Development and validation of the Immune Profile Score (IPS), a novel multiomic algorithmic assay for stratifying outcomes in a real-world cohort of patients with advanced solid cancer treated with immune checkpoint inhibitors

Abstract

Background: Immune checkpoint inhibitors (ICIs) have transformed the oncology treatment landscape. Despite substantial improvements for some patients, the majority do not benefit from ICIs, indicating a need for predictive biomarkers to better inform treatment decisions.

Methods: A de-identified pan-cancer cohort from the Tempus multimodal real-world database was used for the development and validation of the Immune Profile Score (IPS) algorithm leveraging Tempus xT (648 gene DNA panel) and xR (RNA sequencing) (N=1,707 development cohort; N=1,600 validation cohort). The cohort consisted of patients with advanced stage cancer with solid tumor carcinomas across 16 cancer types treated with any ICI-containing regimen as the first or second line of therapy. The IPS model was developed using a machine learning framework that includes tumor mutational burden (TMB) and 11 RNA-based biomarkers as features.

Results: IPS-High patients demonstrated significantly longer overall survival (OS) compared with IPS-Low patients (HR=0.45, 90% CI (0.40 to 0.52)). IPS was consistently prognostic in programmed death-ligand 1 (PD-L1) (positive/negative), TMB (High/Low), microsatellite status (microsatellite instability (MSI)-High), and regimen (ICI only/ICI+other) subgroups. Additionally, IPS remained significant in multivariable models controlling for TMB, MSI, and PD-L1, with IPS HRs of 0.49 (90% CI 0.42 to 0.56), 0.47 (90% CI 0.41 to 0.53), and 0.45 (90% CI 0.38 to 0.53), respectively. In an exploratory predictive utility analysis of the subset of patients (n=345) receiving first-line chemotherapy (CT) and second-line ICI, there was no significant effect of IPS for time to next treatment on CT in L1 (HR=1.06 (90% CI 0.88 to 1.29)). However, there was a significant effect of IPS for OS on ICI in L2 (HR=0.63 (90% CI 0.49 to 0.82)). A test of interaction was statistically significant (p<0.01).

Conclusions: Our results demonstrate that IPS is a generalizable multiomic biomarker that can be widely used clinically as a prognosticator of ICI-based regimens.

Keywords: Biomarker; Gene expression profiling - GEP; Immune Checkpoint Inhibitor; Next generation sequencing - NGS; Tumor microenvironment - TME.

Figure 1. Study overview. (A) A table showing the biological roles corresponding to each feature from the final IPS algorithm. The IPS features are sorted by feature importance within each biological role. (B) Inclusion/exclusion criteria for the training and validation study cohorts. ECOG PS, ECOG performance status; ICI, immune checkpoint inhibitor; IPS, Immune Profile Score; TMB, tumor mutational burden; 1L, first-line; 2L, second-line.

Figure 2. The HR was 0.45 (90% CI 0.40 to 0.52), p<0.01. Predicted OS from the Cox PH model for (A) 1L monotherapy and (B) 2L monotherapy patients. Predicted survival for 1L and 2L combination therapy patients are similar to above. (C) Predicted OS curves from the Cox PH model in 1L combination therapy (shown in A). (D) Predicted OS curves from the Cox PH model in 2L combination therapy patients (shown in B). (E) The median OS and 95% CI for IPS-H and IPS-L groups for each line of therapy/treatment group combination. IPS, Immune Profile Score; IPS-H, IPS-High; IPS-L, IPS-Low; OS, overall survival; 1L, first-line; 2L, second-line.

Figure 3. Forest plot showing IPS-High versus IPS-Low HRs and CIs across demographics and clinically relevant subgroups. Subgroups may have <1,519 patients due to availability of data. CRC, colorectal cancer; HNSCC, H=head and neck squamous cell carcinoma; ICI, immune checkpoint inhibitor; IPS, Immune Profile Score; MSI, microsatellite instability; MSS, microsatellite stable; NSCLC, non-small cell lung cancer; PD-L1, programmed death-ligand 1; RCC, renal cell carcinoma; TMB, tumor mutational burden; 1L, first-line.

Figure 4. Overall survival is significantly higher in IPS-H versus IPS-L. (A) Forest plot showing univariate (UV) HRs for TMB, PD-L1, MSI and multivariate (MV) HRs that include IPS. A likelihood ratio test between the UV and MV models was significant (p<0.01) for all three biomarkers, indicating that IPS has significant prognostic utility beyond TMB, MSI, and PD-L1. Plots (B–E) show predicted OS from a model stratified by line of therapy and fit on IPS, treatment group, and the MV model with the listed biomarker: (B) TMB pan-cancer, (C) MSI pan-cancer, (D) PD-L1 pan-cancer and (E) PD-L1 in NSCLC patients. The predicted OS curves represent patients treated with monotherapy in 1L for TMB and MSI (B–C), and combination therapy in 1L for PD-L1 and NSCLC (D, E). (F) HR and 90% CI for the most relevant curves shown in the predicted OS plots in (B–E). (G) Predicted OS curves from Cox PH model for NSCLC ICI+combination 1L cohort stratified by IPS result and PD-L1 IHC staining level. PD-L1 ultra high: TPS>90, PD-L1 high: TPS=89–50, PD-L1 low: TPS=49–1, PD-L1 negative: TPS=0. IHC, immunohistochemistry; IPS, Immune Profile Score; IPS-H, IPS-High; IPS-L, IPS-Low; MSI, microsatellite instability; MSS, microsatellite stable; NSCLC, non-small cell lung cancer; OS, overall survival; PD-L1, programmed death-ligand 1; TMB, tumor mutational burden; TMB-H, TMB-High; TMB-L, TMB-Low; 1L, first-line; TPS, tumor proportion score.

Figure 5. (A) Predicted TTNT for 1L CT with no significant effect for IPS (HR=1.06 (90% CI 0.88 to 1.29)). (A) Predicted OS for 2L ICI shows that IPS does have a significant effect (HR=0.63 (90% CI 0.49 to 0.82)). Interaction test p<0.01, indicating that the HR in 2L ICI is significantly different from HR in 1L CT. (C) Kaplan-Meier OS curves for stage IV TCGA patients stratified by IPS result generated from TCGA DNA and RNA data processed through Tempus bioinformatic pipelines. CT, chemotherapy; ICI, immune checkpoint inhibitor; IO, immunotherapy; IPS, Immune Profile Score; IPS-H, IPS-High; IPS-L, IPS-Low; OS, overall survival; TCGA, The Cancer Genome Atlas; 1L, first-line; 2L, second-line; TTNT, time to next treatment.

Figure 6. (A) Prevalence plot showing the percentage of IPS-H patients in a large, representative cohort of Tempus clinical patients. (B) Prevalence plot showing the percentage of IPS-H and IPS-L patients compared with percentage of TMB-H/L patients in a large, representative cohort of patients from the Tempus multimodal database. IPS, Immune Profile Score; IPS-H, IPS-High; IPS-L, IPS-Low; SCC, squamous cell carcinoma; TMB, tumor mutational burden; TMB-H, TMB-High; TMB-L, TMB-Low.