Quantitative structure-activity and structure-property relationship (QSAR/QSPR) modelling uses molecular descriptors to relate chemical structure to physicochemical behaviour and biological activity. In this study, we propose a new family of descriptors, termed sum-connectivity descriptor, obtained by multiplying a classical edge-degree kernel \(\Phi(d_u,d_v)\) by the normalization factor \((d_u+d_v)^{-1/2}\) . This construction generates a systematic family of descriptors induced from Zagreb-type, Sombor-type, Albertson, arithmetic-geometric, geometric-arithmetic, Forgotten, and inverse Nirmala indices. From a theoretical perspective, we establish comparison relations with the corresponding classical descriptors, degree-extremal bounds, and regular-graph proportionality results. The empirical study uses a curated antibacterial E. coli dataset from ChEMBL containing 6,657 unique compounds, with MIC values standardized to \(\mu\) M and transformed to \(pMIC\) . In the QSPR benchmark, the sum-connectivity indices consistently outperformed the classical indices, increasing the mean \(R^2\) from \(0.6951\) to \(0.7015\) . Moreover, the best QSPR performance was obtained by the All Combined descriptor set, which integrates indices, RDKit descriptors, and physicochemical properties, with properties-specific \(R^2\) values ranging from \(0.9349\) to \(0.9997\) . For antibacterial activity, the best model is ExtraTrees + RDKit Descriptors, with \(R^2=0.5959\pm0.0246\) under repeated cross-validation and \(R^2=0.4799\pm0.0252\) under Murcko scaffold validation. The proposed sum-connectivity descriptors are therefore most strongly supported as systematically improved graph-topological descriptors in matched QSPR comparisons, while their added value for the more heterogeneous \(pMIC\) endpoint is complementary and modest.
