Defining therapeutic strategies for boosting T-cell infiltration into cold tumors with spatial proteomics and machine learning

Project Summary Immunotherapies such as immune checkpoint inhibitors and chimeric antigen receptor (CAR-T) cell therapy have been highly successful in reversing cancer progression in a subset of patients. However, immunotherapies fail in patients with “cold tumors,” where T-cell infiltration and function are suppressed by inhibitory signaling

environments generated by cancer and stromal cells. Poor CD8+ T-cell infiltration due to suppressive signaling environments is a primary obstacle to effective immunotherapy in many solid tumors including breast, liver, prostate, and colon cancer. Recent advances in high-resolution molecular imaging technologies, known as

spatial proteomic methods, now allow micron-resolution profiling of signaling environments in cold and hot human tumors across up to 50 molecular channels providing a new data source for identifying signaling cues that promote or suppress T-cell infiltration. There is an urgent unmet need for computational strategies that can

analyze large-scale, spatial proteomic data sets collected from human patient data to identify features of the tumor microenvironment that promote cold vs hot tumor phenotypes. Computational methods must be designed to extract concrete and specific therapeutic strategies that can be tested clinically for reprogramming the tumor

microenvironment to promote T-cell infiltration and function. In this project, we develop a machine learning framework that uses cutting-edge spatial proteomic data to identify signaling molecules and guidance cues that promote the infiltration and function of T-cells into a tumor microenvironment. Our approach first trains a neural

network on spatial proteomic data to predict T-cell infiltration using signaling and guidance cues. We, then, apply “counterfactual reasoning” to the classifier to predict optimal signaling perturbations for increasing CD8 T-cell infiltration into tumors. In preliminary data, we applied our strategy to melanoma and identified a therapeutic

strategy that involves manipulation of five chemokine and signaling molecules in melanoma based on spatial proteomic data from 300 patients. In the work to be performed here, we aim to generalize our approach to a broader range of cancer types and larger patient data sets. We will systematically test neural network

architectures to identify optimal architectures for different cancer types. Since spatial proteomic training data is currently limited, we will collect new training data from human patients across a broader set of tumors, for which we will profile chemokine and signaling molecules through a collaboration between Cedars-Sinai Medical Center

and Caltech. We will generalize our counterfactual reasoning strategy to breast and prostate cancer to identify optimal therapeutic targets and to compare targets for different base tumor types. Broadly, our work will develop a novel machine learning approach for converting large-scale spatial proteomic data into specific molecular

hypotheses for increasing T-cell infiltration into cold tumors across a range of solid tumor types.