DMIL-IsoFun

DMIL-IsoFun predicts isoform-level functions by applying a deep multi-instance learning framework that integrates convolutional neural networks (CNNs) and a class-imbalance graph convolutional network (GCN) to refine annotations using isoform sequences, gene-level annotations, and co-expression networks.

Key Features:

• Deep Multi-Instance Learning Framework: Integrates multi-instance learning with convolutional neural networks (CNNs) and graph convolutional networks (GCNs) to capture relationships between isoforms and their parent genes.
• Sequence and Annotation Integration: Leverages isoform sequences and gene-level annotations to initialize isoform feature vectors for downstream prediction.
• Class-Imbalance Handling: Employs a class-imbalance Graph Convolution Network that refines isoform annotations using co-expression networks and extracted features.
• Performance Improvement: Demonstrates experimental improvements in minimum sensitivity (Smin) by at least 29.6% and maximum F1 score (Fmax) by 40.8% over prior methods.

Scientific Applications:

• Functional Genomics: Provides more accurate isoform annotations to support fine-grained functional genomics analyses.
• Proteomic Diversity Studies: Aids investigation of proteomic diversity arising from alternative splicing and the distinct roles of isoforms.
• Biomedical Research: Supports analysis of disease mechanisms involving specific isoforms and can inform isoform-targeted therapeutic research.

Methodology:

Feature extraction via a multi-instance learning convolutional neural network trained on isoform sequences and gene-level annotations to obtain initial feature vectors; annotation initialization using the extracted features to assign isoform functional annotations; refinement of annotations using a class-imbalance Graph Convolution Network that incorporates co-expression relationships and the initial feature vectors.