AUC-Maximized Deep Convolutional Neural Fields for Protein Sequence Labeling | JoVE Visualize

Area of Science:

Computational Biology
Machine Learning
Bioinformatics

Background:

Deep Convolutional Neural Networks (DCNNs) excel in machine learning but struggle with imbalanced datasets.
Existing training methods like maximum-likelihood and maximum labelwise accuracy are suboptimal for imbalanced sequence labeling.
Conditional Random Fields (CRFs) are effective for sequence labeling but require robust training strategies.

Purpose of the Study:

To introduce Deep Convolutional Neural Fields (DeepCNF), integrating DCNNs and CRFs for sequence labeling.
To develop and evaluate a novel training algorithm, maximum-AUC, specifically designed for imbalanced label distributions.
To improve the performance of sequence labeling models on biological prediction tasks with skewed data.

Main Methods:

Integration of Deep Convolutional Neural Networks (DCNN) with Conditional Random Fields (CRF) to form DeepCNF.
Development of a maximum-AUC training algorithm that directly optimizes the Area Under the ROC Curve (AUC).
Formulation of AUC within a pairwise ranking framework, approximated by a polynomial function, and optimized using gradient-based methods.

Main Results:

The maximum-AUC training method significantly outperforms maximum-likelihood and maximum labelwise accuracy on imbalanced datasets for 8-state secondary structure and disorder prediction.
Maximum-AUC shows comparable performance to other methods on balanced datasets like solvent accessibility prediction.
DeepCNF models trained with maximum-AUC demonstrate superior performance over existing predictors for the evaluated biological tasks.

Conclusions:

The proposed maximum-AUC training algorithm is highly effective for DeepCNF in sequence labeling tasks with imbalanced label distributions.
DeepCNF with maximum-AUC offers a powerful new approach for biological sequence analysis, outperforming current state-of-the-art methods.
The developed method provides a robust solution for challenges posed by imbalanced data in machine learning applications.