For decades, geneticists have relied on the classical genetic code to interpret how DNA sequences translate into proteins. However, a profound mystery persisted: the regulatory framework governing gene expression plasticity remained elusive. While regulatory elements like promoters are known to modulate gene activity, the complex ‘grammar’—or regulatory syntax—that orchestrates precise gene toggling had not been deciphered. This regulatory system is responsible for spatiotemporal gene expression, determining cell identity, behavior, and response to environmental cues.

The urgency of decoding this genomic control system cannot be overstated, especially given that many cancer-related mutations reside in non-coding regions traditionally deemed as “junk DNA.” These mutations often disrupt gene regulation and contribute to tumor development and progression. Historically, interpreting the pathogenic potential of such mutations was a major bottleneck in cancer research. PARM directly addresses this challenge by providing a sophisticated computational tool that can interpret regulatory DNA sequences and predict their effects on gene activity with exceptional accuracy.

The development of PARM was made possible through a collaborative initiative, the PERICODE project, which united seven research groups under the Oncode Institute. Utilizing cutting-edge experimental techniques pioneered in the Bas van Steensel laboratory at the Netherlands Cancer Institute (NKI), researchers employed massively parallel reporter assays (MPRA) to generate millions of quantitative measurements. These experiments systematically tested how myriad short DNA sequences influenced gene expression levels in specific cell types, thereby creating an unprecedented data repository linking promoter architecture to functional output.

Yet, possessing vast data alone is insufficient for biological insight. Here, Jeroen de Ridder’s group at UMC Utrecht harnessed advanced artificial intelligence algorithms to model these experimental results. Unlike conventional AI models that rely on imperfect proxy data, PARM benefited from precisely engineered, high-fidelity datasets explicitly crafted for deciphering gene regulation. This intentional synergy between experimental design and machine learning empowered the creation of an ultra-efficient model fine-tuned for specific cellular environments. By training on meticulously controlled datasets, PARM captures nuanced, cell-type specific regulatory logics that previous models missed.

Demonstrating extraordinary predictive power, PARM elucidates how gene regulation varies not only between cell types but also dynamically changes under environmental stimuli, such as exposure to drugs or hormones. This dynamic modeling revealed the detailed architecture of regulatory elements—effectively exposing each gene’s “on” and “off” control switches and their combinatorial interactions. Importantly, the scientific team subjected every prediction to rigorous experimental validation, assuring the robustness and biological fidelity of the model’s insights.

PARM also innovates through its remarkable computational efficiency. Previous state-of-the-art models, like Google DeepMind’s AlphaGenome, while powerful, demanded colossal computational resources making them less accessible to many research laboratories worldwide. PARM’s architecture requires approximately one thousand times less computing power, making it achievable for typical academic environments. This efficiency was achieved without sacrificing performance, meaning researchers worldwide can now simulate complex regulatory landscapes using modest laboratory setups and conventional computing hardware within a single day.

This breakthrough has profound implications for cancer biology and therapeutic development. By enabling accurate prediction of regulatory mutation impacts in specific cell types and conditions, PARM opens novel avenues for precision oncology, such as designing patient-specific diagnostics and stratified treatments. The ability to forecast how tumor cells may adapt or resist therapeutics at the level of gene regulation provides an invaluable resource for drug discovery and personalized medicine.

The success of PARM underscores the power of multidisciplinary collaboration bridging genomics, computational biology, and experimental biophysics. Funded by notable institutions such as the Oncode Institute and the AVL Foundation, this collective effort amalgamated expertise from Bas van Steensel’s group at NKI, Jeroen de Ridder’s team at UMC Utrecht, and several other leading genomic research labs. Such integration of experimental high-throughput approaches with deep learning signifies a paradigm shift in decoding biological complexity.

Importantly, PARM’s design also bridges the gap between scalability and interpretability, two features often mutually exclusive in AI. By tailoring its predictive models to highly specific cellular states, PARM manages to retain mechanistic interpretability—insight into the regulatory grammar—while scaling analyses across millions of variants. This combination promises to accelerate functional genomics research across a broad spectrum of diseases and biological systems beyond oncology.

Looking forward, the research community anticipates PARM’s versatility to expand substantially. Researchers can now systematically map gene regulatory changes across diverse human tissues, developmental stages, and disease contexts. The model’s adaptability to incorporate different stimulus-response patterns also sets the stage for unraveling how environmental factors and pharmacological agents reshape epigenetic landscapes, further enriching our understanding of gene control in health and disease.

As the frontiers of genomics advance deeper into the realm of regulatory DNA, tools like PARM will be indispensable for translating vast sequence data into actionable biological knowledge. This model not only demystifies how non-coding DNA dictates cellular phenotypes but also empowers a new generation of genomic medicine that integrates predictive, customizable insights into clinical workflows.

In sum, the advent of PARM signifies a scientific milestone: the ability to ‘read’ the language of gene regulation at unparalleled resolution and scale. By transforming gene regulatory decoding from an enigmatic black box into an interpretable and computable framework, PARM promises to accelerate breakthroughs in cancer biology, therapeutic design, and fundamental genomics, heralding a new era of precision in biomedical science.

Subject of Research: Cells

Article Title: Regulatory grammar in human promoters uncovered by MPRA-based deep learning

News Publication Date: 3-Feb-2026

Web References:

• Nature article
• PARM model portal

References:
Bas van Steensel, Jeroen de Ridder, et al. Regulatory grammar in human promoters uncovered by MPRA-based deep learning. Nature, 2026. DOI: 10.1038/s41586-025-10093-z

Image Credits: ©Netherlands Cancer Institute / Sanne Hijlkema

Keywords: Gene expression, Machine learning

Nucleosomes are the fundamental units of chromatin, composed of DNA wrapped around histone proteins. This structure plays a crucial role in regulating gene expression by controlling the accessibility of DNA to transcriptional machinery. The placement and occupancy of nucleosomes can dramatically influence transcription, making the accurate prediction of their positioning a compelling challenge. The advent of models like HDGS-Net could catalyze not only basic genomic research but also applied fields such as synthetic biology and gene therapy.

Researchers Shi, Wang, and Teng, along with their colleagues, have meticulously engineered HDGS-Net to learn directly from high-dimensional genomic data. Utilizing advanced deep learning techniques, the model efficiently captures intricate patterns tied to nucleosome occupancy. By merging dilated convolutions with gated mechanisms, the architecture allows for finer control over information passage, enhancing both the accuracy and computational efficiency of predictions concerning nucleosome placements.

This sophisticated approach emerges from recognizing that traditional models often succumb to limitations due to their inability to factor in long-range dependencies and interactions present in genomic datasets. The hybrid nature of HDGS-Net enables it to consider broader spatial contexts, thereby increasing the model’s predictability across diverse genomic regions. The result stands not only in improved accuracy but also in the model’s generalizability across various organisms, opening doors for extensive comparative genomic studies.

Training the HDGS-Net model involved a comprehensive dataset encompassing a wide array of epigenomic signals, with particular focus placed on features that influence nucleosome positioning. This process engaged both supervised and unsupervised learning strategies, allowing the model to develop a robust understanding of the underlying biological processes. The flexibility of this hybrid architecture significantly enhances its ability to adapt and learn from varying data conditions, promising exceptional outcomes in nucleosome modeling.

Moreover, the researchers conducted robust validation of HDGS-Net, employing several benchmark datasets against which they meticulously compared their predictions. These tests yielded remarkable improvements, showcasing HDGS-Net’s ability to outperform traditional nucleosome prediction methods. Statistical analyses demonstrated that the model could reduce prediction errors significantly while simultaneously enhancing the biological relevance of its outputs.

Beyond its technical merits, the implications of HDGS-Net resonate deeply within the broader scientific community. As researchers grapple with the complexities of genetic regulation and chromatin dynamics, tools that provide clear insights into nucleosome occupancy are invaluable. HDGS-Net stands to not only enrich our understanding of gene regulation but also expedite the discovery of novel therapeutic targets by elucidating epigenetic modifications that influence disease states.

Furthermore, the model’s design encourages future enhancements, allowing for integration with multi-omics data. This capability paves the way for complex models that could incorporate transcriptomic, proteomic, and even metabolomic data, establishing a more holistic view of the genomic landscape. By creating a comprehensive mapping of the epigenetic landscape, researchers can cultivate insights that lead to more precise and personalized medical treatments.

HDGS-Net also carries significant implications for the future of genomic research. As more researchers adopt artificial intelligence and machine learning methodologies, the accumulation of knowledge from tools like HDGS-Net will propel the field forward. By fostering collaborative environments where bioinformaticians, geneticists, and machine learning specialists can interact, the potential for revolutionary discoveries becomes ever more attainable.

Furthermore, the ease of access to such advanced computational tools is crucial for democratizing genomic research. The availability of HDGS-Net’s predictions can potentially bolster research efforts in laboratories worldwide, including those in resource-limited settings. This democratization of technology reinforces the notion that breakthroughs in genetics should not be confined to well-funded institutions.

In the broader context of technological advancement, HDGS-Net epitomizes how artificial intelligence can yield significant strides in specialized scientific fields. It serves to bridge the gap between computational techniques and biological inquiry, illustrating the profound potential of interdisciplinary collaboration in driving scientific innovation. As researchers delve deeper into the functionalities of HDGS-Net, a cascade of discoveries across diverse biological disciplines is poised to emerge.

The introduction of HDGS-Net is poised to become a cornerstone in the fields of computational genomics, providing researchers with a powerful tool to explore the complexities of nucleosome occupancy and its implications on gene regulation. As the exploration of genomic interactions continues to unfold, the future looks bright for computational models that harness cutting-edge technologies to unlock the mysteries of the biological world.

In this exciting age of genomic research, HDGS-Net stands as a hallmark of innovation, paving the way for a deeper understanding of the fundamental mechanics governing life at a molecular level. As human capacity to decode genetic information expands, the ramifications of such advancements ripple through medicine, biotechnology, and beyond, shaping the very fabric of future biological discoveries.

As the team behind HDGS-Net continues to refine and disseminate their findings, the scientific community awaits with bated breath at the prospect of further advancements. The true potential of such models lies not only in their capacity to predict nucleosome occupancy but also in their ability to inspire new generations of researchers to explore, innovate, and transform the possibilities innate within genomic science.

Subject of Research: Nucleosome occupancy prediction

Article Title: HDGS-Net: nucleosome occupancy prediction based on a hybrid dilated gated separable convolutional neural network

Article References:

Shi, F., Wang, M., Teng, Z. et al. HDGS-Net: nucleosome occupancy prediction based on a hybrid dilated gated separable convolutional neural network.
BMC Genomics (2026). https://doi.org/10.1186/s12864-026-12523-2

Image Credits: AI Generated

DOI: 10.1186/s12864-026-12523-2

Keywords: Nucleosome occupancy, computational genomics, artificial intelligence, hybrid dilated gated separable convolutional neural network, gene regulation

Traditional genetic research has predominantly focused on the approximately 2% of the genome that encodes proteins—those molecular workhorses indispensable for myriad biological functions. Yet, extensive genome-wide association studies (GWAS) have unearthed compelling evidence that variants lurking outside these coding sequences wield significant influence over health and disease. These noncoding variants often operate within regulatory domains, orchestrating when and how genes are expressed through the modulation of DNA-protein interactions. However, deciphering this “regulatory code” has proved profoundly challenging due to the complex interplay of transcription factors—the proteins that bind specific DNA sequences to control gene activity—and the subtle nature of their genomic footprints.

The pioneering study addresses a critical bottleneck in genetic analysis: distinguishing causative variants from a constellation of nearby candidates within noncoding loci associated with disease. Because many of these variants cluster around transcription factor binding motifs, the precise delineation of where these proteins latch onto the genome can illuminate which variant actually disrupts gene regulation. The research hinges on a nuanced understanding of the transcription factor “footprint,” a term denoting the localized suppression of DNA accessibility at binding sites following protein attachment. This footprint acts like a molecular signature, detectable through sophisticated sequencing techniques, that reveals the exact coordinates where transcription factors exert their influence.

To capture this elusive footprint, the researchers employed ATAC-seq (Assay for Transposase-Accessible Chromatin using sequencing), a powerful experimental technique that maps open, accessible regions of the genome amenable to protein binding. By coupling ATAC-seq data from 170 distinct human liver tissue samples with PRINT, a novel deep learning framework designed to discern subtle DNA-protein interactions, the team generated an unprecedented high-resolution map of transcription factor footprints. These maps enabled the identification of “footprint quantitative trait loci” (fpQTLs)—specific genomic sites where variation in DNA sequence correlates with altered footprint patterns, implicating differential transcription factor binding attributable to genetic variants.

This integrative methodology represents a quantum leap in resolving the “needle in the haystack” problem that has bedeviled geneticists. By refining the localization of functional regulatory variants within broad disease-associated regions, scientists can now approach the heretofore nebulous noncoding genome with surgical precision. Beyond liver tissue, the researchers envision extending this approach to a variety of organs and cellular contexts, which could vastly accelerate the identification of disease-driving variants across diverse conditions including metabolic disorders, psychiatric illnesses, and beyond.

Senior author Dr. Struan F.A. Grant eloquently analogized the challenge, likening it to a police lineup where multiple suspects—genetic variants—appear similar, yet only one bears culpability. Through the ability to map precise transcription factor footprints influenced by DNA sequence changes, the team has effectively enhanced the investigative toolkit, enabling confident identification of the “culprit” variants most likely to contribute to pathogenesis. This capability promises to fill a critical knowledge gap in the translation of GWAS findings into actionable biological insight.

The research initiative was made possible in part by the multidisciplinary integration of computational modeling, high-throughput experimental genomics, and biostatistical rigor. PRINT, the deep learning algorithm central to footprint detection, exemplifies the transformative power of artificial intelligence in genomics, as it can parse complex and noisy biological data far beyond human capacity. The application of such computational sophistication to ATAC-seq datasets has dissected the nuanced interplay between nucleotide variation and transcription factor binding strength, uncovering patterns invisible to conventional analyses.

Furthermore, the study’s focus on liver samples holds particular pertinence given the organ’s central role in metabolism, detoxification, and disease susceptibility. By characterizing liver-specific fpQTLs, the team has laid a vital foundation for understanding how regulatory variants may influence conditions such as metabolic syndrome, liver fibrosis, and other prevalent disorders. Moreover, the principles delineated here are broadly transferable, as regulatory mechanisms mediated by transcription factors are a universal feature of cellular function.

First author Max Dudek highlighted the profound implications of these findings for precision medicine. The capacity to accurately pinpoint noncoding variants that actively modulate gene expression shifts the paradigm from correlation to causation in genetic risk assessment. With ongoing expansion to larger cohorts and diverse tissue types, this approach might ultimately enable bespoke intervention strategies, wherein patients are treated based on the precise regulatory variants driving their disease—a leap towards truly personalized therapeutics.

In addition to its clinical potential, this research also advances fundamental biological understanding. Noncoding DNA has traditionally been understudied relative to coding counterparts, yet it harbors myriad regulatory elements governing cellular identity and responsiveness. The fine-scale mapping of transcription factor footprints offers a window into the dynamic regulatory architecture of the genome, elucidating how genetic variation sculpts the gene expression landscape underpinning health and disease.

Funding for the investigation was provided by prestigious institutions including the National Science Foundation Graduate Research Fellowship Program and multiple National Institutes of Health grants. The confluence of public investment and academic ingenuity underscores the societal value attributed to decoding the regulatory genome, which stands as a frontier of modern biomedical science.

Looking ahead, the researchers aspire to augment their footprint QTL atlas with integrative multi-omics data, including chromatin conformation, epigenetic modifications, and transcriptomics, to construct a holistic model of gene regulation perturbed by noncoding variants. Such composite frameworks could dramatically enhance predictive modeling of disease risk and responsiveness to therapy.

This seminal work, published in the American Journal of Human Genetics on April 17, 2025, marks a watershed moment in genomics research. By illuminating the shadows of the noncoding genome, the study opens new avenues for discovery and innovation, promising to transform our approach to diagnosing, preventing, and treating a kaleidoscope of human diseases through the lens of genetic regulation.

Subject of Research: Human tissue samples
Article Title: Characterization of non-coding variants associated with transcription factor binding through ATAC-seq-defined footprint QTLs in liver
News Publication Date: 17-Apr-2025
Web References: Children’s Hospital of Philadelphia, Perelman School of Medicine at the University of Pennsylvania, American Journal of Human Genetics
References: Dudek et al, “Characterization of non-coding variants associated with transcription factor binding through ATAC-seq-defined footprint QTLs in liver.” Am J Hum Genet. Online April 17, 2025. DOI: 10.1016/j.ajhg.2025.03.019.
Keywords: Genetic variation, Psychiatric disorders, Discovery research, Basic research