The study, titled “A novel multiomics machine learning signature identifies rapid progression in clinically low risk prostate cancer,” was published on 14 September 2026. Its central finding is both technically elegant and clinically consequential: by integrating multiple layers of molecular data through machine learning and network analysis, the researchers uncovered biological differences within tumors that had been grouped together under a single low-risk label. These differences, the authors suggest, may explain why some patients who appear indolent on paper nonetheless experience disease progression, and they could ultimately point the way toward new therapeutic strategies tailored to the underlying biology of each tumor.

At the heart of the research lies a multi-omics approach, a strategy that has become one of the most powerful frameworks in modern precision medicine. Rather than examining a single type of molecular measurement, multi-omics analysis looks across several distinct but interconnected layers of tumor biology simultaneously. As Arian Lundberg, assistant professor at KTH Royal Institute of Technology and leader of the study, explained in the university’s announcement of the findings, each layer tells a different part of the story. “For example, genomics tells us about DNA alterations, transcriptomics tells us which genes are active, and epigenomics tells us about regulatory changes that can influence gene activity,” Lundberg says.

The crucial insight, and the methodological advance that distinguishes this study, is that these layers are not analyzed in isolation. “Instead of looking at these layers separately, we integrate them to build a more complete, patient-specific picture of tumor biology,” Lundberg says. This integration matters because cancer is not driven by a single mutation or a single misbehaving gene. A DNA alteration may only become consequential when it is amplified by epigenetic changes that silence protective regulatory mechanisms, and those combined effects may only manifest as aggressive behavior when the resulting transcriptional programs rewire the cell’s internal communication networks. By modeling these relationships across layers, the researchers were able to detect patterns of aggressive biology that no single data type would have revealed on its own.

Machine learning and network analysis were central to both the method and the findings. The study’s machine-learning model was developed by a team led by Assistant Professor Golnaz Taheri at the Department of Computational Science and Technology at KTH. The computational challenge is formidable: multi-omics datasets are high-dimensional, meaning they contain measurements of thousands of genes, regulatory marks, and molecular interactions for each patient, while the number of patients available for training is comparatively small. Machine-learning models of the kind developed by Taheri’s team are designed to find meaningful structure within this complexity, identifying signatures that distinguish clinically relevant subgroups of tumors without being overwhelmed by noise or spurious correlations.

Network analysis adds a complementary dimension to this computational toolkit. Where machine learning can identify predictive patterns, network analysis maps the relationships between molecular components, revealing how genes, proteins, and regulatory elements interact as systems rather than as isolated parts. In the context of this study, the combination allowed the researchers to move beyond simple classification and toward a mechanistic understanding of what distinguishes the aggressive molecular profile within the low-risk group. The result is what the authors describe as a novel signature capable of identifying rapid progression in tumors that clinical criteria would have flagged as slow-growing.

The clinical implications of this work are significant. Current risk stratification for prostate cancer relies primarily on clinical measures such as prostate-specific antigen levels, Gleason grading of biopsy tissue, tumor stage, and the number of positive biopsy cores. These measures have served patients well for decades, but they are, by design, proxies. They capture the visible consequences of tumor biology rather than the biology itself. The KTH study demonstrates that within the population of patients who look identical under these clinical measures, there exist molecularly distinct subgroups, some of which harbor features linked to more aggressive disease. For the subset of patients whose tumors carry these hidden signatures, the current classification may underestimate their risk, potentially delaying treatment that could be beneficial.

It is important to note, as the researchers themselves emphasize, that the new method remains at the research stage and is not yet ready for clinical use. The path from a computational discovery to a validated clinical tool runs through extensive independent validation, and the KTH team has been explicit about the next steps. The immediate priority is to test some of the potential therapeutic vulnerabilities suggested by the analysis through laboratory experiments. These experiments will help Lundberg’s team determine whether the molecular changes they identified are useful for risk prediction alone, or whether they may also point toward new therapeutic strategies. This distinction is critical: a biomarker that predicts progression is valuable for guiding surveillance intensity and treatment timing, but a biomarker that reveals a druggable vulnerability could change the treatment options available to patients altogether.

“In the longer term, the goal is to translate these findings into more precise risk stratification, and to potentially identify patients who could benefit from earlier or more targeted treatment,” Lundberg says. That vision reflects a broader shift in oncology toward data-driven precision medicine, in which treatment decisions are informed not only by how a tumor looks under the microscope but by the integrated molecular portrait of the disease in each individual patient. The study is, in Lundberg’s words, “a good example of the type of data-driven precision medicine research we are building at KTH, combining large-scale molecular data with machine learning to address clinically relevant questions in cancer.”

The research infrastructure behind the study is itself noteworthy. Lundberg’s A. Lundberg Lab has launched one of the largest national and international efforts to study the tumor microbiome, drawing on extensive Swedish and global clinical cohorts. Access to large, well-characterized patient cohorts is a prerequisite for the kind of multi-omics machine-learning analysis performed in this study, since models must be trained and tested on sufficient numbers of patients with long-term outcome data to be meaningful. The work was supported by the SciLifeLab-Knut and Alice Wallenberg Data-Driven Life Science (DDLS) program, Vetenskapsrådet, the Swedish Research Council, Prostatacancerförbundet, and Digital Futures, reflecting the collaborative and interdisciplinary funding environment required to sustain research at this scale.

For patients and clinicians, the study’s message is one of cautious optimism rather than immediate change. No patient should expect the new molecular signature to appear in the clinic tomorrow, and active surveillance remains the appropriate standard for men classified as low risk under current guidelines. But the research adds to a growing body of evidence that the molecular era of cancer medicine is moving from promise to practice. If the laboratory experiments now planned by the KTH team confirm that the identified molecular features are both predictive of progression and indicative of therapeutic vulnerabilities, the study could lay the groundwork for a future in which a prostate cancer diagnosis comes with a complete molecular dossier, one that tells clinicians not just how the tumor appears today, but how it is likely to behave tomorrow and what weapons might be deployed against it. In that future, the distinction between a truly indolent tumor and a deceptively aggressive one would no longer depend on proxies, but on the biology itself.

Subject of Research: Multi-omics machine learning identification of aggressive molecular signatures in clinically low-risk prostate cancer

Article Title: Researchers identify aggressive molecular features in low-risk prostate cancer

Article References: Researchers identify aggressive molecular features in low-risk prostate cancer. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: prostate cancer, multi-omics, machine learning, precision medicine, risk stratification, npj Digital Medicine, KTH Royal Institute of Technology, genomics, transcriptomics, epigenomics, network analysis, tumor biology

The research team, working in collaboration with the Swedish University of Agricultural Sciences (SLU), constructed a gene atlas of the Norway spruce cone, a map that records when and where individual genes are switched on as the tree’s seed-bearing reproductive organs take shape. The atlas identifies key genes involved in cone development and, in a surprise to the researchers themselves, uncovered an entirely previously unknown gene, designated DAL55. Together, these results begin to decode the molecular switch that tells a spruce to stop investing in vegetative growth and start building the structures that carry seeds.

What makes the achievement remarkable is the sheer difficulty of the study organism. Conifers such as Norway spruce carry among the largest genomes in the plant kingdom, orders of magnitude bigger than the human genome, and their long life cycles make classical genetic experiments slow and cumbersome. Generations cannot be crossed quickly, and the infrequency of cone production means that even obtaining the right tissue at the right developmental stage is a logistical challenge. These obstacles have left fundamental questions about cone development unanswered, despite the species’ enormous ecological and economic importance across Northern Europe.

The methodological heart of the study is a technology called spatial transcriptomics, developed at KTH and since commercialized as Visium by 10x Genomics. Conventional gene-expression measurements tell researchers which genes are active in a sample, but they destroy the spatial context, blending the signals of many different cell types into a single average. Spatial transcriptomics preserves that context. It allows scientists to see which genes are active in a tissue and exactly where that activity occurs, effectively producing a molecular map layered directly onto the anatomy of the organ being studied.

Stefania Giacomello, associate professor at the Department of Gene Technology at KTH, emphasized the power of this approach for the spruce work. The technology, she said, enables the study of the expression patterns of all genes simultaneously. To achieve this, the researchers cut sections of spruce cone tissue to extraordinary thinness, just 10 micrometers, or 0.01 millimeters, and measured gene activity across those sections. The result is a spatiotemporal record of reproductive development: not merely a list of genes involved in cone formation, but a picture of how genetic programs unfold across the tissue as the cone grows and differentiates.

The practical motivation behind the research is straightforward. Forest tree breeding depends on the ability to produce seed, and in spruce the long juvenile period and erratic flowering make that production inefficient. Breeders seeking to develop improved varieties, whether for timber quality, growth rate or resilience, are constrained by how rarely their breeding material sets cones. By identifying the genes that govern the transition from vegetative to reproductive development, the new atlas provides molecular targets that could eventually help breeders induce or accelerate cone formation, shortening breeding cycles that currently stretch across decades.

Giacomello pointed to the climate dimension of this goal. By learning more about the molecular mechanisms that regulate cone formation, she said, the team hopes to accelerate breeding efforts and facilitate the production of climate-adapted spruce seedlings for forest owners across the country. As warming shifts the conditions under which Northern European forests grow, the demand for seedlings bred from trees suited to future climates is expected to rise. A faster, more controllable route to reproductive maturity in spruce would directly support that effort, allowing desirable genetic combinations to be captured in seed more quickly than natural flowering schedules permit.

The study also reaches beyond forestry into one of the deepest questions in plant evolutionary biology. Flowering plants, the angiosperms, dominate the modern landscape, but cone-bearing trees such as spruce belong to the gymnosperms, a far older lineage. Both groups descend from a common seed-plant ancestor, and a central open question is whether some of the genetic mechanisms that control reproduction in flowering plants, including familiar crops and trees such as apple, were inherited from that shared ancestor rather than invented independently along the flowering lineage. Comparing the gene programs active in spruce cones with those known from flowers offers a way to test this.

Jens Sundström, a researcher in plant biotechnology at SLU, framed the findings in these evolutionary terms. The results, he said, improve understanding of the evolutionary processes that contributed to the development of all living seed plants, including both flowering plants and conifers. In other words, the cone atlas is not only a breeding tool; it is a window into the ancient genetic toolkit that has governed seed production for hundreds of millions of years. Shared patterns of gene activity between cones and flowers would suggest deep conservation of reproductive control, while differences would point to lineage-specific innovations in each group.

Sundström also stressed the relevance of the work to economies that depend on forestry, singling out Sweden and Finland, where Norway spruce underpins the sector. For those countries, the gap between a fundamental genetic discovery and a commercial seed orchard may still be long, but the direction of travel is clear. A species that once resisted molecular study because of its giant genome and glacial life cycle has now yielded a high-resolution map of its reproductive development, complete with a new gene to investigate. The decades-long wait for a spruce to reproduce may not be shortened overnight, but the genetic instructions behind that wait are, for the first time, coming into view.

Subject of Research: Genetic regulation of the vegetative-to-reproductive transition and cone development in Norway spruce

Article Title: Why do spruce take decades to reproduce? A genetic study offers some clues

Article References: Why do spruce take decades to reproduce? A genetic study offers some clues. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Norway spruce, spatial transcriptomics, cone development, gene atlas, conifer reproduction, forestry breeding, DAL55, gymnosperms, plant evolution, KTH Royal Institute of Technology, SLU, climate-adapted seedlings