A multitask interpretable model with graph attention mechanism for activity prediction of low-data PIM inhibitors

Area of Science:

• Biochemistry
• Computational Chemistry
• Drug Discovery

Background:

• Aberrant expression of proviral integration site for Moloney murine leukemia virus (PIM) kinases is linked to various cancers and chemotherapy resistance.
• Designing selective PIM inhibitors is challenging due to high homology among PIM1, PIM2, and PIM3 isoforms and limited bioactivity data.

Purpose of the Study:

• To develop a multitask regression model for simultaneous prediction of PIM kinase inhibitor activity (IC50 values).
• To enhance the accuracy and interpretability of predicting inhibitors for highly similar protein targets, especially in low-data scenarios.

Main Methods:

• Constructed a multitask regression model incorporating an attention mechanism to analyze local atomic group effects and inter-group interactions.
• Employed weight sharing to leverage correlated data from PIM1 and PIM2 for improved PIM3 inhibitor prediction.
• Utilized node weight visualization for enhanced model interpretability, linking molecular features to prediction outcomes.

Main Results:

• The multitask regression model accurately predicts half-maximal inhibitory concentration (IC50) values for PIM kinase inhibitors.
• Attention mechanisms and weight sharing improved prediction accuracy, particularly for the PIM3 isoform.
• Model interpretability was enhanced through visualization of atom-level contributions to predictions.

Conclusions:

• This work presents a novel approach for activity prediction of inhibitors targeting multiple similar protein kinases.
• The developed model offers valuable insights and methods for drug discovery in low-data environments.
• The strategy addresses the challenge of designing selective inhibitors for highly homologous targets like PIM kinases.