How Close Are We to Treating Cancer with MicroRNAs in the Age of AI?
In this article
Oncology's Quiet Revolutionaries in the Age of AI: MicroRNAs
DOI: 10.1038/s41571-025-01114-x · Publication date: 15 January 2026
Institution: MD Anderson Cancer Center (Houston, US), Charité (Berlin), German Cancer Consortium (DKTK)
In 1993, in Caenorhabditis elegans worms, the discovery of a small gene-regulating molecule began a three-decade journey with the potential to transform oncology through a quiet revolution: microRNAs (miRNAs). These noncoding RNA molecules, just 18-25 nucleotides long, are the true "conductors" that manage the final fine-tuning of gene expression. They are abnormally expressed in many cancers — some act as oncogenes and promote tumors, while others function as tumor suppressors. Despite thirty years of intensive research, no miRNA biomarker has yet become established in clinical practice. However, in Nature Reviews Clinical Oncology, this comprehensive review published in January 2026 makes an important observation: the wave of artificial intelligence (AI) and machine learning has begun to overcome the barriers that held miRNAs back for years. The team led by George Calin at MD Anderson carefully maps this transformation, from miRNA biology to clinical translation, and what can be expected in the coming decade.
What Is a MicroRNA? A Revolution in Three Sentences
MicroRNAs are very short RNA molecules produced by the cell using its own genes. Their task is not to become proteins; instead, regulating gene expression by silencing other mRNAsis what they do. A single miRNA can target hundreds of different mRNAs at once — so a small molecule can control a large gene network. This extensive regulatory power explains both the importance and the complexity of miRNAs in cancer.
A DUAL ROLE: THE SAME miRNA, DIFFERENT FACES IN DIFFERENT CONTEXTS
One of the most surprising aspects of miRNA biology is context dependence. The same miRNA can act as a tumor suppressor in one cell type and an oncogene in another. For example, miR-200cfunctions as a tumor suppressor in invasive breast and colorectal cancer (inhibiting EMT by suppressing ZEB1 and ZEB2 mRNAs); however, high expression of miR-200c in the initial stages of breast cancer can increase cell proliferation and chemotherapy resistance. This paradox shows why placing miRNAs into "good" or "bad" categories is inadequate — the real story depends on which cell they are in, at which moment, and within which regulatory network.
The same regulatory molecule can promote or suppress a tumor depending on the cellular context
The Unfulfilled Side of a Thirty-Year Promise: The Dream of a Single miRNA Biomarker
The initial hope was enormous: could a single miRNA measured in circulating blood detect tumors not yet medically visible? That this simple dream had not come true was demonstrated by thirty years of research. A single circulating miRNA biomarker could not achieve the diagnostic accuracy needed to detect cancer. The reasons are multilayered: miRNA levels are influenced by age, sex, ethnicity, comorbidities, and the tumor microenvironment; a single miRNA lacks sufficient sensitivity and specificity to signal amid all these variables.
But the field accepted this failure and changed direction. The central message emphasized in the review by Calin and colleagues is this: cancer diagnosis should use miRNA panels rather than a single miRNA. This is where artificial intelligence comes in — to evaluate hundreds or even thousands of miRNAs simultaneously and create a "signature."
How Artificial Intelligence Is Transforming miRNA Research: 25 Years of Acceleration
One of the review's most illuminating aspects is its chronological mapping of the integration of artificial intelligence and machine learning into oncology research. The journey from early machine-learning applications for cancer classification in 1999 to foundation models under investigation in 2024 offers important clues about the direction of clinical oncology's evolution.
25 Years of Artificial Intelligence and Machine Learning in Oncology
The journey from early SVM applications to foundation models
Striking Results from AI-Assisted miRNA Studies
Several studies featured in the review demonstrate the real potential of combining miRNA + artificial intelligence:
Perhaps the most surprising of these figures is the last. In self-organizing map (SOM) analysis — analysis using a self-organizing artificial neural network — even in serum samples collected 32 years before cancer, the possibility that the miRNA signature had already changed was demonstrated. Blood samples collected years earlier from 30 patients who developed lung, breast, or colon cancer carried miRNA signatures distinguishable from the control group. This finding is exciting for cancer prevention (could very early diagnosis be possible?), but also raises unsettling biological questions: do the changes that trigger cancer really begin decades beforehand?
EMERALD: Lessons from an AI-Assisted Multicenter miRNA Study
The EMERALD study, specifically highlighted in the review, is a large-cohort study conducted in four countries. Overexpressed in patients with esophageal adenocarcinoma (EAC), a 6-miRNA panel (miR-106b-5p, miR-146a-5p, miR-15a-5p, miR-18a-5p, miR-21-5p, and miR-93-5p) was discovered and validated using machine-learning models. This panel could distinguish patients with EAC or premalignant Barrett's esophagus from controls with an AUC of 0,919.
Even more impressively, this 6-miRNA panel could identify individuals with Barrett's esophagus among people with gastroesophageal reflux symptoms with an AUC of 94,8%. The clinical implication: identifying, through blood, the at-risk group of reflux patients who develop Barrett's esophagus — potentially reducing the burdens of endoscopy in terms of time, cost, and patient comfort.
miRNA Therapies: Bold Trials and Painful Lessons
The idea of using miRNAs as treatment is very appealing — targeting a single molecule to reset a broad gene network. But reality proved far more complex. According to ClinicalTrials.gov, as of June 2025, in oncology, 472 clinical trials related to miRNA were registered; however, none has yet become an approved treatment. One of the most instructive sections of the review is its account of the clinical research landscape, honestly analyzing failure:
The first human trial, MRX34 (a synthetic miR-34a mimic), was tested in 85 patients with advanced cancer during 2014-2016. The outcome:
- Partial response in 3 patients
- Stable disease in 16 patients (median 19 weeks)
- Binding to the target miRNA was demonstrated
- ⚠ The trial ended early with 4 deaths
- Cause: severe immune-mediated adverse effects (similar to a cytokine storm)
The miR-155 inhibitor cobomarsen (an LNA-modified antisense oligonucleotide), in cutaneous T-cell lymphoma (CTCL):
- Phase 1 was successfully completed in 66 patients
- Clinical response in >50% of patients
- Dose-dependent lesion reduction
- No dose-limiting toxicity
- SOLAR phase 2 was initiated...
- ⚠ ...but was terminated early by the sponsor's commercial decision (not a safety/efficacy problem)
The MRX34 tragedy taught miRNA therapy research an important lesson: double-stranded RNA mimics and liposomal carriers can unpredictably stimulate the immune system. For this reason, the field moved toward more selective and less immunogenic LNA-modified antisense inhibitors. Cobomarsen is the product of this change of direction.
The Landscape of Ongoing Clinical Trials
🟢 LNA-i-miR-221 (for HCC, phase 1 completed)
A miR-221 inhibitor was tested in patients with hepatocellular carcinoma. There was no grade ≥3 toxicity, inhibition of the target miRNA was demonstrated, and a partial response was observed.
🟡 INT-1B3 (miR-193a-3p mimic, phase 1 terminated early)
The LNP-formulated mimic tested in solid tumors was terminated early because of insufficient funding — a financial reason, not a scientific one.
🟡 TargomiRs (mesothelioma + NSCLC, phase 1 completed)
EGFR-targeted delivery of a miR-16 mimic using EnGeneIC bacterial minicells; disease stabilization and rare tumor shrinkage were observed.
🟡 TTX-MC138 (miR-10b inhibitor, triple-negative breast cancer + liver metastasis)
An iron oxide nanoparticle-formulated inhibitor; phase 0/1 completed, phase 1/2 about to begin. No dose-limiting toxicity.
Barriers to Clinical Translation: Three Major Fronts
- One miRNA targets hundreds of mRNAs
- One mRNA is regulated by many miRNAs
- Context-dependent dual roles
- Compensatory mechanisms (lncRNA, circRNA)
- Multiple regulatory feedback loops
- RNA degradation and side effects of extraction
- Incompatibility between platforms (qPCR, sequencing, microarray)
- Lack of a normalization strategy
- Exclusion of isomiRs from standard analysis
- High false-positive rate of target-prediction algorithms
- The challenge of targeted drug delivery
- Immunogenicity (TLR activation, cytokine storm)
- Off-tissue side effects
- Insufficient endosomal escape
- Absence of a standard pharmacodynamic endpoint
New Approaches: AI and Innovative Delivery Systems
Despite these barriers, the field is evolving rapidly. The review by Calin and colleagues highlights several promising approaches that could enable future clinical translation:
Developed with inspiration from the success of antibody-drug conjugates (ADCs), antibody-oligonucleotide conjugates (AOCs) are attracting particular attention. This approach has the potential to increase effectiveness and reduce toxicity by enabling the specific delivery of a miRNA mimic or inhibitor to particular cancer cells. Current AOCs mostly carry siRNA or antisense oligonucleotides, but the same principle can also be adapted to miRNA therapies.
The Future Turning Point: Spatial miRNA Mapping
Patho-DBiT AND NEXT-GENERATION SPATIAL TRANSCRIPTOMICS
Perhaps the review's most exciting section concerns new technologies that make it possible to map the spatial distribution of miRNAs within a tumor. Patho-DBiT (pathology-compatible deterministic barcoding in tissue) and similar platforms can reveal miRNA distribution even in paraffin-embedded tumor samples by overcoming technical barriers such as mature miRNAs lacking tails (lacking polyadenylation). The first application involved reconfirming miR-142, miR-150, miR-146a, and miR-21 in extranodal marginal zone B-cell lymphoma. Adapting this technology to common tumors such as lung, breast, and colorectal cancer will map the role of miRNAs in the tumor microenvironment along the axis of "tumor cells vs T cells vs stroma" — so the question of where they have an effect will be answered more clearly in the near future.
Previously discussed on drozdogan.com, the immunosuppressive microenvironment of tumor-draining lymph nodes is also influenced by miRNAs — particularly through their effects on increased Tregs and dendritic cell maturation. Spatial transcriptomics technologies could map the molecular pathways of immune manipulation in lung cancer TDLNs.
Implications for Practice in Turkey and Strategic Positioning
This review has several important messages for oncology research and practice in Turkey. First, a national biobank infrastructure for miRNA biomarker research is becoming increasingly critical. Turkey's participation in multicenter, multinational studies such as EMERALD could be strategically important, especially for esophageal and gastric adenocarcinoma (high incidence in Turkey). The Turkish population's miRNA signatures may differ from Western cohorts, making local validation studies essential.
Second, integrating artificial intelligence and machine-learning competencies into oncology research is an important area of opportunity for Turkey. Through cooperation between TÜBİTAK and major university oncology centers, shared platforms could be developed for AI-assisted miRNA biomarker discovery. Turkish scientists have much they can initially do, particularly with models based on freely available data (such as TCGA and GEO).
Third, how will liquid biopsy be positioned within the SGK reimbursement system? At present, circulating tumor DNA tests receive limited reimbursement; miRNA panels will also become the subject of a similar debate as their evidence base develops. Proactive monitoring of developments in this field by health-policy decision-making bodies could accelerate our patients' access to innovative diagnostic tools.
Fourth, the concrete message for clinical oncologists is: miRNA-based therapies are not yet ready for approved clinical use; however, this is a field in which the direction can change. Referring our patients to international clinical research trials is an option for approaches such as cobomarsen in CTCL or LNA-i-miR-221 in HCC. The comprehensive molecular oncology infrastructure of centers such as Memorial Göztepe provides a foundation for integration into future trials of miRNA-based personalized treatments.
🎓 DROZDOGAN Academy — Commentary and Discussion
MICRORNA · LIQUID BIOPSY · ARTIFICIAL INTELLIGENCE · MACHINE LEARNING · CLINICAL TRANSLATION
The Importance of This Review
The past thirty years of oncology literature have taught the world that miRNAs play a central role in cancer biology. The journey from the miR-15a/16-1 deletion in CLL discovered by Calin in 2002 to today's AI-assisted pan-cancer models is a case study in scientific history. Yet over the same period, investment of tens of millions of dollars has still failed to produce a single miRNA biomarker used in clinical practice. This paradox is the review's central question: how much of these thirty years of intensive research will we bring into clinical practice? The extraordinary development of artificial intelligence and machine learning over the last few years may be helping overcome the technical barriers underlying earlier failures (high-dimensional data analysis, multi-omics integration, and evaluation of multiple miRNA signatures). Calin and colleagues present tangible evidence of this change of direction through examples from multiple studies and seriously examine miRNA research's potential to move beyond the category of "successful as science, unsuccessful in the clinic." This matters not only for the miRNA field but also for the overall philosophy of cancer biomarker research — the transition from searching for a single molecule to signature-based models is a fundamental paradigm shift in modern molecular oncology.
Strengths
The review's methodological strength lies in its multilayered perspective: it presents an integrated picture spanning basic biology (miRNA biogenesis pathways, dual roles, context-dependent functions), translational applications (liquid biopsy, therapeutic delivery systems), and AI-based modeling techniques (classical ML algorithms, deep learning, foundation models). Joint authorship by scientists at globally leading institutions such as MD Anderson, Charité, and the German Cancer Consortium (DKTK) strengthens the review's academic authority. Its presentation of clinical trials is particularly valuable — failed studies (MRX34) and promising results (cobomarsen) are treated with equal care, preserving "transparency," a fundamental principle of medical ethics. Concrete innovations that could be ready for future application (Patho-DBiT, sChemNET, AOC platforms) are also presented to the reader; this approach strengthens the review's value as a guide to research directions over the next five to ten years. Its self-critical approach to AI's limitations also deserves recognition — issues such as the "black box" problem, the risk posed by small samples, and the lack of external validation are discussed without exaggerated hope.
Limitations and Open Questions
Several important points emerge from a critical assessment. First, the review generally offers a hopeful view of the future of miRNA-based medicine, but we must remember that similar hopes have failed to become reality for thirty years. Claims that artificial intelligence will "definitively" overcome these barriers are not yet supported by sufficient clinical data. Second, the AI studies in the review largely involve models using data from single-center, retrospective cohorts; independent multicenter validation is limited. This keeps the risk of "overfitting" (a model fitting only its training data) high. Third, the review has not sufficiently discussed the cost and technical challenges of miRNA test panels that can be applied in a standardized way in clinical laboratories — a vital issue, especially for low- and middle-income health systems. Fourth, the economic reality of miRNA-based treatment trials is missing: the early termination of INT-1B3 because of "insufficient funding" shows that the field remains largely dependent on small biotechnology companies and has not attracted major pharmaceutical investment. Fifth, although the review discusses AI models' "black box" problem, its practical solution in clinical decision-making remains unclear — can a physician accept a patient's miRNA-based cancer diagnosis without being able to explain why it was made? Sixth, ethnic and geographical diversity is insufficiently addressed — current AI models have largely been trained on European and American cohorts, with validation lacking for Turkish, Asian, or African populations. Finally, even though the dual-role reality is described in the review, its significance as a barrier to clinical decision-making is not sufficiently emphasized — the question of which miRNA panel could serve as a biomarker in which tumor type and at which stage reflects the tension between the ideal of "personalized oncology" and practical clinical applicability.
Implications for Clinical Practice
The concrete messages for oncology practitioners in Turkey are as follows. First, miRNA-based tests are not yet part of routine clinical practice; however, developments in liquid biopsy have the potential to change this in the near future. We should provide realistic information when discussing these subjects with our patients — caution about commercial claims for tools such as an as-yet-unapproved "miRNA test" is important. Second, for our patients with rare diseases such as cutaneous T-cell lymphoma, access to international clinical trials of cobomarsen and similar miRNA-based treatments is an option to consider. Third, trials such as LNA-i-miR-221 for patients with hepatocellular carcinoma may also open in Turkey in the coming years; strategic importance lies in centers preparing for these studies. Fourth, paying close attention to patient intake is necessary: obtaining high-quality tumor tissue samples and storing them appropriately (RNA integrity is highly sensitive) provide a foundation for future miRNA analyses. The biobank infrastructure of centers such as Memorial Göztepe is a strategic resource for Turkey in this respect. Fifth, we are entering an era when miRNA-related decisions in patient assessment could be supported by artificial intelligence; this should support, rather than replace, physicians' clinical judgment. Clinicians should approach these new tools with a mindset of critical assessment, rather than treating them as a "black box." Sixth, miRNA signature-based screening approaches, especially for esophageal and gastric cancers and hepatocellular carcinoma (tumors with relatively high incidence in Turkey), could be integrated into population health programs in the future; this is an area already worth feasibility studies. Finally, the dual-role reality of miRNAs shapes clinical decision-making — the same miRNA can have different functions in different regions even within the same tumor, underscoring once again the importance of tissue heterogeneity and multiple sampling strategies. This connects with our earlier discussion on drozdogan.com of tumor-draining lymph nodes and PD-L1/NGS heterogeneity as a shared theme: in cancer diagnosis and treatment, the logic of "one sample, one decision" is giving way to a philosophy of "multiple samples, multilayered decisions."
Future Questions
The critical questions for the coming period are these: Can AI-assisted pan-cancer miRNA panels (such as the 39-miRNA study by Wu and colleagues) demonstrate effectiveness in multicenter prospective validation studies and obtain regulatory approval? How will the potential to "detect early-stage cancer 32 years in advance" (Keller and colleagues' SOM study) be managed ethically — how should telling someone decades beforehand that they are "high risk" be framed in terms of psychological and insurance effects? Can spatial transcriptomics technologies such as Patho-DBiT bring miRNA distribution within the tumor microenvironment into a format that can be integrated into routine pathology practice? When will foundation models (MDformer, SeNMo) enter clinical decision-support systems, and how will we overcome the "black box" problem through explainable AI (xAI) approaches? Can AOC platforms (antibody-oligonucleotide conjugates) repeat the commercial success of ADCs for miRNA-based therapies? Beyond cobomarsen in CTCL and LNA-i-miR-221 in HCC, for which oncologic diseases do miRNA-based therapies have the greatest chance of approval? Specifically for Turkey: Can a domestic miRNA biomarker research platform, national biobank infrastructure, and AI-assisted analysis center be established through cooperation between TÜBİTAK and the Ministry of Health, and will this become a health-policy priority? Ethically, how will the risk that a miRNA test capable of identifying cancer risk decades in advance could lead to discrimination in personal insurance and employment be managed? Finally, the deepest biological question: Does a change in miRNA signatures before cancer mean that the biological processes triggering cancer really begin decades earlier — or do miRNAs merely reflect early biological imbalances that we do not yet understand? These questions will shape the discussions guiding the next decade, not only for the miRNA field but for oncology as a whole.
Scientific References
- Jurj A, Dragomir MP, Li Z, Calin GA. MicroRNAs in oncology: a translational perspective in the era of AI. Nat Rev Clin Oncol 2026;23(4):239–259. doi:10.1038/s41571-025-01114-x
- Lee RC, Feinbaum RL, Ambros V. The C. elegans heterochronic gene lin-4 encodes small RNAs with antisense complementarity to lin-14. Cell 1993;75:843–854.
- Calin GA et al. Frequent deletions and down-regulation of micro-RNA genes miR15 and miR16 at 13q14 in chronic lymphocytic leukemia. Proc Natl Acad Sci USA 2002;99:15524–15529.
- Hong DS et al. Phase 1 study of MRX34, a liposomal miR-34a mimic, in patients with advanced solid tumours. Br J Cancer 2020;122:1630–1637.
- Querfeld C et al. Phase 1 trial of cobomarsen, an inhibitor of Mir-155, in cutaneous T cell lymphoma. Blood 2018;132:2903.
- Tassone P et al. Safety and activity of the first-in-class LNA miR-221 selective inhibitor: a first-in-human phase 1 study. J Hematol Oncol 2023;16:68.
- Wu P et al. Pan-cancer characterization of cell-free immune-related miRNA identified as a robust biomarker for cancer diagnosis. Mol Cancer 2024;23:31.
- Miyoshi J et al. (EMERALD). Liquid biopsy to identify Barrett's oesophagus, dysplasia and oesophageal adenocarcinoma: the EMERALD multicentre study. Gut 2025;74:169–181.
- Pardini B et al. A fecal microRNA signature by small RNA sequencing accurately distinguishes colorectal cancers: results from a multicenter study. Gastroenterology 2023;165:582–599.e8.
- Keller A et al. Competitive learning suggests circulating miRNA profiles for cancers decades prior to diagnosis. RNA Biol 2020;17:1416–1426.
- Bai Z et al. Spatially exploring RNA biology in archival formalin-fixed paraffin-embedded tissues (Patho-DBiT). Cell 2024;187:6760–6779.e24.
- Galeano D et al. sChemNET: a deep learning framework for predicting small molecules targeting microRNA function. Nat Commun 2024;15:9149.
- Dong B et al. MDformer: a transformer-based method for predicting miRNA-disease associations. Comput Biol Med 2023;167:107585.
- Drula R, Calin GA. MicroRNAs: where brilliance, perseverance, and ambition converged. J Clin Invest 2024;135:e189625.
This article was created using various editorial tools, including artificial intelligence, as part of the process. It was reviewed by human editors before publication. The miRNA-based diagnostic and treatment approaches discussed here are largely at the research stage, and there is not yet an approved miRNA-based cancer test or therapy available for standard clinical use in Turkey.
Was this article helpful?
Your feedback helps us improve our content.
Reader comments
Share your comment
Your name and comment are published only after administrator approval. Your email address is never shown to visitors.
Be the first to comment.