Search for: All records

Pep-TCRNet is a novel approach to constructing a prediction model that can evaluate the probability of recognition between a TCR and a peptide amino acid sequence while combining inputs such as TCR sequences, HLA types, and VJ genes.Pep-TCRNet operates in two key steps:Feature Engineering: This step processes different types of variables:TCR and peptide amino acid sequencing data: The model incorporates neural network architectures inspired by language representation models and graph representation model to learn the meaningful embeddings.Categorical data: Specialized encoding techniques are used to ensure optimal feature representation for HLA types and VJ genes.Prediction Model: The second step involves training a prediction model to evaluate the likelihood of a TCR recognizing a specific peptide, based on the features generated in the first step. 

Exploring features associated with the clinical outcome of interest is a rapidly advancing area of research. However, with contemporary sequencing technologies capable of identifying over thousands of genes per sample, there is a challenge in constructing efficient prediction models that balance accuracy and resource utilization. To address this challenge, researchers have developed feature selection methods to enhance performance, reduce overfitting, and ensure resource efficiency. However, applying feature selection models to survival analysis, particularly in clinical datasets characterized by substantial censoring and limited sample sizes, introduces unique challenges. We propose a robust ensemble feature selection approach integrated with group Lasso to identify compelling features and evaluate its performance in predicting survival outcomes. Our approach consistently outperforms established models across various criteria through extensive simulations, demonstrating low false discovery rates, high sensitivity, and high stability. Furthermore, we applied the approach to a colorectal cancer dataset from The Cancer Genome Atlas, showcasing its effectiveness by generating a composite score based on the selected genes to correctly distinguish different subtypes of the patients. In summary, our proposed approach excels in selecting impactful features from high-dimensional data, yielding better outcomes compared to contemporary state-of-the-art models.