ScRNA data from healthy donors, as well as scRNA + VDJ and experimentally derived tumor-reactivity data for the 50 most frequent TCR TIL clonotypes from sample BT21, were used to train an intermediate model using XGBoost. Due to the sparse nature of scRNA data, we optimized this intermediate model by first performing Bayesian optimization to tune hyperparameters with stratified k -fold cross-validation. Subsequently we identified the top features (that is, genes) in this intermediate model using explainable AI SHAP, and then trained a simpler model using only these features to prevent overfitting to the training data. This simpler model was retuned as before and then applied to the remaining BT21 TIL data. Per-cell reactivity probabilities calculated by the classifier were averaged for each TCR clonotype, and the Fisher-Jenk natural break was used to determine the appropriate minimum threshold for calling TCRs as tumor reactive.
