Answers are grounded in this paper and cite their sources.
Chimeric antigen receptor T (CAR-T) cell therapy has transformed the treatment of hematological malignancies, but its translation to solid tumors remains limited by antigen heterogeneity, poor tumor infiltration, immunosuppressive tumor microenvironments, and on-target/off-tumor toxicity. Here, we present a safety-constrained artificial intelligence framework for rational CAR-T design in solid tumors. The workflow integrates four linked stages: CD-antigen prioritization, safety-by-design CAR architecture mapping, constrained large language model (LLM)–assisted reasoning, and statistical feature-structure analysis. Curated CD-antigen functional annotations and tumor-associated expression descriptors were encoded into a multi-criteria scoring system that stratified targets into high-confidence, conditional, and unsafe classes. Across the analyzed CD-antigen space, 38% of markers were classified as high-confidence candidates, 34% as conditional targets requiring logic-gated or controllable architectures, and 28% as unsafe targets because of predicted off-tumor risk. External benchmark analysis using MSLN, HER2/ERBB2, EGFR/EGFRvIII, CEA/CEACAM5, GD2, B7-H3/CD276, and CD47 supported the framework’s ability to distinguish direct targets from safety-constrained or unsuitable candidates. An updated four-model LLM benchmark, including GPT-5.6 Sol, Gemini 3.5, Grok 4.5, and Claude Sonnet 5, showed convergence toward dual-antigen logic-gated, trafficking-enhanced, and armored CAR designs, while revealing model-dependent differences in safety awareness, traceability, and hallucination risk. Claude Sonnet 5 and GPT-5.6 Sol showed the strongest safety-oriented performance, whereas Grok 4.5 required greater expert oversight. Correlation and clustering analyses further identified redundant safety-liability features and candidate antigen combinations for logic-gated CAR design. This study provides a reproducible, human-in-the-loop decision-support framework for generating safety-aware CAR-T hypotheses. The framework is computational and hypothesis-generating and does not constitute experimental validation of CAR-T efficacy or safety.
No takes yet. Share an insight, caveat, or question.
Taştan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: