DIFFUSE

DIFFUSE predicts isoform functions from isoform sequences, domain information, and expression profiles using a hybrid deep neural network and conditional random field framework to annotate functions of alternatively spliced isoforms.

Key Features:

• Integration of Diverse Data Types: Combines isoform sequences, domain information, and expression profiles for function prediction.
• Hybrid Deep Learning Framework: Uses a deep neural network (DNN) to predict functions from sequences and a conditional random field (CRF) to refine predictions by incorporating co-expression relationships.
• Iterative Semi-Supervised Learning: Implements iterative semi-supervised learning with self-training to train the DNN and CRF models despite limited isoform-level annotations.
• Performance: Reports average area under the receiver operating characteristic curve (AUC) of 0.840 and area under the precision-recall curve (PR AUC) of 0.581 across 4184 Gene Ontology (GO) functional categories.

Scientific Applications:

• Isoform Function Annotation: Predicts GO functional annotations for alternatively spliced isoforms to enable isoform-level functional catalogs.
• Gene Regulation and Protein Interaction Studies: Links predicted isoform functions to sequence and expression features to investigate isoform-specific roles in gene regulation and protein–protein interactions.
• Disease Mechanism Analysis: Identifies functionally distinct isoforms to explore molecular bases of diseases associated with alternative splicing events.

Methodology:

Computational steps comprise data integration of sequence, domain, and expression profiles; initial function prediction by a deep neural network (DNN) from isoform sequences; refinement of predictions using a conditional random field (CRF) that incorporates co-expression relationships; and iterative semi-supervised self-training to improve model training with limited labels.