstrating their co-labeling CD3⁺ TILs. d K-M curves (top) from CD20⁺ TIL-Bs, PNAd⁺ HEV, and CD68⁺ TAMs, with representative close-up IF images from cores (bottom). e Graph of average proportions of IIC subsets relative cell count (DAPI), where percentages represent IIC subset abundance relative to CD45⁺ IICs. Percentages are relative to CD45 content, and error bars represent mean ± s.d.. f Graph of correlations between IIC subsets and quantified effector molecules (IFN-γ, GZMB, HLA-DR). Percentages represent IIC subset attribution to effector molecule expression, as calculated from proportions of individual IIC subsets infiltrating cores expressing effector molecules. g K-M curves (top) of GZMB, IFN-γ, and HLA-DR effector markers, with representative close-up IF images from cores (bottom) of these markers and TILs. The number of patients (n) for each group is given on K-M curves, and remainder are in Additional file 1: Figure S2b. Algorithm design, normalization and analyses were performed by two independent operators. Norm., normalized; hi, high marker expression, lo, low marker expression; μm, micron; P, Log-rank test; ns, not significant; * P < 0.05; ** P < 0.01; *** P < 0.001; HR, hazard ratio (Log-rank); CI, confidence interval of ratio; NA, not applicable

IFN-γ expression by activated TILs increases PD-L1 expression [32]. IFN-γ is also correlated with the expression of other ICPs, including BTLA [33], TIM-3 [34], LAG-3 [35], and PD-1 [36]. Since ICPs are expressed by various cell types, their usage as mono-CDx will lead to assay inconsistencies exemplified by PD-L1 [37]. Indeed, on our TMA, certain ICPs labeled numerous cells types (PD-L1, TIM-3, TIGIT, LAIR-1, CD73), whereas others almost exclusively labeled TILs (BTLA, LAG-3, PD-1, CD39, 2B4, CD57, CD26, CLTA-4) (Additional file 1: Figure S3a to e). Despite this, principal component analysis (PCA) demonstrated that relative to patients, tight clustering of ICPs and cognate CD3-ICPs indicated that they were mostly labeling TILs, and not other cells of the tumor microenvironment (Additional file 1: Figure S3f).

The only ICP associated with positive OS independently of TILs was TIM-3 (P = 0.0448), and this was augmented by it co-labeling CD3⁺ TILs (P = 0.0151) (Fig. 3a). Association with OS for other ICPs was only met by their co-labeling CD3⁺ TILs: CD3-TIGIT (P = 0.0188), CD3-LAG-3 (P = 0.0251), CD3-BTLA (P = 0.0167), and CD3-PD-1 (P = 0.0189) (Fig. 3a). While mean fluorescence intensities (MFI) of ICPs or all other markers tested showed no association with OS, some correlated with clinicopathological characteristics (Additional file 1: Table S4).Fig3Fig. 3

Effects of ICP expression on NSCLC patients. a Summarizing graph of P-values generated from K-M survival analyses of IIC-enriched TMA, where significance indicates positive associations ICP and CD3-ICP co-labeling cells with OS (top left). K-M curves and representative close-up IF images from cores (right and bottom) of CD3 dense core areas, demonstrating colocalization between CD3 and TIGIT, TIM-3, LAG-3, BTLA, or PD-1. b Summarizing graph of P-values generated from K-M survival analyses of IIC-enriched TMA, where significance indicates positive associations of combinations of CD3, CD8, PD-1, PD-L1 and TIM-3 with OS (top). Representative close-up IF images from cores (bottom) of CD3 dense core areas, demonstrating colocalization between these antigens. The number of patients (n) for each group is given on K-M curves, and remainders are (high and low, respectively): CD3 n = 34hi, 32lo; CD8 n = 41hi, 21lo; TIGIT n = 26hi, 38lo, TIM-3 n = 21hi, 26lo, LAG-3 n = 29hi, 33lo, BTLA n = 30hi, 30lo, PD-1 n = 36hi, 30lo, CD3-PD-1 n = 29hi, 24lo; CD8-PD-1 n = 36hi, 32lo; TIM-3-PD-1 n = 34hi, 30lo; PD-L1 n = 25hi, 33lo; CD3-PD-L1 n = 18hi, 30lo; CD8-PD-L1 n = 19hi, 24lo; TIM-3-PD-L1 n = 25hi, 34lo. Algorithm design, normalization and analyses were performed by two independent operators. hi, high marker expression, lo, low marker expression; μm, micron; merge, merge of all IF channels; P, Log-rank test; ns, not significant (implied when no asterisk is present); * P < 0.05; ** P < 0.01; HR, hazard ratio (Log-rank), CI, confidence interval of ratio

In correlative analyses between global ICP or CD3-ICP expression and IIC subsets, IIC subset infiltration of patient cores were used to demonstrated that expression of ICPs and CD3-ICPs effector markers could be most associated with the presence of CD8⁺, CD20⁺ and CD4⁺ IIC subsets (Fig. 4a and b). We tested whether the IIC-density of biopsies influenced CD3 and ICP distributions. CD3⁺ TILs were highly correlated with CD45⁺ IICs (P < 0.0001, r = 0.3428), but global ICP expression was not (Fig. 4c), with the exception of CD3-PD-1, CD3-PD-L1, CD3-BTLA and CD3-LAG-3 (Additional file 1: Table S5). This also supports that ICPs are not uniquely expressed by TILs (ICP vs CD3-ICP; P < 0.001) (Fig. 4c and Additional file 1: Figure S3a to e) [38, 39]. ICPs correlating with CD3 were BTLA, LAG-3, TIM-3 and CD26, and CD73 and CD3-CD73 correlated with the ADC subtype [40] (Additional file 1: Table S5). Despite their clear effects on outcomes (Additional file 1: Figure S4), there was no correlation between treatments and ICP expression. We also observed that CD3-ICPs were inversely correlated with tumor size and extent (Fig. 4d and Additional file 1: Table S5). K-M performed using global expression of ICP or CD3-ICP revealed that both positively correlated with OS (Fig. 4e and f), and global CD3-ICP expression also correlated with female gender (P = 0.0321, r = 0.0701).Fig4Fig. 4

Effects of IIC density on global ICP expression and validation of global ICP prognostic effects on various cancers. a-b Graphs demonstrating correlations between TMA IIC subsets and (a) ICP or (b) CD3-ICP expression (%, IIC subset attribution; n = 73). Percentages represent IIC subset attribution to ICP or CD3-ICP expression, as calculated from proportions of individual IIC subsets infiltrating cores expressing ICPs or CD3-ICPs. (c) Graph demonstrating correlation between IIC-density of biopsies and CD3⁺ TILs, ICPs and CD3-ICPs. Two-way ANOVA with Bonferroni’s multiple comparisons test; n = 73, CD3, P < 0.0001; ICP vs CD3-ICP P = 0.005; F = 12.06, df = 1/219; error bars represent mean ± s.e.m.. d Graph demonstrating correlation of advancing T-stages with ICP expression (T2, n = 206; T3, n = 106; T4, n = 511) and CD3-ICPs (T2, n = 199; T3, n = 120; T4, n = 496) expression on TMA (two-way ANOVA with Tukey’s multiple-comparison; CD3-ICP T2 vs T4, F = 2.97, df = 2/1632, P = 0.0085; error bars represent mean ± s.e.m.). e K-M curve of total TMA ICP (P = 0.0273, HR [95% CI] = 0.514 [0.248–0.883], n = 32hi, n = 34lo) overlaid with number of ICP/patient relative to survival in months (green circles and right axis; dotted line, high vs low); linear regression of overlay F = 9.41, df = 1/62, P = 0.0032, R² = 0.132. f K-M curve of total TMA CD3-ICP (P = 0.0472, HR [95% CI] = 0.546 [0.270–0.952], n = 30hi, n = 36lo) overlaid with number of CD3-ICP/patient relative to survival in months (green circles and right axis; dotted line, high vs low); linear regression of overlay F = 5.56, df = 1/63, P = 0.0215, R² = 0.081. g Graph demonstrating correlation of advancing stage with ICP expression levels from LUAD dataset. Two-way ANOVA with Bonferroni’s multiple comparisons test, Stages I, n = 274; II, n = 121; III, n = 81; IV, n = 26, where stages I vs IV from both ICP hi or lo are P < 0.0001, F = 9.78, df = 3/996; error bars represent mean ± s.d.. h Graph demonstrating correlation of survival with ICP expression from TCGA LUAD and LUSC datasets. Two-way ANOVA with Bonferroni’s multiple comparisons test, P < 0.0001, F = 29.94, df = 1/828; ICP DCD, n = 172hi, 254lo; ICP SURV, n = 228hi, 178lo; error bars represent mean ± s.d.. a-h Algorithm design, normalization and analyses were performed by two independent operators. i-l K-M plots validating effects of global ICP expression on new cohorts of (i) NSCLC (n = 783hi, 362lo), (j) breast (n = 386hi, 240lo), (k) gastric (n = 265hi, 366lo), and (l) ovarian (n = 275hi, 380lo) cancer patients. Two-way ANOVA with Tukey’s post-test; norm., normalized; n, number of patients; SURV, surviving; DCD, deceased; P, Log-rank test; ns, not significant; ** P < 0.01; *** P < 0.001; **** P < 0.0001; HR, hazard ratio (Log-rank), CI, confidence interval of ratio

Using TMAs, we assessed minimal ICP combinations on TILs maximizing prognostic value (Additional file 1: Table S12). Indeed, the TIM-3/CD26/CD39 combination had a stronger association with OS than these did independently (P = 0.0139), and was superior when co-labeling with CD3 (P = 0.0051) (Fig. 5a). The positive effect on OS was maintained with ICPs and CD3-ICPs co-labeling for TIM-3/BTLA/LAG-3 combinations (P = 0.0018 to P = 0.0033), as it was for the 2B4/PD-1/CD57 combination (Fig. 5b and c). As supported by imaging (Additional file 1: Figure S6), comparisons of ICP and CD3-ICP K-M curves validated that these ICP combinations were specifically labeling TILs, and that the difference in prognostic association using duplex or triplex ICP panels was dependent on ICP combinations.Fig5Fig. 5

MP-IF panels for ICP combinations stratifying NSCLC patients. a-c Summarizing graph of P-values generated from K-M survival analyses (left), of ICP alone, and in combination with each other and with CD3 TILs, where significance indicates positive associations of combinations with OS. From top to bottom, panels interrogate combinations of CD3⁺ TILs and ICPs (a) TIM-3, CD26 and CD39, (b) TIM-3, BTLA and LAG-3, and (c) 2B4, PD-1, and CD57. K-M plots (right) illustrate similarities of curves of ICP combinations ± CD3 co-labeling. The number of patients (n) for each group is given on K-M curves, and others are either previously reported in Fig. 3, or are (high and low, respectively): CD26 n = 34hi, 32lo, CD39 n = 32hi, 26lo, CD26-CD39 n = 37hi, 35lo, TIM-3-CD26 n = 35hi, 37lo, TIM-3-LAG-3 n = 36hi, 32lo, TIM-3-BTLA n = 39hi, 31lo, 2B4 n = 31hi, 33lo, CD57 n = 29hi, 34lo, 2B4-CD57 n = 30hi, 35lo, PD-1-CD57 n = 27hi, 38lo, and 2B4-PD-1 n = 24hi, 44lo; associated confidence intervals are listed in Additional file 1: Table S12. Algorithm design, normalization and analyses were performed by two independent operators. Representative images of CD3-ICP colocalization-dense core areas can be found in Additional file 1: Figure S6. hi, high marker expression, lo, low marker expression; P, Log-rank test; ns, not significant; * P < 0.05; ** P < 0.01; HR, hazard ratio (Log-rank); CI, confidence interval of ratio

From the demonstration that specific combinations of ICPs could efficiently stratify patients, we performed correlation studies between all ICPs from RNA and TMA datasets to reveal ICP coexpression dynamics (Additional file 1: Table S13). Correlograms showed that for both RNA datasets, a majority of ICPs were highly correlated in expression (Fig. 6a), with the most highly correlating pairs being TIM-3 and LAIR-1, and CTLA-4 and TIGIT. TMA cohort correlograms reveal strongest associations between 2B4 and CD57, and BTLA, TIM-3 and LAG-3; this group conserved across all four datasets, and positively associating with OS.Fig6Fig. 6

RNA and protein conserved ICP coexpression groups ranked for NSCLC patient stratification. a-c Graphs depicting R package generated correlation studies made between all ICPs from RNA and TMA datasets to reveal ICP coexpression dynamics stratifying patients. From left to right, RNA expression of ICPs from the TCGA LUAD (n = 504) and LUSC (n = 494) patient samples (left two graph columns), were compared to that of ICP and CD3-ICP expression from all TMA dataset patient (n = 73) samples (right two graph columns). a Correlograms demonstrating ICP coexpression clustering, where black boxes demarcate most highly correlating ICP. b PCA for visualization of multi-dimensional ICP coexpression, relative to distributions patient data (blue circles), where yellow shaded PC quadrants are occupied by ICP coexpressing groups having positive associations with OS, defined by Additional file 1: Figure S7. c Mean correlations of ICP coexpression demonstrate those most abundantly expressed relative to all other ICPs in NSCLC patients. Analyses were performed using alternative software (see Online Methods) by two independent operators. PC1, principal component 1; PC2, principal component 2

From the observation that ICPs positively associating with OS were increased in expression in tumor samples (Additional file 1: Table S8), we used the Integrated Interaction Database (IID) to identify 1750 key ICP-protein interactions from 40,555 possible interactions between all identified ICP-interacting proteins. Key ICP-interactors were refined for those that were 1) experimentally validated to interact with ICP, 2) redundantly interacting with more than one ICP, 3) associated with OS, and 4) had supporting evidence for their interactions in lung tissues (Additional file 1: Table S15). NAViGaTOR software was used to visualize all ICP-interactors, their characterized molecular functions, and supported interactions in lung tissues; demonstrating that 10 of the 13 signature ICP interacted with each other (Additional file 1: Figure S10, Table S16, and Additional file 4). Interaction networks were expanded to visualize defined groups from refined ICP-interactors (Fig. 7). The majority of ICP-interactors had a positive association with OS (64.6%); most of which also had increased gene expression in tumors (85.4%). The majority of ICPs in these two categories were also those ranking highest in interactions with other ICPs. Both increased in expression in tumors and associated with positive OS, BTLA and TIM-3 were observed to interact with a majority of these proteins (Fig. 7 and Additional file 1: Table S15). The pathDIP portal was used for comprehensive pathway enrichment analyses of ICP-ICP interactions and refined ICP-interactors lists (Fig. 7 and Additional files 3 and Additional file 4), and word-cloud analysis was used to compile the most significant ICP-interactors and associated pathways (Additional file 1: Figure S11). Together, these results demonstrate that most ICP-interactors are increased in expression and are associated to positive outcome, further suggesting that ICPs are positive prognostic NSCLC biomarkers.Fig7Fig. 7

ICP-interacting proteins associated with NSCLC patient survival. Visualization of complete comprehensive and interactive ICP-ICP and ICP-proteins interaction mapping by NAViGaTOR. ICP interactors with (a) increased gene expression in tumors and positive association with OS, (b) decreased gene expression in tumors and positive association with OS, (c) increased gene expression in tumors and negative association with OS, and (d) decreased gene expression in tumors and negative association with OS