Akita

Akita ingests raw DNA sequence at single-nucleotide resolution across megabase-scale genomic windows. The input is one-hot encoded, representing each nucleotide (A, C, G, T) as a binary vector. This eliminates the need for hand-crafted features and allows the model to learn regulatory motifs and sequence grammar directly from the data. The input window is typically 1–2 Mb, providing sufficient context for predicting local and long-range chromatin interactions.

The core of Akita is a residual network with exponentially increasing dilation rates. This design allows the model's receptive field to grow exponentially with depth, capturing interactions between genomic loci separated by hundreds of kilobases without requiring quadratic self-attention mechanisms. Key properties:

• Exponential dilation: Dilation rates double at each layer (1, 2, 4, 8, ...), enabling the network to integrate information across the entire input window.
• Residual connections: Skip connections preserve gradient flow and allow the model to refine predictions iteratively.
• Symmetry: The architecture is designed to produce symmetric contact map outputs, respecting the inherent symmetry of Hi-C matrices.

Akita's decoder transforms the latent representation into a predicted Hi-C contact map. The output tower uses transposed convolutions to upsample the representation to the target resolution. The model predicts contact probability at multiple resolutions simultaneously through separate output heads:

• Fine-scale head: Predicts contacts at high resolution (e.g., 2 kb bins) for local interaction detail.
• Coarse-scale head: Predicts contacts at lower resolution (e.g., 32 kb bins) for long-range compartment-level structure. This multi-scale supervision ensures the model learns both fine-grained loop structures and broad A/B compartment organization.

Akita is trained to minimize the Poisson negative log-likelihood between predicted and experimental Hi-C contact maps. The loss function incorporates genomic distance normalization to account for the expected power-law decay of contact frequency with linear genomic separation. Without this normalization, the model would be dominated by short-range contacts and fail to learn long-range interaction patterns. The loss is computed as:

• Loss = Σ (predicted_ij - observed_ij * log(predicted_ij))
• Weighted by inverse distance-dependent expected contact frequency. This ensures the model allocates learning capacity to biologically meaningful deviations from the background distance decay.

Akita's architecture supports transfer learning from data-rich cell types to data-scarce contexts. The model is pre-trained on high-coverage Hi-C data from well-characterized cell lines (e.g., GM12878, H1-hESC) and fine-tuned on target cell types with limited data. The dilated convolutional backbone learns universal sequence determinants of folding, while the output tower adapts to cell-type-specific chromatin organization. This approach dramatically reduces the amount of experimental Hi-C data required for accurate prediction in new biological contexts.

Akita's performance is evaluated using the Stratum-Adjusted Correlation Coefficient (SCC), a metric specifically designed for Hi-C data. SCC measures the correlation between predicted and observed contact maps while controlling for genomic distance strata. This prevents inflated performance estimates from the trivial correlation of short-range contacts. Akita achieves SCC scores exceeding 0.8 on held-out chromosomes, demonstrating its ability to generalize across genomic regions unseen during training.

The core computational task that Akita solves: directly predicting genome-wide chromatin interaction frequencies from raw linear DNA sequence using deep learning. This bypasses the need for costly and time-consuming experimental Hi-C assays. The model learns to map nucleotide patterns to spatial proximity, implicitly capturing the binding logic of architectural proteins like CTCF and cohesin. Akita's architecture uses a deep convolutional neural network with dilated convolutions to capture long-range dependencies up to megabases away, outputting a predicted contact map for a given genomic locus.

A genome-wide matrix quantifying the interaction frequencies between all pairs of genomic loci, serving as both the training target and primary output for Akita. Derived from chromosome conformation capture assays, these maps reveal the 3D architecture of the genome. Key features Akita learns to predict include:

• Topologically Associating Domains (TADs): Self-interacting blocks visible as triangles along the diagonal
• Chromatin loops: Off-diagonal punctate peaks indicating distal element interactions
• A/B compartments: Plaid patterns reflecting open and closed chromatin segregation

Akita builds directly upon the Basenji framework, a deep convolutional neural network originally designed for predicting gene expression and epigenomic tracks from DNA sequence. Akita adapts this architecture by replacing the final regression head with a layer that predicts a symmetric contact probability matrix. The model uses dilated residual convolutions to exponentially expand its receptive field without losing spatial resolution, enabling it to integrate sequence information across the megabase distances relevant to 3D genome folding.

The mechanistic biological process that Akita implicitly learns to simulate. In this model, cohesin complexes load onto chromatin and actively reel DNA to form progressively larger loops until blocked by CTCF boundary elements. Akita's predictions recapitulate loop extrusion patterns without being explicitly programmed with the physics. The model learns that convergent CTCF motifs create loop anchors and that motif orientation determines extrusion directionality, validating that the sequence code contains sufficient information to predict 3D structure.

The primary evaluation metric used to benchmark Akita's prediction accuracy. Unlike standard correlation, SCC accounts for the distance-dependent background signal in Hi-C data—contacts decay predictably with genomic distance. SCC stratifies contact predictions by genomic distance and computes correlation within each stratum, preventing models from achieving high scores simply by recapitulating distance decay. Akita achieves high SCC scores across multiple cell types, demonstrating its ability to predict cell-type-specific folding patterns from sequence alone.

A key capability of Akita: the model trained on Hi-C data from one cell type can predict 3D folding in unseen cell types without retraining. This zero-shot transfer works because the DNA sequence is largely invariant across cell types—what changes is the activity of transcription factors and chromatin regulators. Akita learns a sequence-to-structure mapping that captures the intrinsic folding potential encoded in the genome, making it a foundational tool for predicting structural changes in contexts where experimental Hi-C data is unavailable, such as rare cell types or disease states.