DeePhase

DeePhase predicts protein propensity for homotypic liquid-liquid phase separation (LLPS) from amino acid sequence to inform study of biomolecular condensates and cellular compartmentalization.

Key Features:

• In Silico Strategy: Performs global in silico analysis of associations between protein sequences and phase behavior.
• Machine Learning Models: Integrates machine learning approaches, including neural network-based language models, to predict LLPS propensity.
• Sequence Feature Analysis: Identifies LLPS-prone sequences as more intrinsically disordered, less hydrophobic, and lower in Shannon entropy compared with entries from the Protein Data Bank (PDB) and Swiss-Prot, and emphasizes a balance between polar and hydrophobic residues.
• Integrated Model: Combines knowledge-based sequence features with unsupervised embeddings from neural network models to distinguish LLPS-prone sequences from structured proteins and from unstructured proteins with lower LLPS propensity.
• High-Accuracy Prediction: Demonstrates high accuracy in identifying LLPS-prone sequences within the human proteome.

Scientific Applications:

• Cellular compartmentalization: Predicts proteins that may form biomolecular condensates to study mechanisms of cellular compartmentalization and regulation.
• Mechanistic hypothesis generation: Provides sequence-based hypotheses about molecular determinants of phase separation (disorder, hydrophobicity, Shannon entropy, polar/hydrophobic balance) for experimental testing.
• Proteome- and disease-focused screening: Enables proteome-scale identification of candidate phase-separating proteins to investigate roles in biological processes and diseases.

Methodology:

Trains classifiers on knowledge-based sequence features and unsupervised embeddings derived from neural network-based language models, performs global in silico association analyses between sequence and phase behavior, and validates predictions against established databases such as the Protein Data Bank and Swiss-Prot.