F‑CPI:\nA Multimodal Deep Learning Approach\nfor Predicting Compound Bioactivity Changes Induced by Fluorine Substitution

Fluorine (F) substitution is a common method of drug\ndiscovery\nand development. However, there are no accurate approaches available\nfor predicting the bioactivity changes after F-substitution, as the\neffect of substitution on the interactions between compounds and proteins\n(CPI) remains a mystery. In this study, we constructed a data set\nwith 111,168 pairs of fluorine-substituted and nonfluorine-substituted\ncompounds. We developed a multimodal deep learning model (F-CPI).\nIn comparison with traditional machine learning and popular CPI task\nmodels, the accuracy, precision, and recall of F-CPI (∼90,\n∼79, and ∼45%) were higher than those of GraphDTA (∼86,\n∼58, and ∼40%). The application of the F-CPI for the\nstructural optimization of hit compounds against SARS-CoV-2 3CL<sup>pro</sup> by F-substitution achieved a more than 100-fold increase\nin bioactivity (IC<sub>50</sub>: 0.23 μM vs 28.19 μM).\nTherefore, the multimodal deep learning model F-CPI would be a veritable\nand effective tool in the context of drug discovery and design.