BayesPI-FLY: a Bayesian neural network approach for inferring feature weighted TF-DNA interaction

Area of Science:

• Computational Biology
• Epigenetics
• Genomics

Background:

• Transcription factor (TF) DNA motif recognition is crucial for gene regulation.
• DNA methylation variability poses a challenge for integrating multi-omics data and understanding TF binding.
• Accurate characterization of TF-DNA interactions requires accounting for epigenetic modifications.

Purpose of the Study:

• To develop a computational framework, BayesPI-FLY, for de novo motif discovery that integrates DNA sequence and methylation data.
• To quantitatively characterize the influence of DNA methylation on TF binding at nucleotide and motif levels.
• To provide a versatile platform for analyzing TF-DNA interactions within complex epigenetic landscapes.

Main Methods:

• Developed BayesPI-FLY, a Bayesian neural network based on a biophysical model of TF-DNA interactions.
• Employed a two-layer inference architecture for joint estimation of model parameters and hyperparameters.
• Implemented core algorithms in C with Python parallelization for computational efficiency.
• Integrated DNA sequence information with DNA methylation status data.

Main Results:

• BayesPI-FLY successfully integrates DNA sequence and methylation data for motif discovery.
• The framework quantitatively characterizes methylation effects at single-nucleotide and motif levels.
• Validation using synthetic and whole-genome bisulfite sequencing data confirmed recapitulation of known methylation-associated TF-binding patterns.
• Inferred strand-specific TF-DNA associations within the modeling framework.

Conclusions:

• BayesPI-FLY is an effective computational platform for de novo motif discovery incorporating DNA methylation.
• The tool facilitates the interpretation of TF-DNA binding patterns influenced by epigenetic modifications.
• BayesPI-FLY advances the study of gene regulation in diverse epigenetic contexts.