Abstract:
CAR T cell therapy has emerged as a promising approach for treating various forms of
cancer. Despite its success, the effectiveness of CAR T cell therapy can be influenced
by the specific configurations of signaling motifs within the CAR architecture. This
study focuses on optimizing CAR architecture using strategic motif combinations and
predictive modeling to enhance therapeutic outcomes. The research utilized a dataset
of CAR T cell configurations, each characterized by different motif combinations, to
explore the impact on cytotoxicity and stemness. Various machine learning models,
including Random Forest, Support Vector Machine, Neural Networks, Linear
Regression, Decision Tree, and Gradient Boosting, were employed to predict
cytotoxicity and stemness based on these configurations. The transformer-based model
was also implemented to predict cytotoxicity from protein sequences, showcasing the
potential for integrating generative AI models into the healthcare domain. The results
demonstrated that traditional machine learning models, such as Decision Tree and
Gradient Boosting, effectively captured key features of cytotoxicity and stemness,
particularly at moderate levels, and did so with significantly less computational power
and time compared to more complex models like CNNs and LSTMs. This highlights
the first major objective of the research: to show that machine learning models can
achieve comparable performance to CNNs and LSTMs in feature extraction, while
being more efficient in terms of computational resources. This study contributes to the
growing field of CAR T cell therapy by providing a detailed analysis of how different
motif combinations influence therapeutic outcomes. The research offers key insights
that can lead to the refinement of CAR designs, making them more effective and safer.
These findings support the development of improved CAR T cell therapies and pave
the way for personalized treatment strategies that use predictive modeling to tailor
interventions to individual patients.
