Objective Limited sample sizes in natural anti-aging compound datasets, class imbalance, and redundancy among high-dimensional molecular descriptors can restrict model performance in virtual screening. These problems make it difficult to identify informative molecular features and reliably recognize minority-class active compounds. This combination of data scarcity and imbalance increases the difficulty of establishing robust decision boundaries for compound screening. This study aimed to develop a tabular learning method tailored to small-sample and class-imbalanced settings for virtual screening of natural anti-aging compounds, with emphasis on improving active-compound recognition and candidate ranking.Methods Sq-TabPFN was developed by introducing sequential attention into the Tabular Prior-data Fitted Network (TabPFN). Across successive decision steps, the attention module generates instance-specific sparse masks that emphasize informative molecular descriptors while progressively suppressing redundant features. Thus, feature selection is integrated into model learning rather than performed as a fixed preprocessing step. To further address class imbalance, a strategy termed Wasserstein-Based Dual Secure Sampling (WBDSS) was proposed. WBDSS first removes potentially noisy majority-class samples near the decision boundary and identifies reliable minority-class samples according to model confidence. Representative root samples are then selected from the safe minority subset, and their neighboring samples are weighted according to Wasserstein distance to guide minority-class interpolation. The resulting rebalanced training set is used to fine-tune Sq-TabPFN, allowing adaptive feature selection and distribution-aware resampling to jointly strengthen minority-class learning while reducing the influence of noisy or unreliable samples.Results WBDSS-optimized Sq-TabPFN achieved the best overall screening performance among the compared models. Across 10 runs with different random seeds, its mean F1 score and PR_AUC were 0.557 and 0.792, respectively. Compared with TabPFN, F1 and PR_AUC increased by 19.5% and 21.1%, respectively, indicating improved recognition and ranking of active compounds under class imbalance. Statistical testing indicated that the improvement in PR_AUC was unlikely to result from random variation. Under scaffold-based splitting, F1 and PR_AUC were 0.54 and 0.79, respectively, showing that the model retained favorable predictive performance when structural overlap between the training and test sets was restricted. Experiments on two benchmark molecular classification datasets further supported the cross-task applicability of Sq-TabPFN. The optimized model was subsequently applied to the CMAUP natural product database for virtual screening and identified 10 candidate compounds. Five candidates were supported by published evidence or plausible aging-related mechanisms. Molecular docking analysis suggested potential binding between the candidate compounds and targets associated with anti-aging effects. These results support complementary roles for sequential attention in feature utilization and WBDSS in minority-class representation.Conclusion By integrating sequential-attention-based feature selection with WBDSS-based data rebalancing, the proposed method simultaneously addresses descriptor redundancy and class imbalance in small-sample molecular screening. WBDSS-optimized Sq-TabPFN improves the recognition and ranking of potentially active compounds and provides an effective computational approach for virtual screening of natural anti-aging compounds. The method can help narrow the candidate space and support the prioritization of compounds for subsequent experimental investigation. The screening and docking results are computational predictions rather than direct evidence of anti-aging activity and therefore require further biological validation.