Artificial Intelligence-Guided Design of CAR-T Cells for Solid Tumors through CD-Antigen Prioritization, Safety-by-Design Architectures, and Constrained Large Language Model Reasoning

Abstract

Chimeric antigen receptor T (CAR-T) cell therapy has achieved transformative success in hematological malignancies, yet its extension to solid tumors remains constrained by antigen heterogeneity, limited tumor infiltration, immunosuppressive tumor microenvironments, and on-target/off-tumor toxicity. Addressing these challenges requires integrative design strategies that simultaneously optimize antigen selection, receptor architecture, and safety constraints. Here, we present an artificial intelligence (AI)-guided, end-to-end computational framework for rational CAR-T design in solid tumors that unifies CD-antigen prioritization, safety-by-design CAR architectures, constrained large language model (LLM)-assisted evidence synthesis, and statistical feature validation. The framework operates across four sequential stages. First, structured CD-antigen knowledge, immune functional annotations, surface topology, and tumor-associated expression descriptors are integrated into a multi-criteria antigen prioritization scheme, enabling stratification of targets into high-confidence, conditional, and unsafe classes. Second, prioritized antigens are algorithmically mapped to safety-constrained CAR architectures, including logic-gated, modular, and armored designs that explicitly mitigate off-tumor toxicity and functional exhaustion. Third, multiple LLMs are benchmarked under standardized prompts and quantitative scoring rubrics to evaluate architectural convergence, safety awareness, factual grounding, and translational feasibility. Fourth, correlation analysis and unsupervised clustering of antigen feature spaces identify functional redundancy and synergistic antigen combinations, directly informing rational multi-antigen CAR-T designs. Applying this framework reveals strong convergence of LLMs toward dual-antigen logic-gated and trafficking-enhanced CAR architectures for solid tumors, while also uncovering substantial model-dependent variability in safety rigor and hallucination risk. Statistical validation demonstrates that antigen suitability is an emergent, context-dependent property shaped by feature interactions rather than expression alone. By embedding AI within explicit biological constraints, quantitative validation, and human oversight, this work establishes a reproducible, auditable blueprint for translating heterogeneous biomarker data into safety-aware, experimentally testable CAR-T designs for solid tumors.