Designing clinical trials for dystonia poses formidable methodological challenges because the disorder is both clinically and etiologically heterogeneous. A mechanism-based approach requires deeper understanding of underlying pathophysiology and identification of factors that may impact disease progression, while a phenotypic approach focuses on improving dystonic symptoms by discovering new drugs or building upon existing therapies. Primary outcome measures have been variable, relying mainly on patients’ subjective reports of therapeutic benefit, and objective quantification remains difficult due to the diversity of dystonia presentations. In 2013, the Movement Disorders Society commissioned a task force to review 36 potential rating scales, ultimately recommending seven for craniocervical, laryngeal, and generalized dystonia, including the Toronto Western Spasmodic Torticollis Rating Scale for cervical dystonia, the Jankovic Rating Scale for blepharospasm, the Voice Handicap Index for laryngeal dystonia, and the Fahn-Marsden Dystonia Rating Scale for generalized dystonia. The review’s authors emphasize that development of sensitive and objective clinical outcome measures will be a crucial prerequisite for successful future trials.

On the pharmacological front, the past five years have seen trials targeting the neurotransmitter pathways implicated in dystonia, including dopamine, acetylcholine, and GABA. Levodopa remains the cornerstone treatment for dopa-responsive dystonia, which accounts for about five percent of childhood dystonias, but there are currently no active clinical trials studying dopaminergic therapies. Trihexyphenidyl, an anticholinergic drug introduced in the 1950s, continues to be one of the most effective treatments for generalized dystonia, though side effects like confusion, dry mouth, and drowsiness often limit its use in adults. The Children’s Mercy Hospital in Kansas City is running the TRIKE2 trial, a Phase 1, 16-week, single-arm pilot study evaluating how CYP2D6 and CYP2C19 genotypes influence trihexyphenidyl metabolism in 40 pediatric patients with dystonic cerebral palsy, aiming to inform the feasibility of future pharmacogenomic trials. Meanwhile, the PREDICT-ITB trial at Baylor College of Medicine will enroll 65 children with dystonic cerebral palsy through 2029, following a standardized intrathecal baclofen titration protocol with the Barry-Albright Dystonia Scale as the primary outcome measure at baseline, three, six, and twelve months.

One of the most striking results came from the Massachusetts Eye and Ear Infirmary, which completed recruitment in 2024 for a randomized, placebo-controlled crossover trial of sodium oxybate in spasmodic dysphonia and voice tremor. The trial enrolled 117 participants with laryngeal dystonia, both alcohol-responsive and alcohol-nonresponsive, with and without voice tremor. Results published in 2025 showed that sodium oxybate significantly reduced symptom severity in the alcohol-responsive group, with a mean Visual Analog Scale severity score of 28.0 versus 14.2 for placebo, a difference that reached statistical significance at p = 0.008. No significant benefit was observed in the alcohol-nonresponsive group. Adverse events were mild to moderate, including dizziness in 48 percent, daytime sleepiness in 20 percent, and mild nausea in 14 percent, without serious adverse events. This is the first controlled clinical trial demonstrating sodium oxybate as a therapeutic option for laryngeal dystonia and voice tremor, particularly for alcohol-responsive patients.

VMAT-2 inhibitors, which received FDA approval as valbenazine and deutetrabenazine in 2017 for tardive dystonia and dyskinesia, have been repurposed in recent trials. From 2021 to 2023, the Orthopedic Foundation assessed valbenazine for cervical dystonia in a Phase 4 open-label study of 13 participants who completed 16 weeks of treatment titrated to 80 milligrams per day while maintaining botulinum toxin treatment. Preliminary abstract results reported significant improvements in the TWSTRS at p = 0.01 and in the Visual Analog Scale for pulling at p = 0.02. The University of Pennsylvania recently completed the AUDYT trial, an open-label study of deutetrabenazine in 15 adults with non-dopa-responsive isolated dystonia, titrating the drug from 12 to a maximum of 48 milligrams per day over 12 weeks. Only 40 percent of participants tolerated the maximum dose, and although the median change in the Global Dystonia Rating Scale was zero, the variability was notably wide, suggesting a need for a larger cohort in future trials. Glutamate-targeting drugs have also been explored: perampanel, an AMPA receptor antagonist, showed modest improvements in pain relief and sleep quality in cervical dystonia but was limited by tolerability, while dipraglurant, a negative allosteric modulator of the mGlu5 receptor, was well-tolerated in blepharospasm but provided no meaningful clinical benefit.

Botulinum toxin therapy continues to gain traction as a high-yield area for dystonia research, with five commercially available formulations FDA-approved for first-line treatment of cervical dystonia and only two approved for blepharospasm. The most significant recent advance is daxibotulinumtoxinA, whose molecular size and unique stabilizing excipient peptide enhance binding through electrostatic charge and reduce spread of the toxin, leading to a longer therapeutic duration. In the pivotal ASPEN-1 Phase 3 trial completed in 2020 across 60 international sites, 291 participants with cervical dystonia were randomized to daxibotulinumtoxinA 125 units, 250 units, or placebo. The placebo group had a mean TWSTRS reduction of 4.3 points compared to 12.7 points in the 125-unit group and 10.9 points in the 250-unit group, both highly significant. The duration from injection to loss of 80 percent of peak effect was significantly longer than other botulinum toxin type A formulations, lasting 24 weeks and 20 weeks respectively, with no dose-dependent increases in adverse events. The subsequent ASPEN-OLS open-label extension enrolled 357 patients over 52 weeks, confirming that successive treatments were well-tolerated with dysphagia rates remaining stable or decreasing after repeat dosing.

Other botulinum toxin formulations are being validated in clinical trials even without FDA approval for dystonia. PrabotulinumtoxinA, FDA-approved for glabellar lines, completed a Phase 2 randomized placebo-controlled trial in 57 patients with cervical dystonia, reporting no treatment-related serious adverse events up to 20 weeks, with peak efficacy occurring within two to eight weeks and wearing off starting at 16 weeks. NeubotulinumtoxinA, developed in South Korea as a cost-effective alternative, demonstrated comparable efficacy to abobotulinumtoxinA in a 48-week crossover study of 52 Thai patients with cervical dystonia. LetibotulinumtoxinA showed significant reduction in the Jankovic Rating Scale at four weeks in a 220-participant blepharospasm trial, confirming non-inferiority. Beyond new formulations, trials are optimizing injection parameters: a Phase 4 noninferiority study of 207 patients demonstrated that shorter incobotulinumtoxinA injection intervals of eight weeks were safe and effective for patients with early waning of benefit, while a Western University trial in Canada showed that kinematic-guided onabotulinumtoxinA injection achieved faster symptom reduction than visual assessment alone, with the kinematics group reducing TWSTRS from 39.1 to 27.9 points at six weeks.

Deep brain stimulation, FDA-approved for dystonia in 2003 under a Humanitarian Device Exemption, has seen exciting progress with new targets and innovative programming techniques. A 2024 French multicenter trial published by Cuartero and colleagues implanted dual DBS in 12 patients with medically refractory dystonia, comparing GPi, STN, and thalamic centromedian-parafascicular monotherapy against combined targets. The study confirmed the efficacy and safety of GPi-DBS with 44 percent improvement in dystonia severity at three months, but found no additional benefit from dual-target stimulation or alternative targets alone. In China, the MEIGES trial at Beijing Tiantan Hospital is enrolling 110 patients with idiopathic craniofacial dystonia to investigate whether STN-DBS is noninferior to GPi-DBS, while Qilu Hospital is comparing the two targets in Meige syndrome with 100 patients. Peking University People’s Hospital published results from 98 patients with primary Meige syndrome, showing BFMDRS improvements of 61 to 66 percent at one year and 65 to 71 percent at three years for DBS, compared to 58 percent at one year and 50 percent at three years for pallidotomy. Two active trials at Indiana University and Massachusetts Eye and Ear are exploring DBS for laryngeal dystonia using novel outcome measures including acoustic voice recordings, high-speed videoendoscopy, and intraoperative microelectrode recordings of brain signals during speech production.

Innovative DBS programming is advancing toward personalized therapy. The University of California, San Francisco has enrolled 25 patients, including five with dystonia, in a closed-loop DBS study formulating individualized adaptive algorithms based on specific cortical biomarkers, with each patient undergoing four weeks of open-loop versus four weeks of adaptive mode stimulation. Wuerzburg University Hospital is recruiting 80 participants for a prospective randomized double-blind crossover trial comparing computer-assisted image-based DBS programming against conventional clinician-based programming, aiming to optimize stimulation settings and reduce non-responder rates. Novel brain lesioning techniques are also being tested as alternatives to DBS. The University of Maryland launched a Phase 1 trial in 2024 for MR-guided focused ultrasound pallidotomy in task-specific focal hand dystonia, enrolling 10 patients who failed botulinum toxin and DBS, while the University of the Philippines Manila is investigating focused ultrasound pallidothalamic tractotomy for X-linked dystonia-parkinsonism, now available through expanded access until 2026. The All India Institute of Medical Sciences is planning a cohort study and registry of 250 patients undergoing lesioning procedures for movement disorders, with long-term functional outcomes assessed over a year.

Non-invasive brain stimulation, gene-targeted therapies, and rehabilitative approaches round out the therapeutic landscape. Transcranial magnetic stimulation has shown potential in small cohorts but variability in technique, administration, and response remains a barrier to FDA approval for dystonia. A randomized crossover trial at AIIMS New Delhi targeting the inferior parietal lobule in 16 patients with task-specific focal hand dystonia showed significant improvement in the Writer’s Cramp Rating Scale after real stimulation compared to sham, while Duke University is recruiting for accelerated TMS trials using functional MRI to guide individualized stimulation. Transcranial ultrasound stimulation is emerging as a novel modality, with the University Health Network in Toronto studying low-intensity ultrasound to non-invasively stimulate the basal ganglia. Gene therapy has made rapid advances: in November 2024, the FDA approved the first gene therapy for aromatic L-amino acid decarboxylase deficiency based on positive Phase 1/2 results, with the National Taiwan University Hospital trial showing mean Peabody Developmental Motor Scales-2 scores increasing from 10.4 to 80.5 and cerebrospinal fluid homovanillic acid rising from 6.6 to 30.2 nanomoles per liter in 26 patients. In vitro studies are exploring CRISPR-Cas9 disruption of mutant TorsinA in DYT1 dystonia fibroblasts, showing selective reduction of mutant protein with phenotypic recovery, though no active clinical trials yet test gene-targeted therapy in hereditary dystonia patients. Rehabilitative techniques including vibration therapy, transcutaneous electrical nerve stimulation, osteopathic manipulation, behavioral exercises, and neurofeedback brain-computer interface training are all being explored as adjunctive therapies, with a French TENS trial showing an eight-point TWSTRS reduction from baseline in painful cervical dystonia and the University of Minnesota demonstrating that laryngeal vibrotactile stimulation can improve voice quality within 15 to 20 minutes, though the benefit decays within 30 minutes. The review’s authors conclude that successful therapy development should be based on precise understanding of pathophysiology, validated by sensitive clinical outcome measures in well-designed trials, and supported by correlating changes in biomarkers, with a biology-tailored approach essential as understanding of dystonia subtypes advances.

Subject of Research: Neuroscience

Article Title: A narrative review: clinical trials in therapeutic interventions for dystonia (2020 – 2025)

Article References: Ly, A., Mushtaheed, A., Soliman, M. E., Karp, B., & Cho, H. J. (2025). A narrative review: clinical trials in therapeutic interventions for dystonia (2020 – 2025). Dystonia, 4, Article 14547. https://doi.org/10.3389/dyst.2025.14547

Image Credits: AI Generated

DOI: 10.3389/dyst.2025.14547

Keywords: dystonia, clinical trials, botulinum toxin, deep brain stimulation, gene therapy, movement disorders, neurology, transcranial magnetic stimulation, narrative, review, clinical, trials

The distinction between the two delivery platforms identified in the review—lipid nanoparticles carrying CAR-encoding messenger RNA and integrating lentiviral vectors—reflects two fundamentally different philosophies about how in vivo programming should work. Messenger RNA delivered by lipid nanoparticles produces a transient expression of the CAR on the surface of T cells, because the mRNA does not integrate into the genome and is gradually degraded by cellular machinery. This transience is a double-edged sword: it may limit the duration of the therapeutic effect and require repeat dosing, but it also provides an inherent safety mechanism, since any toxicity related to CAR expression should diminish as the mRNA is cleared. Lentiviral vectors, by contrast, integrate the CAR construct into the genome of the transduced cells, producing stable, potentially lifelong expression. This durability is attractive for malignancy treatment, where persistent anti-tumor activity is desired, but it raises the specter of uncontrolled in vivo expansion, since the vector cannot be restricted to a specific cell population with the precision achievable in an ex vivo manufacturing setting.

The targeting strategies used to direct these agents to T cells deserve particular attention. The review’s abbreviation list reveals the use of a CD8-targeted lipid nanoparticle, which exploits antibodies or ligands that bind the CD8 molecule expressed on cytotoxic T lymphocytes. This approach attempts to concentrate the payload on the cell type most likely to mediate tumor killing, while sparing CD4-positive helper cells, B cells, and other immune populations. The in vivo Gene Placement System referenced in the abbreviations suggests an alternative strategy in which the vector itself is engineered to preferentially integrate or express in T cells, perhaps by exploiting T-cell-specific promoters or miRNA-regulated expression cassettes that silence the transgene in non-T-cell lineages. These targeting mechanisms are critical because off-target transduction of non-T cells could produce unintended consequences, including expression of the CAR on antigen-presenting cells, which could theoretically trigger immune dysregulation or fratricidal effects.

The clinical results summarized in the review, while preliminary, offer encouraging signals across both malignant and autoimmune indications. In relapsed or refractory multiple myeloma, the achievement of minimal residual disease negativity is particularly noteworthy, as MRD negativity is one of the most powerful predictors of long-term survival in myeloma and is generally only achieved with deep, sustained anti-myeloma activity. The fact that in vivo-generated CAR T cells could produce this depth of response suggests that the endogenous T cells, once reprogrammed, are capable of mounting effective anti-tumor responses comparable in some respects to those achieved with ex vivo-manufactured products. In B-cell non-Hodgkin lymphoma, objective responses were observed, which is consistent with the known biology of CD19-directed CAR T-cell therapy in this disease, although the specific target antigen used in the in vivo trials is not detailed in the abstract.

The inclusion of autoimmune diseases in the review reflects a broader trend in the CAR T-cell field, where the deep B-cell depletion achieved by CD19-directed CAR T cells has been repurposed to treat conditions such as systemic lupus erythematosus, in which autoreactive B cells drive pathology. The observation of improved disease activity across autoimmune cohorts in this review aligns with case reports and small series published by other groups, suggesting that in vivo CAR T-cell generation may eventually offer a safer, more accessible alternative to ex vivo products for autoimmune indications, where the risk-benefit calculus differs from that in advanced malignancy.

The safety profile reported in the review warrants careful interpretation. Cytokine release syndrome occurred in 24 of 31 patients in studies reporting numerical incidence, which is a high proportion, although the review notes that most events were low grade. This is consistent with the known pharmacology of CAR T-cell therapy, in which the activation of engineered T cells triggers a cascade of inflammatory cytokines including interleukin-6, interferon-gamma, and tumor necrosis factor-alpha. The occurrence of severe events, even if uncommon, underscores that in vivo CAR T-cell generation does not eliminate the fundamental risk of on-target inflammatory toxicity. The rarity of neurotoxicity, with only one grade 1 event reported across the included studies, is notable and may reflect several factors, including the small sample size, the transient nature of mRNA-based CAR expression, or differences in the cytokine milieu generated by in vivo versus ex vivo engineered cells. However, given the small numbers of patients involved, it would be premature to conclude that in vivo CAR T-cell therapy is inherently less neurotoxic than conventional approaches.

The methodological limitations acknowledged by the authors are substantial and deserve emphasis. Seven clinical trials involving 41 patients represent an extremely small evidence base, and the heterogeneity of the included studies—in terms of disease types, delivery platforms, target antigens, doses, and patient populations—makes it impossible to draw meaningful comparative conclusions. The short follow-up reported is particularly concerning for malignancy indications, where durability of response is the key clinical question. A patient with relapsed myeloma who achieves an initial complete response but relapses within three months has not benefited meaningfully from the therapy, and the review provides insufficient information to assess whether the responses observed are durable. The authors’ call for standardized reporting of cellular kinetics is well taken, as the expansion and persistence characteristics of in vivo-generated CAR T cells are likely to differ substantially from those of ex vivo products, and understanding these differences is essential for optimizing dosing schedules and predicting toxicity.

Looking forward, several challenges must be addressed before in vivo CAR T-cell therapy can advance beyond early-phase testing. First, the efficiency and specificity of T-cell targeting must be demonstrated rigorously in larger cohorts, with careful biodistribution studies to quantify off-target transduction. Second, the pharmacokinetics and pharmacodynamics of repeat dosing, particularly for mRNA-based platforms, need to be characterized, including whether repeated administration leads to immune responses against the lipid nanoparticle or the mRNA itself. Third, the question of dose control is paramount: in conventional CAR T-cell therapy, the dose is precisely known because the cells are counted before infusion, whereas in vivo generation means that the actual number of CAR-positive T cells produced depends on transduction efficiency, T-cell fitness, and host factors that vary from patient to patient. Fourth, the regulatory pathway for in vivo CAR T-cell products remains uncertain, as these agents blur the traditional distinction between gene therapy, cell therapy, and drug products, potentially requiring novel regulatory frameworks. Finally, the cost-effectiveness of in vivo approaches, while potentially superior to ex vivo manufacturing in the long run, has not been established, and the infrastructure required for quality control and batch release of in vivo products may prove more complex than anticipated.

The systematic review by Mahmoudi and colleagues thus serves as an important early map of a rapidly evolving landscape. It documents that in vivo CAR T-cell generation is biologically feasible and capable of producing clinically meaningful responses, while honestly acknowledging the limitations of the current evidence base. As more trials mature and additional platforms enter clinical testing, the field will need to move beyond proof-of-concept toward rigorous comparative studies that can determine whether in vivo CAR T-cell therapy can match the efficacy of established ex vivo products while delivering on its promise of broader accessibility, reduced cost, and improved safety.

Subject of Research: In vivo CAR T-cell therapy: a systematic review of clinical trials

Article Title: In vivo CAR T-cell therapy: a systematic review of clinical trials

Article References: Mahmoudi, E., Shokrolahi, F., Khoshkanab, A. H. R., & Namjoo, S. (2026). In vivo CAR T-cell therapy: a systematic review of clinical trials. BMC Cancer. https://doi.org/10.1186/s12885-026-17165-x

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17165-x

Keywords: vivo, T-cell, therapy, systematic, review, clinical, trials, scientific research

Researchers led by Cheewon Ahn and Hyunsuk Jeong at The Catholic University of Korea screened 5,527 records from PubMed, Embase, and the Cochrane Library and ultimately included 57 randomized controlled trials from 53 reports, covering 28 distinct siRNA agents and a total of 11,757 participants. The protocol was registered with PROSPERO, and two reviewers independently performed study selection and data extraction, with disagreements resolved by a third reviewer. Only placebo-controlled or treatment-as-usual randomized trials were eligible, deliberately excluding single-arm and open-label studies, because the diseases targeted by siRNAs are often intractable, with high baseline mortality and hospitalization rates that can masquerade as drug toxicity without a proper control group.

The pooled analysis focused on four hard safety outcomes: death, serious adverse events, severe adverse events, and adverse events leading to treatment discontinuation. Compared with controls, siRNA treatment did not increase the risk of death, with a pooled relative risk of 0.77 and a 95 percent confidence interval of 0.62 to 0.95. Severe adverse events were likewise significantly lower in the treatment group, at a relative risk of 0.85 with a confidence interval of 0.76 to 0.95. The risk of serious adverse events was not increased, with a relative risk of 0.94 and a confidence interval of 0.88 to 1.00, and treatment discontinuation due to adverse events showed no significant increase, at a relative risk of 0.86 with a confidence interval of 0.66 to 1.13. Remarkably, statistical heterogeneity across studies was essentially zero for every outcome, with I-squared values of 0 percent, indicating an unusually consistent safety signal across 28 chemically and pharmacologically diverse molecules.

That consistency matters because siRNA therapeutics work through a mechanism fundamentally different from conventional small-molecule drugs. The guide strand of the siRNA binds to the RNA-induced silencing complex, or RISC, triggering RNA interference within the cell to silence a target gene in a sequence-specific manner. Early siRNA agents suffered from rapid degradation in vivo, off-target effects, and immunogenicity, but modern drugs incorporate a phosphorothioate backbone and ribose modifications such as 2′-O-methyl and 2′-fluoro, which enhance resistance to nuclease degradation and reduce immune activation. In silico sequence design minimizes unintended hybridization, and Good Laboratory Practice toxicology studies determine the initial starting dose for human trials.

Delivery technology has been equally transformative. Triantennary N-acetylgalactosamine, or GalNAc, conjugation targets the asialoglycoprotein receptor, which is predominantly expressed on hepatocytes, facilitating uptake through receptor-mediated endocytosis. This enables highly selective delivery to the liver while minimizing drug distribution to other tissues, and GalNAc has become the preferred delivery system for liver-targeting siRNAs. Six of the seven FDA-approved therapeutics use GalNAc conjugation alongside 2′-O-methyl, 2′-fluoro, and phosphorothioate modifications. The subgroup analysis confirmed that GalNAc-based agents showed no increased mortality risk, at a relative risk of 0.79, and non-GalNAc agents also showed no increase, at 0.66, although the authors caution that unfavorable non-GalNAc drugs may have been discontinued before any data were published, introducing potential publication bias.

The most sobering findings come from the discontinued programs. Of the ten studies in the analysis involving terminated drug development, four were halted because of safety concerns. Three of these were stopped because the delivery vehicle, rather than the siRNA molecule itself, caused mortality in non-human primates. The fourth, revusiran, a first-generation GalNAc-siRNA conjugate with so-called standard template chemistry, was terminated after a mortality imbalance of 23 deaths in the treatment group versus 7 in the placebo group by the end of safety follow-up. Revusiran required weekly 500-milligram subcutaneous injections, whereas vutrisiran, which targets the same indication with second-generation enhanced stabilized chemistry, is administered as a 25-milligram injection once every 12 weeks. The structural stability improvements resulted in markedly different drug exposures, and all other approved and most ongoing siRNA agents now use second-generation chemistry, suggesting that improvements in chemical modification and delivery have translated directly into enhanced clinical safety.

Subgroup analyses by disease category, clinical trial phase, and development status reinforced the overall picture. siRNA treatment did not increase the risk of death in either the rare disease subgroup, with a relative risk of 0.76, or the common disease subgroup, at 0.78. Marketed drugs showed no increased mortality, at a relative risk of 0.76, and agents in ongoing development showed a favorable profile, at 0.47, though most of those data come from early-phase trials with smaller cohorts. The discontinued or terminated subgroup showed a risk similar to control, at a relative risk of 1.01. The quality of adverse event reporting was high, with a median McHarm score of 13 out of a possible 15, and 96 percent of reports specified the number of deaths in each study group, lending credibility to the pooled estimates.

The authors also examined whether the adverse events observed in trials fell within the range anticipated by preclinical safety assessments. Toxicities of siRNA drugs can be classified into three broad types: hybridization- and sequence-dependent effects such as off-target silencing, which are managed preclinically through in silico analysis; hybridization- and sequence-independent effects that occur when plasma concentrations exceed a threshold, which are managed by adjusting dosage and frequency; and hybridization-independent but sequence-dependent effects, including proinflammatory reactions and toxicity to high-exposure organs such as the liver and kidneys. Injection site reactions were the most frequently reported adverse event in the included trials, but the vast majority were mild to moderate and resolved without intervention. Laboratory abnormalities, such as transient liver enzyme or creatinine elevations, were largely not clinically relevant or resolved spontaneously, implying that the adverse events reported in clinical trials generally fell within the anticipated range of preclinical toxicities.

Nevertheless, the researchers emphasize that the reassuring short-term profile does not preclude toxicity in individual patients. Some trials reported serious hepatic or electrocardiogram abnormalities that led to participant withdrawal, and even rare cases where a causal link cannot be excluded underscore the need for rigorous monitoring. The analysis also has limitations: only three trials had follow-up of two years or more, safety was often a secondary endpoint and severe adverse events were not explicitly reported in 24 of the 53 reports, and the study could not evaluate all pharmacological variables, such as dosing schedules or distinct non-GalNAc delivery systems like lipid nanoparticles. Long-term safety will need to be tracked in open-label extension phases, and a five-year extension study of inclisiran found no new safety signals.

For a field that has moved from laboratory curiosity to approved medicine in barely two decades, the meta-analysis represents a milestone: the first systematic quantitative evaluation of key safety outcomes across the entire clinical siRNA portfolio. The findings support the conclusion that chemical modifications and delivery advancements introduced to reduce the toxicity of early-generation siRNAs have successfully translated into enhanced clinical safety. But the authors are careful to note that unforeseen safety issues can still manifest during clinical development, and that thorough safety validation throughout early trial phases, along with extended surveillance for delayed toxicities, remains essential to elucidate the long-term safety profile of these gene-silencing medicines.

Subject of Research: Safety profile of small interfering RNA therapeutics: a systematic review and meta-analysis of randomized controlled trials

Article Title: Safety profile of small interfering RNA therapeutics: a systematic review and meta-analysis of randomized controlled trials

Article References: Ahn, C., Jeong, H., Yim, H. W., Kim, N. J., Han, S., & Oh, I. (2026). Safety profile of small interfering RNA therapeutics: a systematic review and meta-analysis of randomized controlled trials. Frontiers in Drug Safety and Regulation, 6, Article 1855572. https://doi.org/10.3389/fdsfr.2026.1855572

Image Credits: AI Generated

DOI: 10.3389/fdsfr.2026.1855572

Keywords: Safety, profile, small, interfering, therapeutics, systematic, review, meta-analysis, randomized, controlled, trials, scientific research

The appeal of AI in this setting rests on a simple asymmetry. Clinical trials generate enormous volumes of structured and unstructured data—imaging, laboratory values, genomic profiles, clinician notes, wearable-device streams and patient-reported outcomes—yet historically only a small fraction of that information has been used systematically in trial design and conduct. Machine-learning models, particularly modern deep-learning and foundation-model architectures, excel at extracting signal from exactly these high-dimensional, heterogeneous datasets. When trained on curated clinical data, they can identify which patient subpopulations are most likely to respond to a therapeutic candidate, predict which sites will struggle with recruitment, flag data inconsistencies long before database lock, and estimate the probability that an ongoing trial will meet its primary endpoint. The result is a shift from retrospective, intuition-driven trial management toward a prospectively optimised, continuously adaptive process.

Patient recruitment illustrates the potential most vividly. Failed recruitment and poor participant retention are among the leading reasons that trials miss their timelines or terminate early. Traditional approaches rely on broad eligibility criteria, manual chart review and outreach from a limited number of academic centres, which systematically under-enrols patients from rural areas, lower-income communities and historically marginalised groups. AI-driven approaches invert this model. Natural-language processing systems can scan millions of de-identified electronic health records to locate every patient matching a complex eligibility profile, including patients whose relevant conditions are described only in free-text notes rather than coded diagnoses. Matched patients can then be connected with trial sites through their treating physicians or through decentralised trial infrastructure that brings the study to the participant rather than the reverse. When designed carefully, such tools can also be tuned to improve diversity in enrolment, addressing a long-standing scientific weakness as well as an ethical one: a trial population that does not resemble the eventual treated population limits the generalisability of the results.

Trial design itself is being recomputed. Classical phase II and phase III trials typically use fixed designs conceived months before enrolment begins, and any mid-course revision requires protocol amendments that can delay readouts by many months. AI-assisted design tools draw on large repositories of historical trial outcomes, natural-history data and simulation frameworks to model, before a single patient is enrolled, how a trial will behave under different assumptions about effect size, dropout, endpoint variability and site performance. Bayesian adaptive designs, which allow the randomisation ratio and dose allocation to shift as interim data accumulate, become far more practical when machine-learning models supply reliable predictive components. Digital twins—computational replicas of individual patients built from rich baseline data—offer an emerging complement, allowing a portion of the control information in a trial to be estimated rather than observed, and thereby reducing the number of real participants who must be randomised to achieve a given statistical power. Regulators have begun engaging with these methods, and several external control arms constructed from historical or registry data have already supported regulatory submissions in rare-disease oncology.

Once a trial is running, AI changes the economics of monitoring. Risk-based monitoring, in which scrutiny is concentrated on the sites and data elements most likely to harbour errors, depends on recognising patterns across thousands of simultaneous data streams—precisely the task that anomaly-detection algorithms perform well. Centralised statistical monitoring can flag sites whose data distributions deviate from the norm, whether because of fraud, systematic measurement error or simple process breakdown, without the cost of sending monitors to every site on a fixed schedule. Automated adverse-event signal detection can surface safety trends earlier in the data stream, shortening the interval between an emerging risk and a protocol response. Speech-recognition and summarisation models are increasingly used to draft clinical notes, assist with adjudication of endpoints that require expert review, and reduce the administrative burden that currently consumes a large share of investigator time. Because the marginal cost of applying a trained model to new data is close to zero, these efficiencies can scale across an entire portfolio rather than applying to a single study.

The pharmaceutical industry’s interest follows directly from the arithmetic. Industry analyses have repeatedly estimated that bringing a new drug to market costs on the order of one to two billion dollars, with clinical development accounting for the majority of that expenditure and with most of the cost attributable to failed trials. Even a modest improvement in the probability of technical success, achieved through better target-patient matching or earlier detection of futility, translates into hundreds of millions of dollars in expected savings per programme and, more importantly, into faster access for patients to therapies that work. Every month shaved from a development timeline extends the effective patent-protected market life of a medicine, which strengthens the commercial case, but the public-health case is at least as strong: pipelines that iterate faster can respond more quickly to emerging pathogens, to rare diseases that currently have no treatment at all, and to the individualisation of therapy in fields such as oncology where one-size-fits-all efficacy is the exception rather than the rule.

None of this means the transformation is automatic. Machine-learning models inherit the biases of the data on which they are trained, and clinical datasets under-represent precisely the populations in whom trial evidence is weakest. An algorithm trained mostly on data from large academic hospitals may perform poorly for patients managed in community settings, and a recruitment tool optimised narrowly for speed could worsen enrolment diversity if fairness constraints are not built in explicitly. The opacity of complex models also sits awkwardly with regulatory expectations: a sponsor that uses an AI-derived covariate in a statistical analysis plan, or an AI-curated external control arm, must be able to explain to reviewers how the model was built, validated and monitored. Regulators including the United States Food and Drug Administration and the European Medicines Agency have signalled that they will evaluate such tools under existing frameworks for risk-based software validation, but the guidance landscape is still maturing, and sponsors who treat AI components as unexamined black boxes do so at their own regulatory peril.

Data governance presents a second structural challenge. The models that promise the greatest gains in recruitment and design are those trained on the largest and most diverse clinical datasets, yet health data are fragmented across institutions, jurisdictions and incompatible record systems, and privacy law constrains how they can be pooled. Federated learning, in which models are trained across multiple sites without moving the underlying patient data, offers a technically elegant partial solution, but it introduces its own questions about model ownership, auditability and the equitable distribution of the value created. Standard-setting efforts, from common data models to documented provenance for training sets, will determine whether AI-enabled trials become a broadly shared capability or a competitive advantage concentrated in a handful of organisations with the largest proprietary data estates.

The most realistic near-term picture is therefore not one of autonomous AI running trials, but of a human-machine division of labour in which algorithms perform the enumeration, matching, monitoring and simulation that humans cannot do at scale, while clinicians, statisticians and regulators retain judgment over what counts as evidence. Under that division of labour, the measurable signs of change are already visible: screening times measured in days rather than months at leading sponsors, growing numbers of adaptive and model-informed designs entering regulatory review, and decentralised, data-rich trial formats that were logistically implausible a decade ago. If those trends continue, the defining feature of the next generation of clinical trials will not be any single algorithm but the integration of computation into every stage of the evidentiary pipeline—from the first patient matching query to the final submission—producing trials that are faster, smaller where possible, larger where necessary, and ultimately more representative of the patients the resulting medicines are meant to serve.

Subject of Research: AI-enabled clinical trials

Article Title: AI-enabled clinical trials

Article References: Raynaud, M., Trayanova, N., Mannon, R. B., André, F., Doraiswamy, P. M., & Loupy, A. (2026). AI-enabled clinical trials. Nature Reviews Bioengineering. https://doi.org/10.1038/s44222-026-00487-7

Image Credits: AI Generated

DOI: 10.1038/s44222-026-00487-7

Keywords: AI-enabled, clinical, trials, scientific research, peer-reviewed research, research findings