The research team, led by Alaa Baqer Alali and colleagues based in Al-Hasa, Saudi Arabia, followed PRISMA 2020 reporting guidelines and searched five major databases, including PubMed, Scopus, the Cochrane Library, Web of Science, and Google Scholar, up to November 25, 2025. Their screening process winnowed 181 initial records down to eight eligible studies, encompassing an extraordinary total of 3,632,118 pregnancies. Of these, 2,648 pregnancies involved documented exposure to topical retinoids such as tretinoin, adapalene, tazarotene, and topical isotretinoin, while more than 3.6 million served as unexposed controls. The studies spanned three decades of research, drawing on data from Nordic national registries, European teratology information services, California teratology counseling records, and United States health cooperative pharmacy databases.

The headline finding concerns the primary outcome of major congenital malformations, defined according to established classification systems including EUROCAT, the CDC Metropolitan Atlanta Congenital Defects Program, and the International Classification of Diseases. Across four studies contributing data to this outcome, the pooled risk ratio was 0.83, with a 95 percent confidence interval of 0.64 to 1.07 and a P-value of 0.14. In plain terms, the point estimate actually trended slightly below unity, suggesting marginally lower malformation rates among exposed pregnancies, but the confidence interval crossed the null value of 1.0, meaning the association could not be distinguished from no effect at all. Strikingly, statistical heterogeneity across studies was essentially absent, with an I-squared value of 0 percent, indicating remarkable consistency in the direction of findings.

The secondary outcome of spontaneous abortion, defined as pregnancy loss before 20 to 22 weeks of gestation, told an equally reassuring story. Two studies enrolling 341 exposed and 833 control pregnancies yielded a pooled risk ratio of 0.99, with a confidence interval of 0.68 to 1.43 and a P-value of 0.94. Again, heterogeneity was nil. Additional outcomes reinforced the null pattern: birth weight differed by a non-significant 33.7 grams between groups, gestational age at delivery showed no meaningful difference, and preterm birth rates were statistically indistinguishable between exposed and unexposed pregnancies.

What elevates this meta-analysis above a routine evidence synthesis is the sophistication of its statistical machinery. Beyond the conventional frequentist random-effects model using the DerSimonian-Laird method, the researchers deployed Bayesian hierarchical modeling under three distinct prior specifications: a weakly informative Half-Cauchy prior, a skeptical prior that penalizes large effects, and an empirical prior derived from pharmacological intervention meta-analyses. Posterior risk ratio estimates ranged narrowly from 0.86 to 0.89 across all three priors, and every credible interval crossed unity. The Markov Chain Monte Carlo sampling, run with four chains and 4,000 iterations, converged cleanly, with Gelman-Rubin statistics at 1.000 and divergence rates far below accepted thresholds.

The most technically ambitious component was the Robust Bayesian Model-Averaged analysis, or RoBMA, which simultaneously evaluates 36 competing models spanning three dimensions of uncertainty: whether a true effect exists at all, whether heterogeneity between studies is real, and whether publication bias distorts the observed literature. The verdict was unambiguous. The posterior probability for the null hypothesis of no effect reached 74.1 percent, yielding a Bayes factor of 0.35 in favor of no association. Intriguingly, the same analysis assigned a 75.2 percent posterior probability to the presence of publication bias in the evidence base, yet when the trim-and-fill method imputed two hypothetical missing studies, the adjusted estimate of 0.80 actually strengthened rather than weakened the protective direction of the point estimate.

Beneath the reassuring averages, however, lies a structural vulnerability the authors were careful to expose. A single study, the 2025 Nordic registry cohort by Refsum and colleagues, contributed 87.3 percent of the pooled weight for the malformation outcome, drawing on prescription dispensing records across Denmark, Iceland, Norway, and Sweden. Leave-one-out sensitivity analysis revealed that omitting this single study shifted the pooled risk ratio from 0.83 to 1.07, a change of nearly 30 percent. A Dirichlet Process Mixture clustering analysis confirmed the pattern statistically, isolating the Nordic estimate in its own cluster at a risk ratio of 0.80 while grouping the five remaining effect estimates near perfect nullity. The authors interpret this not as inconsistency but as the expected signature of one high-precision study, with a standard error of 0.138, dominating several smaller studies whose standard errors ranged up to 0.78.

The pharmacokinetic rationale for these findings has existed for years. Percutaneous absorption studies have repeatedly demonstrated that plasma concentrations of tretinoin after topical application remain within endogenous physiological ranges, a stark contrast to oral isotretinoin, whose pooled odds ratio for major malformations has been reported at 3.76 in meta-analysis. Yet regulatory agencies, including the European Medicines Agency and the Australian Therapeutic Goods Administration, continue to advise against topical retinoid use during pregnancy as a precaution, and the United States transitioned retinoid products from letter-based pregnancy categories to the more nuanced Pregnancy and Lactation Labeling Rule in 2014 without resolving the underlying uncertainty.

Perhaps the most sobering finding in the entire analysis concerns not fetal biology but human behavior. In the European multicenter prospective cohort, women exposed to topical retinoids underwent elective termination at more than three times the rate of unexposed controls, a risk ratio of 3.20 with a confidence interval of 1.70 to 5.90. The authors attribute this excess not to detected fetal abnormalities but to counseling-related decisions driven by fear of teratogenicity, amplified by precautionary labeling. Prescribing pattern surveys reinforce the point: in one United States study of women trying to conceive, retinoids were prescribed in zero percent of acne cases, and in a French dermatologist survey, only 45 percent of clinicians checked contraception status when prescribing acne treatments to women of reproductive age.

The authors conclude that the evidence supports reassurance for women with inadvertent first-trimester exposure, but stops short of endorsing intentional use during pregnancy, given that the pooled estimates rest heavily on a single registry study and the confidence intervals include the null. E-value analysis confirmed that the findings are not robust to unmeasured confounding, since the intervals already cross unity. The path forward, they argue, lies in larger prospective studies with detailed exposure characterization, formulation-specific analyses, and attention to minor malformations that registry data may under-ascertain. Until then, the study stands as a rare piece of good news in reproductive pharmacovigilance: the most feared creams in the dermatology cabinet appear, on current evidence, far less dangerous than their oral cousins, and the anxiety they generate may itself be the greater clinical hazard.

Subject of Research: Safety of topical retinoid exposure during early pregnancy

Article Title: Topical retinoids and risk of major congenital malformations and spontaneous abortion: a systematic review and meta-analysis

Article References: Alali, A. B., Alhammad, F. A., Almohammed Saleh, J. A., Alshaikh, S. B., Alshayeb, Z. K., Al Nujaidi, A. J., Albinsaleh, A. A., Almutawah, M. R., Alhodibi, M. H., & Albesher, M. A. (2026). Topical retinoids and risk of major congenital malformations and spontaneous abortion: a systematic review and meta-analysis. Frontiers in Drug Safety and Regulation, 6, Article 1864872. https://doi.org/10.3389/fdsfr.2026.1864872

Image Credits: AI Generated

DOI: 10.3389/fdsfr.2026.1864872

Keywords: topical retinoids, pregnancy, congenital malformations, spontaneous abortion, teratogenicity, tretinoin, meta-analysis, Bayesian statistics, isotretinoin, dermatology, drug safety, acne

The study, led by Maja Louise Arendt of the University of Copenhagen and Jennifer R. S. Meadows of Uppsala University and SciLifeLab, focused on cereblon, the protein that thalidomide binds. Cereblon was only identified as the drug’s primary target in 2010, when researchers showed that thalidomide attaches to this substrate receptor of the cullin-4 E3 ubiquitin ligase complex, known as CUL4-RBX1-DDB1. When thalidomide occupies cereblon, the degradation machinery’s specificity changes: it begins recruiting, ubiquitinating and destroying an unnatural set of proteins called neo-substrates, including transcription factors such as SALL4, p63, Ikaros and Aiolos. These proteins govern critical developmental and immunological processes, which helps explain both the drug’s teratogenic tragedy and its therapeutic benefits. At the same time, native substrates such as MEIS2 are no longer degraded, adding further biological consequences. To date, cereblon remains the only described mammalian target of thalidomide.

The crux of the new findings lies in a single amino acid position within the cereblon protein. The C-terminal region of human cereblon, spanning roughly amino acids 318 to 426, forms the thalidomide-binding CULT domain, and variation within this 109-amino-acid stretch is known to determine how different species respond to the drug. In humans, the amino acid at position 388 is a valine, abbreviated V388. In mice, the equivalent position holds an isoleucine instead. This seemingly minor substitution does not prevent thalidomide from binding to cereblon, but it creates steric hindrance that blocks neo-substrates from entering the binding pocket, preventing their ubiquitination and degradation. The result is that thalidomide and its derivatives are pharmacologically ineffective in mice, a fact that contributed to the drug’s original licensing as a supposedly safe sedative, since rodent tests failed to predict its devastating effects in humans.

When the researchers mined the Dog10K consortium datasets, which include single nucleotide and structural variant calls from 1,929 and 1,879 canids respectively, they found that every single dog examined encodes an isoleucine at the position equivalent to human V388. Across the roughly 30-kilobase CRBN gene region, 384 variable sites were detected, but only three had any potential to alter the protein’s coding sequence, and each of these was extraordinarily rare, with allele frequencies below 0.2 percent. Two structural deletions overlapping the gene were also identified, both at frequencies below 0.5 percent. In other words, the canine cereblon protein is essentially uniform across the species, and it uniformly carries the mouse-like isoleucine that has been shown to abolish thalidomide’s neo-substrate degradation activity.

To place this finding in evolutionary context, the team turned to two massive comparative genomics resources: the Zoonomia alignment of 240 mammalian species and the TOGA alignment covering 344 eutherian mammals. At the nucleotide level, the codon encoding position 388 showed little evolutionary constraint, with phyloP conservation scores of 0.40, 4.66 and -1.74 across its three bases; only the second position showed purifying selection. This wobbling tolerance is mirrored by amino acid variability across the mammalian tree. Among 344 species, 72 percent encode an isoleucine at the equivalent position, including all 55 available species of the order Carnivora, to which dogs belong. Only bats, rodents and rabbits, and the order Primates, which includes humans, were found to encode both valine and isoleucine at this site. Notably, rabbits and guinea pigs carry a valine, consistent with decades of toxicology data showing that rabbits, unlike mice, do develop thalidomide-induced birth defects similar to those seen in humans.

The researchers also considered whether cancer itself might change the equation. Tumors accumulate mutations that can alter drug sensitivity, so the team searched for non-silent, protein-altering mutations in the CRBN gene within two published canine tumor-normal sequencing datasets, comprising 55 mammary tumors and 43 diffuse large B-cell lymphoma cases, as well as 723 samples from the cBioportal Canine Cancer Genome Atlas. The search yielded almost nothing: a single non-synonymous mutation, p.CRBN A10T, was found in one lymphoma sample, affecting the N-terminal region far from the thalidomide-binding CULT domain. No mutations were identified that would plausibly make canine cancer cells more susceptible to thalidomide than normal cells. In human multiple myeloma, somatic CRBN mutations are likewise reported to be neutral or to confer resistance, never enhanced sensitivity.

These genetic findings cast a revealing light on the veterinary literature. Thalidomide has been tested in dogs with splenic hemangiosarcoma, mammary carcinoma, lung carcinoma, lymphoma and multiple myeloma, usually in combination with other therapies, making efficacy difficult to assess. One recent multicentre retrospective study reported that 71 percent of seven dogs with relapsed multiple myeloma maintained or achieved complete remission on thalidomide as a single-agent rescue therapy, with moderate lethargy in two dogs as the only recorded side effect. Strikingly, a 53-week toxicology study in beagles using doses up to 1,000 milligrams per kilogram daily, roughly fifty times the human clinical dose for a ten-kilogram dog, found no clinically significant adverse effects, no hematological or biochemical changes and no evidence of the peripheral neuropathy that plagues human patients. The authors argue that this remarkable tolerance is itself a warning sign: a drug with no side effects at massive doses may simply have no pharmacological activity.

There are caveats. The sedative effect of thalidomide appears to be independent of cereblon, and one canine study reported reduced vascular endothelial growth factor staining in metastatic tissue from treated dogs, hinting that anti-angiogenic mechanisms independent of the CRBN-CUL4-RBX1-DDB1 complex might operate in dogs. A pharmacokinetic study in tumor-bearing dogs also reported a range of clinical signs, though the authors of the new analysis note these could reflect advanced cancer rather than drug toxicity. The researchers are careful not to claim that thalidomide is definitively useless in dogs; rather, they conclude that the molecular function of the drug in this species must be directly investigated before its therapeutic role in canine cancer can be understood or justified.

Beyond its immediate veterinary implications, the study is a compelling demonstration of how publicly available genomic data can inform clinical pharmacology. Because DNA-damaging chemotherapy agents tend to work comparably across humans and dogs, veterinary oncologists have reasonably borrowed human anticancer drugs. But for targeted agents such as thalidomide, whose action depends on the precise structure of a single protein, species differences in that target can silently nullify the drug’s effect. Given thalidomide’s tragic history and its strict regulation in human medicine, the authors advocate that molecular evidence of function should be established before the drug is used therapeutically in dogs, a lesson that applies equally to the growing arsenal of cereblon-targeting protein degraders now entering human clinical development.

Subject of Research: Genetic variation in the canine cereblon protein and its implications for thalidomide therapy in dogs with cancer

Article Title: Genetic evidence supports that thalidomide should not be used therapeutically in dogs

Article References: Arendt, M. L., & Meadows, J. R. S. (2026). Genetic evidence supports that thalidomide should not be used therapeutically in dogs. Veterinary Oncology, 3(1), Article 14. https://doi.org/10.1186/s44356-026-00067-2

Image Credits: AI Generated

DOI: 10.1186/s44356-026-00067-2

Keywords: thalidomide, cereblon, CRBN, dogs, canine cancer, veterinary oncology, comparative genomics, neo-substrates, ubiquitin ligase, drug repurposing, teratogenicity, Dog10K