Radiomic and immunophenotypic profiling to decode tumor heterogeneity in the era of immunotherapy: from preclinical evidence to clinical practice

Background: The ground-breaking results of immune checkpoint inhibitors (ICIs) in solid tumors still leave uncovered multiple drawbacks on the clinical translation of cancer-immune interplay, representing a unique window of opportunity for both preclinical and clinical research. Several prognostic and predictive biomarkers of ICI benefit have been proposed, ranging from cancer-intrinsic molecular cues to immune microenvironmental features or advanced imaging, although without reaching robust consensus when tested or validated in clinical setting. By converting imaging data into high-throughput features, radiomics emerged as a non-invasive tool able to decode the tumor immune microenvironment (TIME). Nonetheless, lack of evidence of causal relationship coupled to methodological issues limit its clinical application. Furthermore, scarce radiomic data in preclinical settings have been generated so far. Similarly, the role of liquid biopsy as a source of immunobiologically and clinically relevant information and a potential surrogate of tissue samples, is still an area of intense investigation. Research hypothesis: Hence, we aimed at exploring the role of radiomic and blood immune-inflammatory signatures as non-invasive tools able to dissect tumor heterogeneity and predict ICI benefit, from both preclinical and clinical perspectives. Study design and methodology: The present work consisted of two interdependent tasks: - Task 1 (preclinical): our analyses were addressed to characterize the TIME of different murine models and evaluate whether micro–Computed Tomography (μ-CT) radiomic features entail specific immunophenotypic profiles. To this end, orthotopic head and neck (H&N) and subcutaneous tumor models were established. TIME was selectively modulated to obtain distinct immune contexts, and immunohistochemical (IHC) staining for CD8+ T cells was computed. In vivo preclinical imaging was performed using the Quantum FX µ-CT technology (PerkinElmer); µ-CT scans were explored using the pyradiomics python library for feature extraction. Appropriate statistical analyses were conducted to assess the robustness of radiomic signatures in predicting immune population density. - Task 2 (clinical): high-throughput imaging and immunophenotypic analyses were performed on CT scans, tissue biopsies and peripheral blood (PB) samples to determine whether a radiomic and blood-based approach might non-invasively intercept tumor heterogeneity and treatment outcome in patients affected by non small cell lung cancer (NSCLC). Two different cohorts of NSCLC patients were included: 1. stage I-IIIA surgically resected NSCLC; 2. ICI-treated metastatic NSCLC. Digital microscopy was employed on tissue samples, to assess the density and spatial distribution of different Tumor-Infiltrating Lymphocytes (TILs) subpopulations (CD3+, CD4+, CD8+, PD-1+). PB was analyzed by FACS (Fluorescence Activated Cell Sorting) to quantify the circulating immunophenotypes and by ELISA-based multi-cytokines assay to evaluate soluble mediators. Radiomic features (RFs) were extracted from CT scans using the open-source tool pyRadiomics, following the delineation of the volume of interest (VOI). Statistical analyses were conducted to assess potential correlations between TIME and blood immune descriptors, radiomic parameters and patient outcome. Results: - Evidence from the preclinical task provided definitive proof of the ability of radiomics to assess the presence of functionally active tumor infiltrating lymphocytes in different murine models. Specifically, we first developed a highly efficient radiomic pipeline in preclinical models of H&N and subcutaneous cancers. Then, we demonstrated the feasibility and accuracy of radiomics to detect dramatic changes in T cells within the TIME, identifying differentially regulated radiomic features between T cell enriched and T cell depleted tumors (Mann-Whitney test, P < 0.05), thus offering radio-immune signatures potentially translatable into clinical practice. - These preclinical findings were in line with our clinical results which strengthened the relevance of radiomics-based multivariate classifiers to decipher distinctive TIME characteristics in early-stage NSCLC patients. In detail, we generated 7 machine learning-based radiomic models able to efficiently predict specific TIME characteristics (i.e. PD-L1 levels of expression, TILs subpopulation, hot and cold TIME categories). Accuracy, precision, sensitivity and ROC AUC values ranged from 70% to 80% for all models, while specificity resulted slightly lower. Selected radiomic features varied depending on the endpoint considered, except for Peri_wavelet_HLL_glcm_Correlation which was shared by four models. - Pursuing non-invasive approaches to dissect the cancer-immune interplay and its impact on treatment outcome in metastatic setting, we documented significant correlations between tissue and circulating immune compartments, reinforcing the role of blood as a source of clinically and immunobiologically relevant information. In particular, a significant direct correlation between tissue and PB CD3+ and CD8+ T cells was observed (Spearman Rho coefficients ranging from 0.29 to 0.45, P < 0.05), together with meaningful associations of serum cytokines (IL-62nd gen, IL-1β, IL-2, IFN-γ3rd gen, IL-12p70) with CD8/CD3 and PD-1/CD8 ratios at tissue level (Spearman Rho coefficients ranging from 0.36 to 0.46, P < 0.05). - Finally, our data on acquired resistance (AR) to ICIs in advanced NSCLC demonstrated that dynamic blood immune-inflammatory and radiomic profiling may early identify the occurrence of specific AR patterns (oligo- vs systemic). Indeed, Δpos cytotoxic (NK, CD8+GnzB+) and Δneg immunosuppressive (CD14+ monocytes) dynamic coupled with different modulation of IL-6, TGF-β1, TNFα and sPD-L1 represented distinctive features of oligoAR patients (P < 0.05), which also displayed significantly longer post-progression survival compared to systemicAR (median 20.3 vs 5.6 months; HR:0.22, P < 0.001). Furthermore, delta radiomics outperformed baseline RFs, with 15 ΔRFs sharply discriminating oligoAR from sysAR (P range: < 0.001-0.04). ROC analysis confirmed the optimal performance of top-ranked ΔRFs (AUC range: 0.88-0.99). Conclusion: We identified radio-immune signatures faithfully reflecting tumor imaging-bioarchitecture correlates in preclinical models, potentially translatable into clinical practice. We successfully chased multiparametric approaches integrating radiomic and immunophenotypic cues, thus offering an effective non-invasive tool to guide clinical decision making in early-stage and metastatic ICI-treated NSCLC.