MoBiTDE: A Molecular Bilinear Two-Tower Network for Drug-Excipient Nanoparticle Co-Assembly Prediction

Up to 40% of clinical candidates are poorly-soluble drugs, which rely heavily on nanoparticle formulations for oral and targeted delivery. Successful formulation largely depends on the co-assembly of drug and excipients into stable nanostructures. However, traditional wet-lab trial-and-error approaches require months, even years for discoveries to be made, with a substantial cost in raw materials. From research, there are 1440 high-throughput screening pairs of drug combination, but few AI algorithms have been applied to this task, especially with the rapidly evolving chemistry large language model (LLM), which hasn’t been evaluated on this task yet. Herein, I introduce the novel application of chemistry foundation model representations to drug-excipient nanoparticle co-assembly prediction. Specifically, this essay introduces MoBiTDE, a molecular bilinear two-tower network designed to learn pairwise compatibility between drugs and excipients. MoBiTDE employs frozen SMI-TED embeddings as pre-trained molecular representations, complemented by 24 RDKit-derived physicochemical descriptors. Drugs and excipient representations are processed through two independent molecular towers and coupled by a rank-32 bilinear interaction module to explicitly capture cross-molecular interactions. The resulting pairwise representation is subsequently fed into a single positive-weighted binary classification head for co-assembly prediction. Experimental results demonstrate that MoBiTDE outperforms other methods under different evaluation settings. Notably, its predictive performance further improves as the dataset expands, highlighting the potential of pre-trained molecular representations to benefit from increasing experimental data. In summary, MoBiTDE provides an effective framework for learning drug–excipient compatibility and offers a promising computational approach to facilitate the discovery of self-assembling nanoparticle formulations. It can improve drug discovery efficiency by reducing the number of trials, focusing only wet-lab testing on promising combinations.