Direct AUC optimization of regulatory motifs

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Transcription factor binding site (TFBS) motif discovery is crucial for understanding genetic variation.
• Discriminative motif learning (DML) methods show promise but often sacrifice accuracy for speed.
• Existing DML methods may not fully leverage information from input sequences.

Purpose of the Study:

• To introduce a novel algorithm, CDAUC, for optimizing DML-learned motifs.
• To enhance the accuracy and efficiency of TFBS motif discovery.
• To improve the interpretability of deep learning models for TF sequence specificity prediction.

Main Methods:

• Developed the CDAUC algorithm for motif optimization using the area under the receiver-operating characteristic curve (AUC) criterion.
• Optimized the AUC loss function in a coordinate-wise manner, leading to efficiently solvable sub-problems.
• Solved a key iterative step as a computational geometry problem for efficiency.

Main Results:

• CDAUC significantly outperforms competing methods in refining DML motifs.
• CDAUC is approximately one order of magnitude faster than existing approaches.
• Preliminary results suggest CDAUC can improve the interpretability of deep learning models for TF sequence specificity.

Conclusions:

• CDAUC offers a more accurate and efficient approach to TFBS motif discovery.
• The algorithm effectively refines motifs learned by DML methods.
• CDAUC shows potential for advancing the analysis of TF binding and sequence specificity.