Weakly-Supervised Convolutional Neural Network Architecture for Predicting Protein-DNA Binding.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Convolutional Neural Networks (CNNs) excel at predicting protein-DNA binding sequence specificities.
• Existing CNN models do not fully leverage weakly-supervised information from DNA sequences containing multiple Transcription Factor Binding Sites (TFBS).

Purpose of the Study:

• To develop a novel Weakly-Supervised Convolutional Neural Network (WSCNN) architecture.
• To enhance the prediction accuracy of protein-DNA binding by integrating Multiple Instance Learning (MIL) with CNNs.

Main Methods:

• Proposed WSCNN architecture combining MIL and CNN.
• DNA sequences divided into overlapping subsequences (instances) using a sliding window.
• CNN models each instance; scores fused using Max, Average, Linear Regression, or Top-Bottom Instances methods.

Main Results:

• WSCNN demonstrated improved performance on both in vivo and in vitro datasets.
• Models trained on in vitro data accurately predicted in vivo protein-DNA binding.
• Quantitative analysis highlighted the importance of reverse-complement mode and explained limitations of direct MIL-CNN pooling.

Conclusions:

• WSCNN effectively utilizes weakly-supervised information for enhanced protein-DNA binding prediction.
• The approach shows promise for cross-dataset prediction (in vitro to in vivo).
• The study provides insights into optimal model design for MIL-CNN integration in bioinformatics.