Abstract: Technetium-99(⁹⁹Tc) is a major long-lived fission product with a half-life of 2.1×10⁵ years, posing persistent radiological risks in nuclear waste management. In oxidizing environments, technetium exists as the pertechnetate anion(\mathrmTcO_4^- ), which exhibits high aqueous mobility due to its low charge density and small hydration energy, making its efficient removal from radioactive wastewater a critical challenge. Metal-organic frameworks(MOFs) have emerged as promising adsorbents for \mathrmTcO_4^- , but experimental data on \mathrmTcO_4^- adsorption remain extremely scarce(fewer than 100 publicly reported records), hindering the development of predictive models. To overcome this small-sample limitation, we propose a machine learning framework that integrates multi-model feature screening, ablation experiments, SHAP interpretability analysis, and symbolic regression, aiming for accurate and fully interpretable prediction of \mathrmTcO_4^- adsorption on MOFs. A source domain dataset of 77 adsorption records is manually curated from open literature, covering nine structurally or chemically related oxyanions(\mathrmReO_4^- , \mathrmCrO_4^2- , \mathrmMnO_4^-, etc.) on various MOFs. From corresponding crystallographic files(CIF), 63 structural descriptors are automatically extracted, encompassing geometry, charge distribution, metal electronic properties, and pore characteristics. Seven classification models(logistic regression, decision tree, random forest, support vector machine, K-nearest neighbor, XGBoost, LightGBM) were used for binary classification tasks, and 18 core features were selected through multi model consensus screening, divided into four categories: geometry, charge, metal electron, and pore. The ablation experiment shows that the removal of metal electrons or charge distribution characteristics lead to the most significant decrease in performance, confirming the dominant role of these two categories. SHAP analysis reveals the direction of influence of key features: metal electronegativity is positively correlated with adsorption capacity, and pore size exhibits an inverted U-shaped relationship, consistent with the hard-soft acid-base theory. To achieve full explainability, the symbolic regression(PySR) to each of the four feature categories is applied, obtaining explicit mathematical classification rules. These four rules are then integrated via a majority-voting ensemble. Although single-category rules suffer from overfitting(validation accuracy 0.43-0.57), the voting ensemble achieves an overall test accuracy of 71.4%, a recall of 100% for high-adsorption samples(4/4), a precision of 67% for high-adsorption predictions. Notably, the model misclassifies only low-adsorption samples as high, which is acceptable in high-throughput screening where false negatives(missing a promising MOF) are more costly. This framework demonstrates that \mathrmTcO_4^- adsorption on MOFs is governed by synergistic contributions from geometry, charge distribution, metal electronic states, and pore architecture. Although validated on \mathrmReO_4^- as a non-radioactive surrogate, the method is expected to generalize to \mathrmTcO_4^- given their high chemical similarity. To our knowledge, this is the first study to combine symbolic regression with a voting ensemble for interpretable prediction of radionuclide adsorption, providing a practical and data-efficient tool for high-throughput screening of MOF-based adsorbents in radiochemistry.