Anopheles sinensis is an important malaria vector in China. Clarifying the driving mechanism of meteorological factors on its population density has practical significance for consolidating the achievements of malaria elimination. Objective: to compare the differences between two models, the geographically weighted random forest (GWRF) combined with SHAP interpretability analysis and the geodetector, in identifying the driving factors of vector mosquitoes, and to analyze the effects of multiple meteorological factors, time-lag effects and their interactions on the density of Anopheles sinensis . Methods: based on public datasets of mosquito density and meteorology, 39 meteorological variables covering the current month, 1-month lag and 2-month lag were included, and GWRF‑SHAP and the geodetector were used respectively to carry out single-factor contribution and factor interaction effect analysis, and the outputs of the two models were compared. Results: relative humidity, precipitation, wind speed and temperature at 1-month lag are the key meteorological factors affecting the density of Anopheles sinensis , and the meteorological effects have a time lag of 1‑2 months. The primary driving factor identified by GWRF is relative humidity at 1-month lag, and the strong interaction effects are concentrated in the temperature‑temperature and temperature‑pressure variable combinations at 2-month lag; the geodetector takes precipitation days at 1-month lag as the primary factor, and the strong interactions are mostly short-term variable combinations of humidity with temperature, pressure and wind speed. The difference between the results of the two models comes from the different algorithm principles: GWRF focuses on the joint contribution of variables at the prediction level, and the geodetector focuses on the explanatory power of factors for spatial stratified differentiation. Conclusion: a single model will bring biased results, and a method of combining multiple models will be considered later, so as to provide methodological reference for zonal early warning and refined control of malaria vectors.

Natural killer (NK) cells sit at the center of tumor immune surveillance. Their contribution goes beyond direct killing—they produce cytokines, release chemokines, and help recruit dendritic cells. Yet solid tumors often contain high local levels of prostaglandin E2 (PGE₂) , generated by cyclooxygenase-2 (COX-2) and microsomal prostaglandin E synthase-1 (mPGES-1) , and this PGE₂-rich environment dampens NK-cell responses at several points. This review examines how tumor-derived PGE₂ signals through EP2/EP4 (E-prostanoid receptors 2 and 4) and the downstream cAMP (cyclic adenosine monophosphate)/PKA (protein kinase A) cascade. The consequences include altered NK-cell activation, cytokine output, receptor expression, immune-synapse organization, cytotoxicity, migration, and antimetastatic activity. We also compare ways to interrupt this axis: pharmacological inhibitors of PGE₂ synthesis, selective or dual EP2/EP4 blockade, and cell-intrinsic engineering approaches aimed at making NK cells less sensitive to PGE₂. Attention is paid to inconsistencies among experimental models and to whether interventions restore NK-cell function directly or act more broadly on the tumor. Biomarkers that might support patient selection are discussed as well. Current evidence suggests the PGE₂–NK-cell axis is a druggable immunosuppressive circuit, although clinical translation will likely require pathway-based stratification and prospective pharmacodynamic assessment.

Artificial intelligence (AI) is reshaping drug discovery by shifting molecular design from empirical high-throughput screening toward data- and knowledge-guided workflows. This review summarizes the applications of graph neural networks, variational autoencoders, generative adversarial networks, biology-guided deep learning, and information-theoretic predictive models in vaccine development and cancer immunotherapy. Key tasks include molecular representation, de novo generation, virtual screening, ADMET prediction, antigen structure optimization, small-molecule adjuvant discovery, tumor microenvironment characterization, and precision dosing. Representative applications involving TLR/STING-related adjuvants, lipid nanoparticle delivery systems, and immunotherapy-response prediction are discussed to illustrate the complementarity of these computational approaches. Although AI can accelerate early-stage candidate identification and expand searchable chemical space, its clinical impact remains constrained by heterogeneous datasets, limited labeled samples, insufficient interpretability, weak cross-population generalization, and incomplete regulatory standards. Future research should emphasize standardized multimodal data, mechanism-aware model design, prospective validation, privacy-preserving collaboration, and closed-loop integration of computation, experimentation, and clinical feedback.