TRUHiC

Hi-C sequencing turns chromosome conformation capture into genome-wide contact matrices, and those matrices are where topologically associating domains (TADs), chromatin loops, and A/B compartments become visible. Resolving them depends on read depth: fine-grained features need small genomic bins, small bins need deep sequencing, and deep sequencing is expensive. Most published Hi-C datasets are therefore too sparse for reliable loop calling, which has driven a decade of computational "super-resolution" methods that treat a contact map like an image and try to sharpen it.

TRUHiC, developed at Temple University with collaborators from the University of Southern California, the University of Washington, and the Human Genome Structural Variation Consortium, attacks the weakness those image-based methods share. CNN- and GAN-based predecessors such as HiCPlus, DeepHiC, HiCARN, and DFHiC rely on convolution, which captures local texture well but cannot represent the long-range dependencies that define chromatin architecture. TRUHiC embeds a transformer block inside a U²-Net so that the same network models fine local structure and global interaction context. It was posted to bioRxiv in April 2025 and revised in October 2025, and has not yet been peer reviewed.

The work also proposes a second, more practical contribution: TRUHiC-LCL, a cell-line-specific model trained across 43 real lymphoblastoid Hi-C samples rather than a single artificially downsampled one. That aggregation strategy is the same reasoning that motivates 3D-genome foundation models such as MIX-HIC and Evo2HiC, narrowed deliberately to one well-characterized cell type.

#Key Features

• Transformer bridge inside a U²-Net: A customized multi-head self-attention gating block replaces the RSU-4F bridge of the standard U²-Net, letting the model capture long-range contacts that stacked convolutions miss.
• Raw contact counts, no normalization: The network trains and predicts directly on unnormalized matrices, removing the normalization and denormalization steps that DeepHiC and HiCARN require.
• Cross-cell-line and cross-species transfer: A model trained only on GM12878 was applied unchanged to K562, IMR90, NHEK, and the mouse CH12-LX line, scoring highest on three of the four.
• Recovery of real sparse samples: On GM12329, a lymphoblastoid sample discarded from an earlier study for poor sequencing quality, enhancement raised detected chromatin loops from 158 to 4,177 while retaining all 158 originals.
• Cell-line-specific variant: TRUHiC-LCL, trained on 43 HGSVC lymphoblastoid samples, outperforms the equivalent model trained on 10 samples across all vision and biological metrics on three held-out biological replicates.

#Technical Details

The backbone is a U²-Net — a U-Net whose layers are themselves residual U-blocks (RSU). TRUHiC trims the encoder/decoder to two RSU-L blocks, removes their max-pooling and upsampling operations to preserve spatial resolution, adds a direct skip connection from input to output for pixel-level fidelity, and weights each auxiliary output's loss by 1/eˣ so early blocks reinforce rather than duplicate the final prediction. Training uses a composite loss of mean absolute error and signal-to-noise ratio with equal weights, L1 and L2 regularization on convolutions, the LAMB optimizer, learning-rate reduction on plateau, and early stopping. Inputs are 40 × 40 non-overlapping blocks of 10 kb-resolution contact matrices.

Benchmarks used GM12878 chromosomes 1–17 for training and 18–22 for testing at a 1/16 downsampling rate, with identical preprocessing applied to every competing method. TRUHiC led on all metrics, reaching PSNR 26.23, Spearman 0.692, Pearson 0.738, and GenomeDISCO 0.922 against DFHiC's 26.10, 0.684, 0.732, and 0.921. On NA19317 from HGSVC3, benchmarked against an independently generated ~3 kb reference for the same individual, TRUHiC raised detected loops from 1,210 to 15,926 and achieved the highest Jaccard and F1 scores for both TAD boundaries and loops. All models were trained on a single 80 GB A100.

#Applications

TRUHiC targets laboratories sitting on Hi-C data that was never sequenced deeply enough — population cohorts, archived samples, and comparative studies across tissues or species where re-sequencing is impractical. The GM12329 and NA19317 results show the concrete payoff: samples previously set aside for low coverage become usable for loop and TAD analysis, extending regulatory and structural-variant work to cohorts assembled under budget constraints. TRUHiC-LCL is aimed specifically at the lymphoblastoid cell lines that dominate population-scale functional genomics.

#Impact

TRUHiC is a preprint awaiting peer review, and the authors are candid about the field's central limitation: deep learning enhancers trained on artificially downsampled data lose substantial accuracy on experimentally sparse data, with reported drops of up to 57%, and TRUHiC's own metrics fall on GM12329 relative to the downsampled benchmarks. The cell-line-specific training strategy is the paper's answer, demonstrated so far only on chromosome 22. The GitHub repository carries training and inference code, benchmark reimplementations of the competing methods, and example GM12878 data, but no license file, and the pretrained TRUHiC-LCL checkpoint the paper describes is not among the deposited files — so the inference-only --mode denoise path currently runs against a checkpoint the user trains themselves.