Polymer Physics-Informed Neural Network

Enforces the power-law relationship between genomic distance and contact frequency. The model learns to predict Hi-C maps that obey the characteristic decay exponent observed in real chromatin.

• Mechanism: A custom loss function penalizes deviations from the expected distance-dependent contact probability.
• Physical Basis: Derived from the fractal globule model of chromatin organization.
• Benefit: Prevents the model from predicting physically impossible long-range interactions.

Incorporates the physical principle that two segments of the chromatin fiber cannot occupy the same space simultaneously. This prevents the generation of structurally impossible, interpenetrating conformations.

• Implementation: A repulsive potential term is added to the structural optimization objective.
• Biological Relevance: Ensures realistic chromatin packing density within the nucleus.
• Validation: Predicted structures are checked for steric clashes against known nuclear volume constraints.

Models the intrinsic stiffness of the chromatin fiber by penalizing sharp, energetically unfavorable bends. The model learns a persistence length consistent with the 30nm fiber or nucleosome-level flexibility.

• Parameter: A bending energy term controls the angular distribution between consecutive segments.
• Impact: Produces smooth, biologically realistic chromatin trajectories rather than jagged, random walks.
• Scale: Operates at the resolution of the model's bead-spring representation.

Directly simulates or approximates the action of loop-extruding factors like cohesin. The network's architecture or training objective is biased to form the hallmark corner peaks and domain structures of loop extrusion.

• Architectural Bias: Attention mechanisms can be constrained to favor local, processive interactions.
• Objective: Reconstructs the characteristic 'dots' and 'stripes' in Hi-C maps.
• Result: Accurately predicts dynamic chromatin loop formation and release.

Instead of predicting a single static structure, the model learns to sample from the Boltzmann distribution of possible chromatin conformations. This captures the dynamic, heterogeneous nature of the genome.

• Method: Variational autoencoders or generative adversarial networks are trained with an energy-based loss.
• Output: A population of structures, each consistent with the input data and physical constraints.
• Significance: Reflects the true biological reality of cell-to-cell variation in genome folding.

Represents the chromatin fiber as a coarse-grained polymer chain, where each bead corresponds to a genomic region of a defined resolution. This bridges the gap between atomic-level physics and whole-genome modeling.

• Resolution: Beads can represent 1kb, 10kb, or larger segments, depending on the model's scale.
• Connectivity: Springs between adjacent beads enforce chain connectivity.
• Advantage: Makes computationally intensive polymer simulations tractable for megabase-scale genomic loci.