Dense neural networks for predicting chromatin conformation.

Area of Science:

• Genomics
• Computational Biology
• Biophysics

Background:

• Eukaryotic DNA folds into higher-order structures like chromatin fibers and loops, influenced by architectural factors.
• Predicting DNA folding from bound factors (sequence) is a complex, unsolved problem in polymer physics.
• Existing polymer models can reproduce some structural features but lack a direct link between sequence and structure.

Purpose of the Study:

• To develop a predictive model for chromatin folding based on bound factors.
• To create an inverse model predicting chromatin factor sequences from structural data.
• To leverage machine learning for understanding DNA higher-order structure.

Main Methods:

• Trained a dense neural network to predict chromatin structure (contact map) from bound chromatin factor sequences.
• Incorporated a convolutional filter to compress factor data into a 1D sequence representation.
• Developed an inverse neural network to predict factor sequences from contact maps.

Main Results:

• The neural network successfully predicts local chromatin folding, linking factor sequences to contact maps.
• The inverse network predicts likely chromatin factor sequences from structural contact map data.
• Sensitivity analysis revealed the importance of chromatin context and neighborhoods in regulating long-range contacts.

Conclusions:

• The developed neural networks provide intuitive and informative physical insights into chromatin folding.
• The models highlight critical factors and alterations affecting DNA contact formation.
• This approach offers a new way to rapidly assess the effects of varying DNA sequence or structure.