CellChem: Cellular transcriptional responses reshape molecular representation space for efficient and multi-scale drug discovery
Chen, J.; Lin, L.; Wang, Y.; Lin, Y.; Li, Y.; Zhang, W.; Fu, Y.; Xie, J.; Zhu, J.; Sun, C.; Shi, G.; Wang, Z.; Lin, H.; Wang, L.; Deng, M.; Lai, L.; Pei, J.
Show abstract
Despite decades of progress in computational drug discovery, deep learning-based molecular representation models remain largely structure-centric, assuming that chemical similarity approximates functional similarity. However, drug effects in cells are shaped not only by chemical similarity but also by molecular interactions in the cellular context. To capture this complexity, we introduce CellChem, a cellular-chemical representation-learning framework that learns cell-guided molecular representations of small-molecule actions within cells. By incorporating large-scale cellular transcriptional profiles during pretraining, CellChem reshapes molecular representation space from structure-centric to a balanced integration of structure and function. The learned CellChem molecular representations exhibit biologically meaningful geometric organization, such that distances between molecules encode not only structural similarity but also similarity in the cellular responses they elicit, independent of downstream tasks. Using downstream derivative models such as cell-guided compound-protein interaction prediction and drug-induced transcriptional response profile generation, CellChem supports highly efficient, multi-scale drug discovery, achieving significantly better performance than traditional models.
Matching journals
The top 7 journals account for 50% of the predicted probability mass.