The aberrant expression of proviral integration site for Moloney murine leukemia virus (PIM) kinases is closely related to various tumors and chemotherapy resistance, making them attractive targets for cancer therapy. However, due to the extremely high homology among the three PIM isoforms (PIM1, PIM2, PIM3) and the limited availability of existing bioactivity data, screening and designing selective PIM inhibitors remain a daunting challenge. To address this issue, this study constructed a multitask regression model that can simultaneously predict the half-maximal inhibitory concentration (IC50 values). The model utilizes an attention mechanism to capture effects within local atomic groups and the interactions between different groups of atoms. Through weight sharing, the model enhances the accuracy of predicting PIM3 inhibitors by leveraging the rich and highly correlated data from PIM1 and PIM2 isoforms. Additionally, visualizing the weights of nodes (atoms in the molecule) in the model helps us to intuitively understand the relationship between molecular features and prediction outcomes, thereby enhancing the interpretability of the model. In summary, this work provides new insights and methods for performing activity prediction tasks for multiple similar targets in low-data scenarios.
Keywords: Graph Attention Mechanism; Multitask interpretable model; PIM inhibitors; Transfer learning.
© 2024. The Author(s), under exclusive licence to Springer Nature Switzerland AG.