picnic-bio

picnic-bio predicts proteins localized to biomolecular condensates using a machine-learning classifier that integrates sequence-based amino acid pattern analysis, intrinsic disorder properties, and Alphafold2-derived structural features to identify condensate members across proteomes.

Key Features:

• Machine learning classifier: Uses a supervised machine-learning algorithm to classify proteins by localization to biomolecular condensates irrespective of their role in condensate formation.
• Sequence-based features: Analyzes amino acid patterns and sequence-derived features as predictive inputs.
• Structure-based features: Incorporates Alphafold2-derived structural features from predicted protein models.
• Intrinsic disorder properties: Uses intrinsic disorder metrics as features and maintains performance across proteins with varying disorder content.
• Training and validation: Trained on known condensate-localized proteins and experimentally validated, reporting approximately 82% accuracy for positive condensate members in cells.
• Proteome-wide, cross-species applicability: Applies to whole proteomes and supports comparative analyses across species.
• PICNIC-GO (GO-term features): Provides an extended model integrating Gene Ontology (GO) term–based features to highlight property enrichment in condensates while acknowledging annotation-driven bias.
• Disorder versus complexity analysis: Enables comparative analyses showing increased disorder with organismal complexity but no direct correlation between predicted condensate proteome content and disorder across species.

Scientific Applications:

• Condensate member identification: Identify candidate proteins that localize to biomolecular condensates for experimental follow-up.
• Unbiased proteome annotation: Generate proteome-wide annotations of condensate membership independent of phase-separation driver status and disorder bias.
• Comparative evolutionary analysis: Compare predicted condensate proteomes across species to study relationships with organismal complexity and intrinsic disorder.
• Hypothesis generation via GO features: Use PICNIC-GO's GO-term features to generate hypotheses about functional properties enriched in condensate proteins.
• Experimental prioritization: Prioritize proteins for cellular experiments informed by reported experimental validation accuracy.

Methodology:

Extracts sequence-based amino acid pattern features and intrinsic disorder properties, derives structure-based features from Alphafold2 models, trains a supervised machine-learning classifier on known condensate-localized proteins, and optionally extends features with Gene Ontology (GO) term annotations for PICNIC-GO.