The human genome contains roughly 3 billion DNA base pairs, yet it must fit inside a cell nucleus that is only a few micrometers across. DNA does not simply crumple randomly - it folds into a precise three-dimensional architecture where specific regions of different chromosomes come into close physical contact. These contacts, called chromatin interactions, are not passive folding accidents but active regulatory decisions that determine which genes get turned on or off.
The most important chromatin interactions are those between enhancers (distant DNA sequences that boost gene activity) and promoters (the control switches directly beside genes). When an enhancer physically contacts a promoter by looping through three-dimensional space, it activates that gene. When the DNA architecture changes - as frequently happens in cancer - wrong enhancers contact wrong promoters, turning on cancer-driving genes or silencing tumor suppressors.
Two laboratory techniques can measure chromatin interactions across the entire genome: Hi-C (which maps all physical contacts between DNA regions) and ChIA-PET (which maps contacts specifically organized by proteins like CTCF or RNA Polymerase II). Both are expensive, technically demanding, and require large amounts of tissue - making it impossible to run them on hundreds of individual cancer patient samples. This creates a critical knowledge gap: we know chromatin architecture matters in cancer, but we cannot measure it at scale.
This study introduces ChINN (Chromatin Interaction Neural Network), a machine learning approach that predicts chromatin interactions from DNA sequence alone - without any laboratory experiments. If successful, ChINN would allow researchers to computationally infer the 3D genome organization of hundreds or thousands of patient samples using only their DNA sequence or accessible chromatin data, which is far cheaper to generate.
ChINN is built as a two-stage machine learning pipeline. In the first stage, a convolutional neural network (CNN) - the same type of deep learning architecture used for image recognition - is applied to the raw DNA sequence at both ends (anchors) of a potential chromatin interaction. The CNN scans the sequence with learnable filters called kernels that automatically detect short DNA patterns, such as transcription factor binding sites, without being told what to look for.
The CNN produces a set of numerical features - essentially a compressed description of the sequence context at each genomic anchor. These features are then fed into a second algorithm called a gradient boosted tree (GBT) classifier, which combines the sequence features with genomic distance between the two anchors to produce a probability score: how likely is it that these two genomic regions are physically touching inside the nucleus?
The model was trained on three types of experimentally measured chromatin interactions: CTCF ChIA-PET (interactions organized by the CTCF architectural protein that creates genomic boundaries), RNA Polymerase II ChIA-PET (interactions at actively transcribed genes), and Hi-C (all physical chromatin contacts). Separate models were trained for each interaction type because their sequence determinants differ substantially.
A critical design choice was how to handle the positive-versus-negative sample imbalance. Chromatin interactions represent a tiny fraction of all possible pairings of genomic regions, so the researchers used distance-matched negative samples - non-interacting pairs at the same genomic distances as real interactions - to prevent the model from simply learning that nearby regions interact more often. This ensures the model learns true sequence determinants of chromatin contact rather than distance effects.
ChINN was trained and evaluated on well-characterized cell lines from the ENCODE project, including GM12878 (a B lymphocyte line, closely related to CLL), K562 (a leukemia cell line), HelaS3 (a cervical cancer line), and MCF-7 (a breast cancer line). Performance was measured using auPRC (area under the precision-recall curve), which is the appropriate metric for imbalanced datasets where real interactions are rare. A random classifier scores near zero while a perfect classifier scores 1.0.
For CTCF-associated chromatin interactions, ChINN achieved auPRC scores of 0.77 for GM12878 and 0.75 for HelaS3 in within-sample testing - substantially outperforming predictions based on conventional functional genomics data (auPRC 0.69). The better performance on CTCF data reflects a key finding: CTCF interactions are determined largely by the DNA sequence motif that CTCF recognizes and its orientation, information that is precisely captured in raw sequence but lost when summarized as ChIP-seq peak counts.
A particularly important finding emerged from examining what sequence patterns ChINN learned. For CTCF models, the most important sequence features on the two interacting anchors came from opposite DNA strands in a convergent orientation - exactly the pattern previously observed in experimentally measured CTCF loops. This means ChINN independently rediscovered a known biological rule of CTCF loop formation, providing confidence that it is capturing real chromatin biology rather than statistical artifacts.
For Hi-C interactions across eight cell lines, auPRC ranged from 0.52 to 0.77 within-sample. Critically, ChINN models trained on one cell line generalized well to other cell lines for Hi-C prediction - meaning the sequence rules for chromatin contact are broadly shared across cell types. All Hi-C models captured the CTCF binding motif in their first convolutional layer, as well as general transcription factor binding sites like FOS/AP-1 that are universally found at gene promoters.
The central application of ChINN was to six newly generated chronic lymphocytic leukemia (CLL) patient samples. CLL is the most common adult leukemia in Western countries, and its prognosis is strongly influenced by a genetic marker called IGHV mutation status. In this study, two patients had IGHV-unmutated CLL (uCLL, more aggressive) and four had IGHV-mutated CLL (mCLL, less aggressive). Each patient's leukemia cells were profiled by Hi-C, ATAC-seq (which measures open chromatin regions), and RNA-seq (gene expression).
When the GM12878 Hi-C model (trained on a related B-lymphocyte cell line) was applied to the six CLL samples, it achieved auPRC values ranging from 0.28 to 0.44 - a meaningful predictive signal given that these were completely new patient samples never seen during training. The K562 leukemia model achieved slightly lower auPRC (0.26 to 0.40), suggesting that the B-lymphocyte training data was more relevant to CLL biology.
CLL-specific models trained using each patient's own Hi-C and ATAC-seq data achieved auPRC values of 0.37 to 0.58 - the best performance in the study. When these CLL-trained models were tested across different CLL patients, they showed comparable performance, demonstrating generalizability within the disease context. The CLL models captured CTCF and AP-1 (FOS/JUN family) transcription factor motifs in their first convolutional layer - motifs known to be important in B-cell biology and CLL specifically.
Predicted chromatin interactions were validated experimentally using 4C-seq, a targeted chromosome conformation technique that measures all contacts of a specific genomic region. ChINN correctly predicted validated 4C interactions at several gene loci in MCF-7 cells, including some that were not captured by the Hi-C experiment at those sites - suggesting ChINN may find real interactions that were missed by insufficient sequencing depth in the original Hi-C experiments.
The most clinically significant finding from the CLL analysis was the extent of patient heterogeneity in chromatin interactions. Among the six CLL patients, 37% of Hi-C identified chromatin interactions could be found in only one patient - meaning more than a third of chromatin contacts are unique to individual patients rather than being shared features of CLL as a disease. This level of personalized genome organization has major implications for understanding why CLL patients respond so differently to the same treatment.
When ChINN was applied at scale to a published cohort of 84 CLL patient samples (34 uCLL and 50 mCLL) using only ATAC-seq data as input, it predicted 48,443 CTCF-associated and 23,633 RNA Pol II-associated open chromatin interactions across the cohort. This demonstrates ChINN's ability to generate chromatin interaction maps for large clinical cohorts where Hi-C is not feasible.
The predicted chromatin interactions distinguished uCLL from mCLL patients, and the differences in chromatin architecture aligned with known gene expression differences between the subtypes. Genes expressed at higher levels in uCLL showed more chromatin interactions at their promoters - consistent with the biological principle that active gene promoters accumulate more regulatory contacts. Examples at known CLL prognostic markers including LPL, ZAP70, and CD38 showed subtype-specific interaction patterns.
These findings suggest that the aggressive behavior of uCLL - which has worse clinical outcomes than mCLL - may be partly driven by fundamentally different chromatin architectures that create different patterns of gene activation. ChINN provides a way to study these architectural differences computationally across large patient cohorts, potentially identifying new regulatory targets for leukemia therapy.
ChINN has genuine limitations that the authors acknowledge directly. The auPRC scores for cross-sample prediction on CLL samples (0.26 to 0.44) reflect the inherent difficulty of predicting chromatin interactions in completely new patient samples from DNA sequence alone. When the number of negative samples (non-interacting pairs) vastly outnumbers positives (actual interactions), auPRC scores are structurally low even for good models - a fundamental challenge in this type of prediction task.
The model uses only DNA sequence and open chromatin accessibility as inputs. Adding more data types - such as histone modification profiles, transcription factor occupancy, or methylation data - would likely improve prediction accuracy but at the cost of requiring more expensive experiments. Future versions of ChINN may benefit from multi-modal inputs as multi-omics profiling of cancer samples becomes more routine.
A key practical advantage of ChINN over its competitors is its ability to generate genome-wide predictions rather than predictions limited to specific genomic windows. Tools like Akita and DeepC are restricted to 1 megabase regions and cannot capture long-range interactions. ChINN predicts interactions between any pair of open chromatin regions across the entire genome, making it the only currently available method for genome-wide chromatin interaction prediction from sequence in both ChIA-PET and Hi-C formats.
Looking ahead, ChINN provides a scalable framework for studying how cancer mutations alter genome architecture at a population scale. As more patient cohorts acquire ATAC-seq data - which is cheaper and more scalable than Hi-C - ChINN can convert those datasets into predicted three-dimensional genome maps, enabling studies of chromatin architecture heterogeneity in leukemia and other cancers that would otherwise be impossible.