IDRdecoder: a machine learning approach for rational drug discovery toward intrinsically disordered regions.

Area of Science:

• Computational Biology
• Drug Discovery
• Protein Science

Background:

• Intrinsically disordered regions (IDRs) are increasingly recognized as crucial in biological activity and disease.
• IDRs present a promising avenue for drug discovery, yet rational design methodologies are underdeveloped.
• A significant gap exists in experimental data for IDR-targeted drug discovery.

Purpose of the Study:

• To develop a machine learning approach for predicting IDR functions and drug interaction sites.
• To identify molecular substructures within IDRs that interact with drugs.
• To address the scarcity of experimental data in IDR drug discovery.

Main Methods:

• Utilized stepwise transfer learning with a neural network model named IDRdecoder.
• Trained an autoencoder on over 26 million predicted IDR sequences.
• Fine-tuned the model on 57,692 ligand-binding PDB sequences with high IDR content.

Main Results:

• IDRdecoder successfully predicted functions and enriched relevant Gene Ontology (GO) terms for known IDR drug targets.
• Achieved an Area Under the Curve (AUC) of 0.616 for predicting drug interaction sites and 0.702 for ligand types.
• Demonstrated moderately improved performance compared to existing methods like ProteinBERT.

Conclusions:

• IDRdecoder is the first tool for predicting drug interaction sites and ligands in IDR sequences.
• Identified key characteristics for IDR-drug design, favoring Tyr and Ala residues as targets and flexible alkyl groups in ligands.
• Provides valuable insights for advancing rational drug design targeting IDRs.