Immune checkpoint inhibitors targeting the PD-1/PD-L1 axis lead to durable clinical responses in subsets of cancer patients across multiple indications, including non-small cell lung cancer (NSCLC), urothelial carcinoma (UC) and renal cell carcinoma (RCC). Herein, we complement PD-L1 immunohistochemistry (IHC) and tumor mutation burden (TMB) with RNA-seq in 366 patients to identify unifying and indication-specific molecular profiles that can predict response to checkpoint blockade across these tumor types. Multiple machine learning approaches failed to identify a baseline transcriptional signature highly predictive of response across these indications. Signatures described previously for immune checkpoint inhibitors also failed to validate. At the pathway level, significant heterogeneity is observed between indications, in particular within the PD-L1+ tumors. mUC and NSCLC are molecularly aligned, with cell cycle and DNA damage repair genes associated with response in PD-L1- tumors. At the gene level, the CDK4/6 inhibitor CDKN2A is identified as a significant transcriptional correlate of response, highlighting the association of non-immune pathways to the outcome of checkpoint blockade. This cross-indication analysis reveals molecular heterogeneity between mUC, NSCLC and RCC tumors, suggesting that indication-specific molecular approaches should be prioritized to formulate treatment strategies.

Anti-PD-L1 antibodies benefit many cancer patients, even those with “non-inflamed tumor”. Determining which patients will benefit remains an important clinical goal. In a non-inflamed tumor mouse model, we found that PD-L1 was highly expressed on antigen-presenting cells (APCs) especially on CD103+ CD11c+ dendritic cells in tumor-draining lymph nodes (dLNs), suppressing T-cell priming by APCs. In this model, anti-PD-L1 antibodies enhanced T-cell priming and increased CXCR3+ activated T-cells in dLNs, which was followed by the trafficking of T-cells to tumors in response to CXCR3 ligands. As predictive biomarker, each APCs-related gene expression (AP score) alone or T-cells trafficking-related chemokine gene expression (T score) alone were still less than perfect among the 17 mouse models examined. However a combining score of AP score and T score (AP/T score) precisely identified anti-PD-L1-sensitive tumors. In the phase 3 trial of atezolizumab vs docetaxel in advanced NSCLC patients (OAK), the AP/T score could identify atezolizumab-treated NSCLC patients who achieved significant improvement in overall survival. This biomarker concept would be a clinically valuable for prediction of anti-PD-L1 antibody efficacy.

While blood gene signatures have shown promise in tuberculosis (TB) diagnosis and treatment monitoring, most signatures derived from a single cohort may be insufficient to capture TB heterogeneity in populations and individuals. Here we report a new generalized approach combining a network-based meta-analysis with machine-learning modeling to leverage the power of heterogeneity among studies. The transcriptome datasets from 57 studies (37 TB and 20 viral infections) across demographics and TB disease states were used for gene signature discovery and model training and validation. The network-based meta-analysis identified a common 45-gene signature specific to active TB disease across studies. Two optimized random forest regression models, using the full or partial 45-gene signature, were then established to model the continuum from Mycobacterium tuberculosis infection to disease and treatment response. In model validation, using pooled multi-cohort datasets to mimic the real-world setting, the model provides robust predictive performance for incipient to active TB risk over a 2.5-year period with an AUROC of 0.85, 74.2% sensitivity, and 78.3% specificity, which approximates the minimum criteria (>75% sensitivity and >75% specificity) within the WHO target product profile for prediction of progression to TB. Moreover, the model strongly discriminates active TB from viral infection (AUROC 0.93, 95% CI 0.91–0.94). For treatment monitoring, the TB scores generated by the model statistically correlate with treatment responses over time and were predictive, even before treatment initiation, of standard treatment clinical outcomes. We demonstrate an end-to-end gene signature model development scheme that considers heterogeneity for TB risk estimation and treatment monitoring.