Generative AI in antibody engineering: emerging concepts and translational bottlenecks in lung cancer

The discovery of biotherapeutics is increasingly being driven by generative artificial intelligence (generative AI), marking a paradigm shift from random combinatorial screening to structure-based computational design. Within the field of non-small cell lung cancer (NSCLC), therapeutic development is persistently challenged by biophysical barriers, particularly those associated with neoantigens derived from intracellular oncogenic proteins (e.g., KRAS G12C), conformationally dynamic drug-resistant variants (e.g., EGFR), and structurally complex transmembrane receptors (e.g., CXCR4). Potential pathways for programmable epitope targeting are now being offered by emerging models based on diffusion and multimodal architectures (such as RFdiffusion and Chroma). In this narrative review, the core mechanisms and recent advancements of generative AI in antibody scaffold generation, sequence-structure co-evolution, and the early-stage screening of multi-parameter developability (e.g., solubility, viscosity, and immunogenicity) are evaluated. The utilization of computational methods to address challenging targets is analyzed, and the significant translational gap remaining between algorithmic confidence and clinically relevant biological fidelity is explored. Structural hallucinations, the absence of glycosylation and membrane lipid environment characterizations in training datasets, and binding-function discordance are identified as the primary current technical barriers. Ultimately, the translation of generative AI into clinical oncology will depend on the deep integration of "lab-in-the-loop" platforms, a paradigm where computational physical priors are continuously calibrated by high-throughput experimental data, progressively bridging the gap between in silico design and in vivo efficacy.