A roadmap for building predictive models from the T cell receptor repertoire feature analysis

Abstract

Although cancer therapy provides a vast repertoire of medicines and treatments, many cancers develop ways to escape and continue to proliferate. Immunotherapy, in particular, has proved efficient in destroying some types of cancers, but it is not an infallible option. Predicting the efficiency of each treatment option would be a valuable tool for the decision-making process in clinical practice. Immunotherapy enhances the patient’s T cells to attack cancer cells. T cells use a receptor protein from their surface to identify possible targets, such as cancer cells. The advent of NGS (Next Generation Sequencing) brought considerable speed to sequencing large amounts of genetic material, such as TCR (T Cell Receptor). The diversity of receptors is colossal, up to 1019 unique receptors, and understanding these highly complex repertoires might be the key to deciphering the immune system’s behavior. Here, we evaluated the process of extracting meaningful features of TCR repertoire data to build predictive models to distinguish healthy controls from cancer patients or patients treated with different drugs. In light of that, it is essential to develop tools that can easily and quickly generate insights from TCR repertoire data to predict future outcomes. We developed a bioinformatic tool called GENTLE (GENerator of T cell receptor repertoire features for machine LEarning), geared towards any researcher working with TCR repertoire data that aims to explore these data and build prediction tools. GENTLE is open-source, has a web platform, can be installed locally, implements many diversity metrics, builds networks using the Levenshtein distance, calculates the frequency of motifs, transforms the data with dimensional reduction methods, implements normalization methods, performs feature selection, builds, evaluates, and deploys classifiers. Using this tool, one can glean great insights from TCR repertoire data.