O03_05
Predicting Novel Therapeutic Target Molecules Using Neural Networks: Validation and Applicability to Unknown Diseases
Hayato TSUMURA *1, Narumi HATANO1, Mayumi KAMADA2, Ryosuke KOJIMA1, Hiroaki IWATA3, Yasushi OKUNO1, 4
1Graduate School of Medicine, Kyoto University
2School of Frontier Engineering, Kitasato University
3Faculty of Medicine, Tottori University
4RIKEN Center for Computational Science(RCCS)HPC/HPC- and AI-driven Drug Development Platform Division
( * E-mail: tsumura.hayato.22x@st.kyoto-u.ac.jp)
Identifying therapeutic target molecules related to disease causation and progression is a crucial and challenging step in the early stage of drug discovery. Although the space of potential target molecules is vast, the number of experimentally verifiable molecules is limited. Thus, computational approaches, especially machine learning, are increasingly adopted to discover novel therapeutic target molecules.
Gene expression profiles obtained from directional perturbations such as knockdown or overexpression of genes encoding drug target molecules can reflect the functionality of drugs that inhibit or activate these targets. A previous method has demonstrated that gene expression profiles representing the perturbation response of candidate target genes and similarity scores between diseases are beneficial in predicting new target-disease relationships. However, this approach was limited to diseases with known therapeutic target molecules, and there is room for improvement in prediction accuracy (ROC-AUC: 0.63 for inhibitory target prediction, 0.65 for activatory target prediction).
This study presents a neural network-based approach to predict novel therapeutic targets applicable to diseases regardless of with and without known targets. The validation results show a significant improvement in prediction accuracy compared to previous methods (ROC-AUC: 0.93 for inhibitory target prediction, 0.82 for activatory target prediction). Furthermore, A leave-one-out validation strategy was employed to examine the model's predictive performance on each disease, thereby assessing the robustness and generalizability of our approach. The results demonstrate that our model can predict therapeutic target molecules for diseases not included in the training phase. Moreover, our investigation of the predicted candidate target molecules has identified literature supporting their association with diseases. The results indicate the potential of our model to be applied to diseases with previously unknown target molecules.
The model proposed in this study has the potential to predict novel therapeutic target molecules for rare and novel diseases, which have been challenging for conventional approaches.