Background Anti-PD-1 therapies have achieved durable clinical responses in a wide range of malignancies, but responses are limited to a small subset of patients. Expression of PD-L1 on tumor cells by immunohistochemistry (IHC) has been applied as a companion diagnostic for anti-PD-1 therapy. However, recent studies have called in to question the reliability of this method to predict response.

Methods Here we developed a novel platform that integrates in vitro pharmacogenomic and functional data with clinical pharmacodynamic responses to immunotherapy using proprietary in silico approaches. The data originate from a long-term co-culture of primary antigen-specific T cells and cancer cells which drives T cells to a terminally dysfunctional, PD-1 refractory state. T cell effector functions and gene expression changes were monitored in the presence or absence of anti-PD-1 antibody or genetic knockouts. RNA expression signatures were refined with a randomized sliding window approach to generate a deep learning neural network for PD-1 response prediction.

Results We defined five T cell states associated with distinct phenotypic and molecular features - naïve, active, effector, transition and dysfunction. Among the genes that were selectively expressed in the dysfunction state, we identified a 96-gene signature that is closely associated with clinical outcomes to anti-PD-1 therapy. In PD-1 treated patients across multiple solid tumor indications, this signature correlates with objective response rate and outperforms traditional metrics such as tumor mutation burden or PD-L1 IHC signal. Moreover, this signature combines with tumor sequencing data to generate a powerful machine-learning model that predicts anti-PD-1 responses in metastatic melanoma patients with significantly higher accuracy than PD-L1 IHC. Having established that the T cell states in our co-culture relate to clinical outcomes, we leveraged the system to investigate the molecular basis for PD-1 responses. Single cell mapping of transition state T cells in the presence of anti-PD-1 revealed an expanded population of T cells that co-expresses PD-1, TIGIT and activation markers. Likewise, PD-L1 knockout on cancer cells identified the TIGIT ligand, CD155, as a potential tumor escape mechanism to anti-PD-1 therapy. Consistent with this, the combination of PD-1 and TIGIT blockade enhanced T cell cytotoxicity of tumor cells relative to monotherapies.

Conclusions Agenus’ T cell dysfunction platform combines deep in vitro profiling and AI-based approaches to predict clinical outcomes. Here, we defined a predictive biomarker signature that outperforms standard PD-L1 IHC. Further, we identified known (TIGIT) and potentially novel combination partners predicted to enhance the durability of anti-PD-1 responses.

Ethics Approval Not Applicable

Consent Not Applicable