MoActiPred: an AI-Based bioactivity prediction model applied to the development of influenza neuraminidase inhibitors

Accurate prediction of small-molecule bioactivity is critical for accelerating drug discovery. Here, we report MoActiPred, a multimodal AI framework that integrates 2D molecular graphs, physicochemical descriptors, and 3D E (3)-equivariant geometric representations. Using a curated ChEMBL-derived influenza virus neuraminidase (NA) inhibitor dataset (1245 compounds; IC 50 threshold 20 nM), MoActiPred showed the strongest overall performance among the evaluated methods, including traditional machine-learning models (Random Forest, SVM, and XGBoost), SMILES-based language models (ChemBERTa, ChemFM, and PolyBERT), and graph-based baselines (MolCLR and SEGNN). On the public test set, MoActiPred obtained an accuracy of 0.952, MCC of 0.797, and F1 score of 0.898, while retaining high average precision, supporting the value of combining 2D and 3D representations. To evaluate its practical utility, MoActiPred was applied prospectively to 26 oseltamivir analogues designed, experimentally synthesised and evaluated against influenza A (H3N2) NA. The model correctly predicted 6 of 7 high-activity compounds (IC 50 < 20 nM) and 18 of 19 inactive or weakly active derivatives, yielding 92.3% accuracy. Enzymatic assays showed that most derivatives were active, with 5b and 5f emerging as the most potent inhibitors (IC 50 4.31 and 6.49 nM, respectively). In MDCK.2 cell-based assays, 5b and 5f exhibited strong H3N2-selective antiviral activity, with EC₅₀ values of 0.72 nM and 3.04 nM, respectively, and no detectable cytotoxicity (CC 50 > 100 μM). Docking, MD simulations, and water-map analyses indicated that the oseltamivir core of 5b occupies the conserved active site while its N -substituent engages the 150- and 430-cavities, with solvation-driven binding particularly favouring N2. These results support that MoActiPred’s utility in drug discovery research and identify 5b as a promising lead for further optimisation against influenza A (H3N2). Scientific contribution MoActiPred advances bioactivity prediction by integrating 2D molecular graphs, physicochemical descriptors, and 3D E (3)-equivariant geometric representations within a single multimodal framework. Compared with conventional machine-learning models, SMILES-based language models, and graph-based baselines, MoActiPred achieved stronger classification performance on the curated and public neuraminidase inhibitor datasets, demonstrating the value of combined 2D/3D molecular representation learning. Its prospective utility was further supported by in-house synthesis of a focused oseltamivir-analogue series followed by enzymatic and cell-based biological activity testing.