CBInformax: bioactivity-aware self-supervised molecular representation learning for molecular property and drug-drug...

Molecular representation learning plays a central role in molecular property prediction and computer-aided drug discovery. Recent self-supervised pretraining methods have substantially improved molecular representations by exploiting large-scale unlabeled molecular data. However, most methods remain structure-centered and rely mainly on 2D or 3D molecular information, while overlooking bioactivity-related knowledge that can provide complementary functional signals. To address this limitation, we propose CBInformax, a bioactivity-aware self-supervised framework for molecular representation learning. CBInformax models each molecule from two complementary views: a 2D molecular graph and a bioactivity similarity network. A local encoder learns structure-based molecular representations from 2D graphs, whereas a global encoder captures inter-molecular relationships in the bioactivity space. The two views are aligned through a cross-view mutual-information maximization objective, enabling bioactivity-related knowledge to be transferred into structure-based embeddings during pretraining. Importantly, the pretrained molecular encoder can be directly fine-tuned for downstream tasks without requiring explicit bioactivity information at inference time. Experiments on multiple MoleculeNet benchmark datasets show that CBInformax achieves competitive or superior performance across classification and regression tasks. In addition, the learned representations improve drug-drug interaction prediction and exhibit enhanced sensitivity in distinguishing structurally similar molecules with subtle chemical differences. These results suggest that incorporating bioactivity similarity during pretraining can enrich molecular representations.