Sequence-based modeling of low-affinity transcription factor-DNA binding through deep learning.

Area of Science:

• Computational biology
• Genomics
• Molecular biology

Background:

• Transcription factor-DNA binding specificity is crucial for gene regulation.
• Deep learning models, including CNNs and SA transformers, have advanced TF-DNA binding prediction.
• Systematic evaluation of DNA sequence orientation handling in models is needed, especially for low-affinity binding.

Purpose of the Study:

• To compare different strategies for handling DNA sequence orientations in deep learning models for TF-DNA binding specificity.
• To evaluate interpretation methods for these models.
• To identify low-affinity binding sites and their mechanisms.

Main Methods:

• Utilized SELEX-seq data for eight Exd-Hox heterodimers in Drosophila.
• Compared canonical models with data augmentation and reverse-complement weight-sharing models (CNNs and SA transformers).
• Evaluated interpretation methods: Gradient*input, DeconvNet, DeepLIFT, and in silico mutagenesis (ISM).

Main Results:

• Reverse-complement weight-sharing CNN and SA models trained with augmented data outperformed other approaches.
• In silico mutagenesis (ISM) was less sensitive to hyperparameter settings compared to other interpretation methods.
• Identified Exd-Ubx binding at low-affinity sites and proposed biophysical mechanisms.

Conclusions:

• Reverse-complement strategies enhance deep learning models for TF-DNA binding specificity, especially for low-affinity interactions.
• ISM is a robust interpretation method for these models.
• Findings contribute to understanding low-affinity TF binding in gene regulation and inform TF-DNA binding specificity-guided protein design.