DMFpred

DMFpred predicts molecular functions of intrinsically disordered proteins and regions (IDPs/IDRs) at the residue level to identify functionally annotated residues relevant to drug target discovery and rational drug design.

Key Features:

• Function scope: Predicts six disordered molecular function types: entropic chain, assembler, scavenger, effector, display site, and chaperone.
• Residue-level and multi-functional prediction: Provides residue-level predictions and supports identification of functional and multi-functional residues across five evaluation categories.
• Protein Cubic Language Model (PCLM): Implements PCLM as the core model architecture.
• Multi-model integration: Integrates three distinct protein language models that characterize sequence, structural features, and functional attributes.
• Attention-based alignment: Uses an attention-based alignment mechanism to generate a comprehensive joint representation of proteins.
• Training strategy: PCLM is pre-trained on a large-scale dataset of IDR sequences and fine-tuned on functionally annotated sequences for molecular function prediction.
• Addressing prior limitations: Extends predictive coverage beyond methods that predominantly focus on entropic chain functions to the other five categories of disordered functions.
• Performance evaluation: Demonstrated high-quality predictive performance across five categories of functional and multi-functional residues.

Scientific Applications:

• Drug target discovery and rational drug design: Supports identification of disordered protein functions relevant to drug target discovery and rational drug design.
• Computational functional annotation of IDPs/IDRs: Enables research-scale prediction of molecular functions of intrinsically disordered proteins and regions for downstream biological studies.

Methodology:

DMFpred implements the Protein Cubic Language Model (PCLM), which integrates three protein language models (sequence, structural, functional) via an attention-based alignment mechanism; PCLM was pre-trained on a large-scale dataset of IDR sequences and fine-tuned on functionally annotated sequences for molecular function prediction.