What is MIDI
Cancer drug discovery using genetic information is still poorly developed. Precisely locating drug atoms and explaining the targeting effect is crucial in precision medicine since it helps understand the drug's mechanism of action. Much data has been collected regarding drug response against cancer cell lines and many models predict the drug response based on genomic information. However, to our knowledge, none of the data-driven techniques propose to detect the targeting mechanism of small drug molecules against genetic targets. In this work, we propose MIDI (Mechanism Interpretable Drug-Gene Interaction) model to delve deep into the targeting relation between drug molecules against genetic patterns. We show that purely based on a data-driven approach, the attention mechanism in our model could capture the important binding effect of small molecules towards gene targets, we provide both theoretical derivation and experiment results to show the information flow regarding the attention mechanism. In the meantime, we demonstrate that our model presents much higher prediction performance with the interpretation mechanism than the other state-of-the-art drug response prediction models.
How MIDI works
A. Technology pipeline and general model architecture.
B. Learning strategy of incorporating prior knowledge on drug-gene interaction.
C. Detailed illustration of the self-attention GraphFormer architecture for processing drug chemical structure input.
D. Detailed illustration for Drug-Gene cross-attention architecture design for important gene ranking and selection.