For decades, computational genomics has operated in "Flatland." Researchers typically treat the genome as a one-dimensional string of nucleotides, annotating functional elements like promoters and enhancers based on their linear position. However, as Maxwell W. Libbrecht and Ferhat Ay argue in Joint annotation of chromatin state and chromatin conformation, this abstraction ignores the physical reality of the cell nucleus. DNA is a three-dimensional object, folded into complex loops and territories that bring distant genomic regions into direct physical contact.
A new wave of research is moving beyond linear sequences to integrate 3D chromatin architecture directly into deep learning frameworks. By treating the genome as a spatial network rather than a text string, these models capture the long-range regulatory interactions that drive gene expression, DNA repair, and disease progression.
The Spatial Context of Regulation
The core intuition behind this shift is that physical proximity in the nucleus is a primary driver of biological function. An enhancer located 100,000 base pairs away from a gene on a linear map might be sitting directly on top of that gene's promoter in 3D space.
Traditional models, such as those discussed by Osman Abul in Compo, often rely on simplified spatial assumptions. In contrast, newer frameworks like MIX-HIC, developed by Minghao Yang and colleagues, explicitly use Hi-C contact maps—experimental data that measures how often different parts of the genome touch—as a structural scaffold. In Multimodal 3D Genome Pre-training, Yang demonstrates that incorporating this 3D information provides critical, non-redundant signals that linear sequence or accessibility data alone cannot provide.
Representing the Genome as a Graph
If the genome is a 3D network, the natural mathematical language to describe it is Graph Theory. Researchers are increasingly replacing standard convolutional or recurrent layers with Graph Neural Networks (GNNs).
In this framing, genomic regions (bins) are nodes, and the physical contacts between them are edges. For example, Wei-Wei Tu and Hao Li developed GraphDSB to predict DNA double-strand breaks. By modeling chromosomes as 3D graphs, Graph Neural Networks for Double-Strand DNA Breaks Prediction shows that the spatial arrangement of chromosomes is a primary predictor of genomic instability. Similarly, Chen Fang and Zhilong Hu introduced Cell-Graph Compass, which moves away from the "genes as words" analogy to model cells as graphs where edges represent regulatory interactions and spatial transcriptomics data (Cell-Graph Compass).
From Tabular Data to Genomic Images
Another powerful approach treats the 3D genome as a computer vision problem. Frameworks like GENIUS, developed by Mateo Sokač and Benjamin Haibe-Kains, transform multi-omics data into 2D images where pixels represent spatially connected genes.
As explained in GENIUS: GEnome traNsformatIon and spatial representation of mUltiomicS data, mapping genomic data onto a grid based on 3D interactions allows models to leverage advanced Convolutional Neural Networks (CNNs). This approach outperforms traditional tabular formats because it preserves the "neighborhood" context of the nucleus—the fact that genes in close physical proximity often work together to regulate complex pathways. Yuchen Yuan applied a similar logic in DeepCNA, treating copy number aberrations and 3D structure as image channels to achieve a massive performance boost in cancer type prediction (DeepCNA).
Multimodal Foundation Models
The current frontier is the development of Genomic Foundation Models (GFMs) that are pre-trained on both sequence and structure. Models like Evo2HiC (Tangqi Fang et al.) distill the knowledge of massive DNA language models into efficient tools for analyzing 3D architecture (Evo2HiC).
These models aim to solve the "zero-shot" challenge: predicting how the genome will fold and function in a new cell type without needing new experimental 3D data. For instance, CREaTor (Yongge Li et al.) uses hierarchical Transformers to map cis-regulatory elements to target genes up to 2 megabases away, effectively inferring 3D structures from 1D epigenomic signals (CREaTor).
Why 3D Matters for the Future
Transitioning to 3D-aware models is not just about accuracy; it is about biological plausibility. As Wuwei Tan notes in Multimodal learning of noncoding variant effects, sequence-only methods are often "blind" to critical regulatory motifs that only become relevant when the genome folds. By integrating 3D aware-positional encodings and geometric priors—as seen in the work of Yize Zhou on InertialGenome (A Resolution-Agnostic Geometric Transformer)—researchers are finally building models that reflect the physical reality of the cell, enabling a deeper understanding of how the spatial environment of the nucleus governs life and disease.