Engineered nanoparticles (ENPs), defined as nanoscale materials with at least one dimension between 1 and 100 nm, exhibit multifunctional and tunable physicochemical properties, that are at the center of several innovative fields. However, ENPs may induce a variety of biochemical reactions upon entry into organisms that could be a threat to human health. Therefore, a systematic evaluation of the toxicity of ENPs is essential. Quantitative structure–activity relationship (QSAR) is a practical in vitro modeling approach used to evaluate the toxicity of nanoparticles. In this study, we established the nanometric QSAR (Nano-QSAR) modelling based on cell membrane damage of ENPs to HepaRG cells. The toxicity data of ENPs and related 2D descriptor information were collected from the NanoCommons Knowledge Base. Periodic table descriptors of the elements were calculated using the Elemental Descriptor Calculator software. A multiple linear regression (MLR) model was constructed, and subsequently combined with read-across (RA) descriptors to establish the Nano-quantitative read-across structure–activity relationship (Nano-q-RASAR) model. Furthermore, machine learning (ML) algorithms were applied to optimize the predictive performance of the models. All models were validated according to the stringent OECD QSAR validation guidelines. Finally, a series of true external ENPs without experimental values were autonomously designed, and predicted using the best GB-Nano-QSAR model. Overall, this study can provide efficient and reliable predictions for the cell membrane damage of ENPs and a detailed theoretical explanation of their toxicity mechanism, which is of practical value for the toxicity assessment of ENPs.

Background: Respiratory syncytial virus (RSV) RNA-dependent RNA polymerase (RdRp) complex is an essential molecular machine for viral genome replication. The L protein, the catalytic subunit of this complex (L-RdRp), is well-characterized structurally and represents a highly promising target for the development of novel small-molecule drugs against RSV. Methods: To address the limitations of current QSAR-based virtual screening strategies for RSV L-RdRp inhibitor development, we established a multi-dimensional computer-aided drug screening framework integrating activity, toxicity, drug-likeness, and stability. Results: Two OECD-compliant 2D-QSAR models were developed and rigorously validated to predict inhibitory activity and cytotoxicity, respectively. The optimal inhibitory activity model exhibited strong statistical performance, with R² = 0.8281, $Q_{\mathrm{L}\mathrm{O}\mathrm{O}}^2=$ 0.7653, $R_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}^2=$ 0.8713, $Q_{\mathrm{F}\mathrm{n}}^2=$ 0.8594 ∼ 0.8837, CCC_test = 0.9301, MAE_test = 0.1966. Similarly, the best cytotoxicity model achieved R²= 0.8263, $Q_{\mathrm{L}\mathrm{O}\mathrm{O}}^2$ = 0.7422, $R_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}^2$ = 0.8951, $Q_{\mathrm{F}\mathrm{n}}^2$ = 0.8108~0.8530, CCC_test = 0.9081, MAE_test = 0.1685. Based on these models, a four-step screening workflow—QSAR-based filtering and molecular docking (15,758 → 2446 → 162 → 19 compounds), ADMET evaluation (19 → 5), and molecular dynamics simulations (MDSs)—was implemented to identify promising L-RdRp inhibitors. Conclusions: Ultimately, five candidate compounds were selected, all of which demonstrated predicted higher inhibitory activity, lower predicted cytotoxicity, a stable predicted binding mode, and favorable oral bioavailability compared with the reference drug remdesivir. These findings provide valuable in silico-derived lead candidates and a reliable computational workflow for identifying experimental L-RdRp inhibitors targeting RSV.