The central challenge the researchers set out to address is one that has frustrated computational biologists for years: systematically identifying transcriptional and epigenetic regulators, or TERs, in myeloid leukemia. Most existing approaches rely on a single type of data, such as gene expression alone, and they struggle with one of the defining features of human genomic architecture, namely long-range regulation. Enhancers, the DNA elements that boost gene activity, can sit hundreds of thousands of base pairs away from the promoters they control, often separated by vast stretches of non-coding sequence. Without a reliable way to link enhancers to their target promoters, regulatory analysis remains incomplete, and candidate disease drivers slip through the net.

TERfinder takes a fundamentally different approach by integrating multi-omics features within a deep learning architecture. Rather than asking what a single data type says about a genomic region, the framework fuses information across layers of cellular regulation, including histone modification signals that mark active or primed regulatory elements. The model’s task is to predict enhancer-promoter interactions, or EPIs, the physical and functional connections through which regulatory elements control gene expression. By learning the patterns that distinguish true interactions from chance proximity, the model builds a map of the regulatory circuitry operating in leukemia cells.

The performance figures reported in the study are striking. On held-out chromosomes, meaning genomic regions the model had never seen during training, TERfinder achieved an area under the receiver operating characteristic curve of 0.9644 and an area under the precision-recall curve of 0.9584. These numbers indicate that the framework can discriminate interacting enhancer-promoter pairs from non-interacting ones with remarkable accuracy even on unfamiliar territory. Crucially, the authors compared their model against baseline versions stripped of either the autoencoder component or the histone features, and statistical testing using the DeLong method confirmed that the full model significantly outperformed these ablated versions, with p-values below 0.01. That comparison matters because it demonstrates that the multi-omics integration is not decorative; each layer of information contributes measurably to predictive power.

With the interaction predictions in hand, the researchers turned to the question of which regulators matter most in myeloid leukemia. Motif enrichment analysis, which searches for the DNA-binding signatures of transcription factors within predicted regulatory regions, flagged members of the C/EBP and ETV families as candidate master regulators. These findings were not left as abstract computational output. The team performed single-cell regulon analysis, examining regulatory network activity at the resolution of individual cells, and confirmed that these transcription factor families are indeed active in acute myeloid leukemia progenitor populations, the immature cells thought to seed and sustain the disease.

The single-cell analysis yielded one of the study’s most compelling results: regulatory networks centered on the transcription factors SPI1 and CEBPA were found to be active in AML blasts, the malignant cells that crowd out healthy blood production in patients. SPI1, also known as PU.1, is a well-established lineage-determining factor in myeloid cells, and CEBPA is a transcription factor whose mutations are a recognized driver of a distinct AML subtype. The finding that their regulon activity, the collective output of the gene networks they control, is associated with poor overall survival elevates these networks from biological curiosities to clinically meaningful signatures of aggressive disease.

Perhaps the most translationally significant result is a compact four-gene expression signature comprising SPI1, CEBPA, MYC, and PTPN6. When the researchers stratified AML patients according to the activity of this signature, they separated them into high-risk and low-risk groups with a log-rank p-value below 0.01, a standard statistical threshold in survival analysis. A four-gene panel is the kind of lean biomarker that could plausibly be adapted into clinical assays, offering a way to refine risk stratification beyond the cytogenetic and mutational markers currently in use. MYC, the famed oncogenic transcription factor, and PTPN6, a phosphatase gene involved in immune cell signaling, round out a signature that spans both classical cancer biology and myeloid lineage regulation.

The methodological significance of TERfinder extends beyond leukemia. The framework demonstrates a general strategy for regulatory inference in which deep learning serves as the integrator of heterogeneous genomic signals, converting the raw complexity of multi-omics data into interpretable predictions about which regulatory elements talk to which genes. The authors describe it as a framework for multi-omics regulatory inference and candidate transcription factor identification in myeloid leukemia, and the architecture’s reliance on features that are broadly measurable, such as histone marks and chromatin accessibility proxies, suggests the approach could be retrained for other cancer types or developmental contexts where enhancer-promoter wiring goes awry.

The study also reflects a broader shift in how computational genomics is approaching cancer. For decades, the field has catalogued mutations and expression changes, but the regulatory layer, the system that decides when and how strongly genes are turned on, has been harder to read. Tools like TERfinder represent an attempt to make that layer computationally legible. By predicting enhancer-promoter interactions with high accuracy, identifying the transcription factors that occupy the resulting regulatory hubs, and connecting those hubs to patient survival, the framework closes a loop from DNA sequence to clinical outcome that few methods have managed to traverse.

There remain, of course, the usual caveats that accompany any computational framework. The model’s predictions are inferences derived from training data, and experimental validation of individual enhancer-promoter links will be needed before the full interaction map can be treated as ground truth. The four-gene signature, while statistically robust in the analyzed cohorts, would require independent validation before clinical deployment. Yet the open-access publication, the extensive supplementary tables released alongside the paper, and the framework’s demonstrated superiority over simplified baselines give the work a solid foundation. For a disease as relentless as acute myeloid leukemia, where five-year survival remains unacceptably low for many patient groups, a tool that can spotlight the regulatory circuits sustaining the cancer, and distill them into a measurable signature of risk, is a development worth watching closely.

Subject of Research: Deep learning-based multi-omics analysis of transcriptional and epigenetic regulation in myeloid leukemia

Article Title: TERfinder: A deep learning framework for multi-omics regulatory analysis in myeloid leukemia

Article References: Xu, M., Xu, X., Yufei, L., Zhang, G., Liu, J., Cui, T., Guo, B., Huang, J., & Li, C. (2026). TERfinder: A deep learning framework for multi-omics regulatory analysis in myeloid leukemia. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06627-5

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06627-5

Keywords: TERfinder, deep learning, multi-omics, enhancer-promoter interactions, acute myeloid leukemia, transcriptional regulation, epigenomics, gene regulatory networks, SPI1, CEBPA, biomarker, single-cell analysis