New ensemble tool sharpens chromatin interaction calls in 4C-seq data

Chromatin, the packed complex of DNA and proteins inside the cell nucleus, is not arranged at random. Its three-dimensional folding helps decide which genes are switched on and off, and disruptions to this architecture have been linked to severe diseases, including cancer. Because many chromatin changes are potentially reversible, researchers see them as promising targets for future therapies. To study these spatial contacts, laboratories use a sequencing technique called Circular Chromosome Conformation Capture Sequencing, or 4C-seq, which starts from one chosen genomic position, known as the viewpoint or bait, and maps all of its interactions across the rest of the genome in a one-versus-all design. A new study now reports that combining several existing 4C-seq analysis tools into a single optimized ensemble can substantially improve the detection of these chromatin interactions.

The work, published in the journal BioData Mining, introduces fourSynergy, a framework built around a weighted-voting ensemble algorithm. Instead of relying on a single interaction caller, the method runs four established 4C-seq tools—r3Cseq, fourSig, peakC, and 4C-ker—and merges their predictions. Each tool receives a weight between zero and 0.5, and the weighted sum of their significance scores produces an ensemble score. A decision threshold of 0.5 then determines whether a fragment is called as an interacting region. Crucially, the weights are not chosen by hand. The researchers used gradient-free optimization strategies from the Nevergrad library, selecting the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to optimize the F1-score and the OnePlusOne algorithm to optimize the area under the precision-recall curve (AUPRC), with a budget of 100,000 trials per optimization loop.

To train and evaluate the ensemble, the team first had to assemble ground truth data. They searched PubMed for 4C-seq experiments published between 2017 and 2025, requiring at least two conditions, two replicates, and at least one biologically validated interaction from mouse or human samples. Validation could come from reciprocal 4C-seq experiments or from other techniques such as CTCF-binding assays or ATAC-seq. This literature review covered 108 experiments and ultimately yielded 20 curated datasets, each containing between one and eight validated interactions. Nine of these datasets belonged to experimentally related experiments and were treated as groups to prevent information leakage during validation. All validated interactions in the collection were located in the near-bait region, meaning the area close to the viewpoint, so the evaluation primarily reflects near-bait interaction calling rather than the detection of more distant contacts.

Before optimization could begin, the heterogeneous outputs of the four base callers had to be harmonized. The researchers built a Snakemake pipeline covering quality control, preprocessing, interaction calling, and postprocessing. FourCSeq, a previously available R/Bioconductor package, was excluded because it is now deprecated and showed low recall in benchmarking. The remaining tools report significance in different ways: r3Cseq outputs q-values, fourSig and 4C-ker assign categorical classes, and peakC returns peak ends for a given false discovery rate setting. The team mapped all of these outputs onto a common four-level scale, where fragments receive scores of 3 for high-confidence candidate interactions, 2 for candidate interactions, 1 for low-confidence candidates, and 0 for non-interacting regions. Because fourSig does not offer built-in replicate integration, the researchers aggregated its scores across replicates, counting an interaction as valid if it appeared in at least half of them.

The benchmarking results showed how much the individual tools vary. On the 20 curated datasets, fourSig achieved mean F1-scores between 0.02 and 0.04, comparable to 4C-ker at 0.03. r3Cseq performed better, with mean F1-scores between 0.12 and 0.16 depending on window size. peakC was the most sensitive to its window parameter, reaching mean F1-scores of 0.18 and 0.19 at the smaller window sizes but dropping to 0.06 at the largest tested window. The AUPRC values followed the same pattern, with peakC achieving the highest values at 0.16 and 0.12 for its smaller windows while fourSig and 4C-ker remained near 0.01 to 0.02. Overall, r3Cseq and peakC outperformed the other two tools in both metrics on this dataset collection.

The ensemble approach clearly surpassed these individual results. When the tool weights were optimized and evaluated on each single dataset—an intentionally optimistic analysis that estimates the upper bound of the method rather than real-world performance—the ensemble reached a mean F1-score of 0.52 and a mean AUPRC of 0.40. A second, similarly overfitted analysis optimizing across the entire dataset collection yielded a mean F1-score of 0.34 and an AUPRC of 0.36, compared with 0.19 and 0.16 for the best single tool. The most important test, however, was the leave-one-group-out cross-validation, in which the ensemble is trained on all groups except one and then evaluated on the held-out group. In this setting, the ensemble achieved a mean F1-score of 0.31 and a mean AUPRC of 0.34, compared with 0.13 and 0.16 for the best-performing single tools per group. Paired Wilcoxon signed-rank tests with Holm-adjusted p-values confirmed that the ensemble’s F1-scores and AUPRC values were significantly higher at the dataset level, with adjusted p-values of 0.005 for both metrics.

The authors are careful to note the limitations of their validation data. Because biological validation of chromatin interactions is time- and cost-intensive, most datasets contain only a handful of validated interactions, and validated non-interactions are usually absent altogether. The researchers describe this situation as reminiscent of positive and unlabeled learning scenarios, in which only positive examples are available. The lack of confirmed negatives makes it difficult to distinguish true negatives from false positives in the confusion matrix, and unvalidated real interactions may be misclassified as false positives, which can underestimate precision and therefore affect precision-related metrics such as the F1-score and AUPRC. The team interpreted its results in the context of these constraints and used two complementary metrics to gain a fuller picture of performance.

The scope of the method is also deliberately limited for now. The consensus approach is designed for near-bait analysis because datasets with validated far-cis and trans interactions are lacking, which made it impossible to optimize the ensemble function for those regions. The authors caution that transferability to more distant cis or trans regions may be limited, since tool behavior varies with distance from the viewpoint. They suggest that distance-dependent weighting could make the approach more versatile, but developing such an extension would require datasets with validated interactions at diverse distances from the viewpoint. Runtime is another practical consideration: fourSynergy required more time than any individual tool, with an additional pipeline overhead of roughly 64 minutes on the tested dataset, although the authors argue that the small size of typical 4C-seq datasets keeps the absolute runtime comparatively short.

To make the ensemble accessible, the researchers packaged it as a complete framework. fourSynergy includes a Snakemake pipeline for reproducible and scalable analysis, an R/Bioconductor package, and an interactive Shiny application for user-friendly access. The framework also offers an optional downstream workflow for differential interaction analysis, which compares chromatin contacts between conditions, such as disease states and healthy controls. This analysis uses DESeq2, based on the observation that 4C-seq read counts follow a negative binomial distribution, and considers differential interactions statistically significant when the adjusted p-value falls below 0.05. On a set of 20 datasets with validated differential interactions, fourSynergy achieved improved overall performance compared with FourCSeq and 4C-ker in terms of mean F1-score and AUPRC. The weighted-voting implementation corresponds to fourSynergy version 1.0.0, available through Bioconductor 3.23 for R 4.6.

Beyond the immediate results, the study makes a broader argument for ensemble thinking in genomics. The authors point out that combining tools is already common practice in other sequencing fields, for example in merging peak callers for ChIP-seq data or variant callers for DNA-seq data, where consensus approaches have been shown to outperform individual callers. Their work extends this principle to 4C-seq, where, to the authors’ knowledge, no ensemble algorithm existed before. Because the optimization is gradient-free, the objective extends readily to additional callers, and a larger ensemble can be built by running each tool multiple times with different window size settings. The researchers also suggest that the weighted-voting strategy based on gradient-free optimization could be transferable to other next-generation sequencing techniques, potentially offering a general recipe for squeezing more reliable predictions out of collections of imperfect tools.

Subject of Research: An ensemble algorithm combining four 4C-seq interaction callers to improve chromatin interaction detection

Article Title: fourSynergy: ensemble-based interaction calling on 4C-seq data using gradient-free optimization

Article References: Wind, S.-M., Plagwitz, L., Dix, J., Heidtmann, G., Heider, D., & Walter, C. (2026). fourSynergy: ensemble-based interaction calling on 4C-seq data using gradient-free optimization. BioData Mining, 19(1), Article 61. https://doi.org/10.1186/s13040-026-00596-4

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00596-4

Keywords: 4C-seq, chromatin interactions, ensemble methods, weighted voting, gradient-free optimization, bioinformatics, gene regulation, cancer research, Snakemake, R/Bioconductor, Shiny app, near-bait interactions