SafeSense: An open-access safety atlas for antisense oligonucleotides adverse events in human

Abstract

Antisense oligonucleotides (ASOs) therapeutics are a powerful class of RNA-targeted therapy, with applications ranging from traditional development pipelines to individualized drugs. ASO therapeutics share common chemical architecture, including modifications and structure, thus enabling cross-program safety analyses, which provide insights into class-specific safety liabilities. Here, we present SafeSense, a safety atlas specifically dedicated to ASOs, integrating adverse event (AE) in human, encompassing data for 112 ASOs, curated in collaboration with the preclinical working group of the N = 1 collaborative (N1C). This resource was curated by aggregation and harmonization of data from clinical trials, regulatory documents, post-marketing data, and more. The data were used for internal analyses, and is now available as an open-access resource, available for public inquiries. Using advanced analyses, we identify several detriments of ASO safety, including structural configuration, sugar chemistries, administration route, and backbone compositions. Collectively, our findings provide a systematic overview of clinical ASO safety and reveal chemical architecture as a primary determinant of tolerability. By integrating heterogeneous safety data into a publicly available and analysis-ready framework, SafeSense provides a valuable resource for guiding ASO design, safety benchmarking, and clinical monitoring strategies, particularly useful for emerging personalized oligonucleotide therapies.

Graphical abstract

Vaknin, Ziv, Dominissini, Mor, and colleagues, with the N = 1 Collaborative, present SafeSense, an open-access, chemistry-aware atlas harmonizing human adverse-event data across antisense oligonucleotide therapies. By linking safety patterns to ASO structure, sugar chemistry, backbone, route, and indication, SafeSense provides benchmarks to guide safer oligonucleotide design and clinical monitoring.

Introduction

RNA-targeted oligonucleotide therapeutics have emerged as a powerful and rapidly expanding class of precision medicines for genetic and acquired diseases.^1,2,3,4 These therapies, including antisense oligonucleotides (ASOs) and small interfering RNAs (siRNAs), enable direct modulation of RNA abundance, splicing, and translation, providing a versatile platform capable of targeting virtually any transcript.

ASOs^5,6 are short, synthetic strands of nucleic acids designed to bind specifically to target RNA sequences, thereby influencing RNA splicing, promoting degradation via RNase H,⁷ or blocking regulatory sites to modulate gene expression.^8,9 Unmodified ASOs are rapidly degraded by nucleases and exhibit poor cellular uptake, necessitating the incorporation of chemical modifications to improve stability and pharmacokinetic properties. These include sugar modifications such as 2′-O-methyl (2′-OMe), 2′-O-methoxyethyl (2′-MOE),^10,11 and locked nucleic acids (LNAs),¹² as well as backbone modifications, most notably phosphorothioate (PS)¹³ linkages, in which a non-bridging oxygen atom in the phosphate group is replaced with sulfur (Figures 1A and 1B). While these innovations have enabled major clinical successes, safety and tolerability remain central challenges in ASO development.

Figure 1.

Schematic representation of ASO chemical modifications and structure

(A) Common sugar and backbone modifications. (B) Different ASO structures and design. The coding in form of color and shape is defined in the legend.

Adverse events (AEs) associated with ASO therapies span multiple physiological domains and include injection-site reactions,¹⁴ flu-like symptoms,¹⁵ hepatotoxicity,^13,16 thrombocytopenia, coagulation disturbances, hydrocephalus,¹⁷ and immune activation. Importantly, safety profiles are highly heterogeneous and influenced by multiple factors, including chemical composition, sequence design, mechanism of action, route of administration, dosing regimen, target biology, and patient population.¹⁸

Despite the growing number of approved and investigational ASO therapies, safety information remains fragmented across heterogeneous sources, including clinical trial records and publications, regulatory labels, post-marketing databases such as the FDA Adverse Event Reporting System (FAERS¹⁹), scientific articles, and gray literature sources such as investor communications, press releases, conference abstracts, and non-peer-reviewed reports. These sources differ substantially in terminology, reporting granularity, and data structure, limiting cross-study comparisons and hindering systematic safety evaluation. Existing reviews and safe studies of ASO have largely focused on specific toxicity domains¹⁰ or administration routes,²⁰ typically emphasizing approved drugs and published trials without integrating structured treatment-level data.

To address this gap, we developed SafeSense, a comprehensive, chemistry-aware database designed to systematically harmonize AE data across the ASO therapeutic landscape. SafeSense integrates information from clinical trials, post-marketing reports, regulatory documents, FAERS records, and gray literature into a unified, analysis-ready framework. The database links AEs to detailed treatment metadata, including chemical modifications, structural design features, administration routes, and mechanistic classes, while preserving source traceability.

Through a combination of automated text mining and expert manual curation, SafeSense standardizes AE terminology, resolves inconsistencies across data sources, and enables quantitative comparisons of safety patterns across chemistries, delivery strategies, and clinical contexts. By providing an openly accessible and dynamically updatable platform, this resource aims to support evidence-based decision-making in ASO design, preclinical development, and clinical trial planning, particularly for academic and non-profit groups developing personalized oligonucleotide therapies.

Results

Dataset assembly and overall characteristics

A total of 3,396 records were initially identified across multiple data sources, including peer-reviewed publications (n = 2,778), clinical trial registries (n = 533), gray literature sources (n = 62), prescribing information documents (n = 13), and FAERS database queries (n = 10) (Figure 2A). Following eligibility screening, 616 records met inclusion criteria and were retained for the final curated dataset, comprising 323 peer-reviewed articles, 209 clinical trial registry records, 62 gray literature sources, 13 prescribing information documents, and 9 FAERS records (Figure 2A). The database search, eligibility screening, and analyses were last updated in February 2026.

Figure 2.

Identification, screening, and characteristics of data sources included in the SafeSense database

(A) Flow diagram illustrating the identification and screening process of records used to construct the database. Database search, eligibility screening, and analyses were last updated in February 2026. (B) Characteristics of evaluable clinical trial and real-world safety sources after curation and harmonization, showing distribution by source type (top) and by clinical phase or reporting category (bottom). Because multiple records may refer to the same study, treatment, or safety dataset, source-type counts in (B) are not expected to match the initial record counts shown in (A). (C) Pie chart showing the proportion of treated patients contributed by each ASO therapy. Percentages represent the share of the total treated cohort (n = 41,730 patients). Individual ASOs contributing <1% of the total cohort were grouped into a single cumulative category (<1% of subjects), representing 14.5% of treated patients. Treatments administered under N-of-1 protocols are shown separately and account for 0.1% of the total cohort. (D) Pie chart showing the distribution of treated patients stratified by route of administration across all therapies. Percentages represent the share of the total treated population, including systemic administration (s.c. and i.v.), intrathecal (i.t.), intravitreal (i.v.t.), and other routes.

Among peer-reviewed publications, non-human studies and reports lacking evaluable safety data were excluded, including articles focused solely on efficacy, study design, or interim analyses without consistent AE reporting. Clinical trial registry entries were excluded if recruitment had not yet commenced, if safety results were unavailable, or if trials were withdrawn for non-safety reasons. One FAERS dataset for fomivirsen was excluded due to inconsistencies and limited interpretability related to its historical reporting context.

Among the included records, 246 clinical trial and real-world reporting sources provided evaluable safety information, representing 112 ASO treatments. AEs were reported in 221 sources corresponding to 87 treatments, while an additional 25 sources explicitly reported the absence of AEs. The majority of safety data originated from peer-reviewed studies (n = 183), followed by gray literature (n = 41), prescribing information documents (n = 13), and FAERS reports (n = 9) (Figure 2B).

The included safety sources spanned multiple stages of clinical development, with the largest representation from phase 1 and phase 2 trials (n = 145). Additional contributions included unspecified phase, including N-of-1, phase 3 trials, post-marketing data including FAERS reports, labeling documents, and a smaller number of studies from mixed or unspecified categories (Figure 2B).

Overall, the curated dataset encompassed safety data from 41,730 treated patients across 112 ASO therapies (Figures 2C and 2D). A total of 42,092 AEs were reported, comprising 1,734 distinct AE terms. Of these, 7,072 (16.8%) events were classified as severe. AEs were reported for 87 treatments, of which 58 treatments had documented severe AEs (SAEs). The data is available as an open-access resource, available for public inquiries (https://aso-adverse-events.streamlit.app).

Across the dataset, AEs were most frequently observed within general, gastrointestinal, and procedure-related clinical categories, with injection-site reactions representing the most commonly reported individual AE term. Intrathecal (i.t.), subcutaneous (s.c.), and intravenous (i.v.) administration routes accounted for the majority of treatments represented in the database (Figure 2D). Consistent with this distribution, i.t. and systemic administration together accounted for the largest proportions of treated participants and observed AEs, including severe events (Figure 3).

Figure 3.

Overview of adverse event distribution and treatment characteristics in the SafeSense database

(A) Top adverse event (AE) groups ranked by total observed event counts across all treatments. (B) Most frequently reported individual AE terms ranked by total observed counts. (C) Distribution of severe and non-severe AEs across the dataset. (D) Distribution of administration routes among included ASO treatments, shown by treatment count. Abbreviations: i.t., intrathecal; s.c., subcutaneous; i.v., intravenous; i.v.t., intravitreal. (E) Pie chart showing the distribution of observed AEs stratified by route of administration. Percentages represent the proportion of total AE events attributed to each route, including systemic administration (i.v. and s.c.), i.t., i.v.t., and other routes.

(F) Pie chart showing the distribution of severe AE stratified by route of administration. Percentages represent the proportion of total severe AE events attributed to each route.

Safety profiles by ASO structural configuration

Gapmer ASOs are associated with higher adverse event rates than fully modified ASOs

To evaluate whether ASO structural design influences safety outcomes, we compared AE profiles between gapmer and fully modified (FM) configurations across the dataset. Gapmers represent the most common RNase H-activating architecture, whereas FM ASOs typically lack a central DNA gap and are primarily designed for steric-blocking mechanisms.

Across all treatments, gapmers demonstrated significantly higher overall AE prevalence compared with FM ASOs (Figure 4A). Radar-based analyses revealed broadly elevated AE rates across multiple clinical domains among gapmers, particularly within procedure-related and neurologic categories (Figure 4B). FM ASOs exhibited a lower overall AE burden, with modest relative representation in certain non-procedural AE domains.

Figure 4.

Structural determinants of ASO safety: gapmer vs. FM architecture and nucleotide length

(A) Coefficient plots showing weighted least squares (WLS) regression estimates comparing adverse event (AE) rates between gapmer and FM ASO structures within treatment subsets; red = higher AE rate in gapmer ASOs, blue represent higher AE rate in FM ASOs. (AI) Differences in overall AE rates. (AII) Differences in severe AE rates. Coefficients represent the difference in accumulated AE rates (FM-gapmer) with 95% confidence intervals. Negative values indicate higher AE rates associated with gapmer designs, whereas positive values indicate higher rates associated with FM ASOs. Treatment subsets are defined based on route of administration and modification context. The number of included treatments (n) and total event weights are indicated for each comparison. (B) Radar plots illustrating the distribution of AE rates across clinical system-organ classes comparing gapmer and FM ASOs under different treatment contexts. (BI) All reported AEs across all administration routes, (BII) Intrathecal (IT) administration only, (BIII) Systemic administration, (BIV) Intravitreal (IVT) administration, (BV) Subset of treatments containing 2′-MOE sugar modifications. (C) Scatterplot showing the relationship between ASO sequence length and mean safety rate (percent of patients experiencing AEs), with simple linear regression with 95% confidence interval. Associations between ASO length and safety rate were assessed using Pearson and Spearman correlation tests. ∗p <0.05, ∗∗p <0.01, and ∗∗∗p <0.001.

These comparisons were derived from exposure-weighted weighted least squares (WLS) regression models, allowing safety estimates to reflect treatment population sizes rather than assigning equal influence to highly heterogeneous studies. WLS analyses confirmed significantly higher overall AE rates among gapmers across the dataset, including after exclusion of N-of-1 treatments (Figure 4AI).

Stratified analyses indicated that this structural safety difference was most pronounced among intrathecally administered treatments. Within the i.t. subgroup, gapmers exhibited substantially higher AE rates than FM ASOs across multiple AE categories, including pronounced differences in procedure-related and neurologic domains (Figure 4BII), with corresponding WLS models demonstrating significant effect estimates. In contrast, intravitreal (IVT) treatments showed no significant structural differences, likely reflecting limited sample sizes within this subgroup.

Further stratification by chemical modification demonstrated that this structural effect persisted within 2′-MOE-modified ASOs, where gapmers continued to exhibit higher AE rates across multiple clinical domains relative to FM counterparts (Figure 4A, BV).

Because nusinersen accounts for a substantial proportion of treated participants and is concentrated within the FM, 2′-MOE, and i.t. strata, we performed additional sensitivity analyses to evaluate the influence of a single dominant ASO. These included equal-ASO-weighted analyses, random-effects meta-regression with ASO-level structure, and leave-one-out analyses excluding nusinersen, each repeated with and without N-of-1 treatments (Figure S1). Overall AE rates showed larger exposure-weighted differences between gapmer and FM ASOs but were more sensitive to weighting strategy and N-of-1 inclusion. In contrast, SAE outcomes showed smaller and less frequently significant effect estimates, as expected given their lower event frequency, but the direction of effect was more consistent across sensitivity analyses, particularly after exclusion of N-of-1 treatments. Random-effects meta-regression was less informative in sparse strata, where few FM agents beyond nusinersen contributed substantial clinical exposure. These findings support the primary WLS interpretation while indicating that the gapmer-FM comparison is best viewed as an exposure-weighted clinical safety association rather than a definitive agent-independent class effect.

When analyses were restricted to SAEs, gapmers generally continued to show higher event rates than FM ASOs. As expected, absolute effect sizes were smaller, and statistical significance was reduced in several strata because SAEs were less frequent than overall AEs. However, the direction of effect was more consistent across sensitivity analyses, particularly after exclusion of N-of-1 treatments. Thus, while overall AE rates showed larger differences, SAE outcomes provided the more stable and clinically meaningful support for the gapmer-FM association.

Collectively, these results identify ASO structural configuration as an important contributor to reported safety patterns, with gapmer architectures associated with higher AE burden compared with FM designs. However, because several FM strata are strongly influenced by nusinersen, these comparisons should be interpreted as exposure-weighted clinical associations rather than definitive agent-independent class effects.

ASO length is not associated with AE burden

To determine whether sequence length influences ASOs’ safety, we examined the relationship between nucleotide length and AE burden across treatments. ASO lengths showed a narrow distribution centered around ∼20 nucleotides, consistent with typical therapeutic design constraints (Figure S2A). When analyzed as a continuous variable, nucleotide length demonstrated no significant association with mean AE ratio, with both Pearson and Spearman correlation analyses indicating weak and non-significant relationships (Figure 4C).

To further assess potential nonlinear effects, ASOs were stratified into discrete length bins (≤19 nt, 20–24 nt, and ≥25 nt). Comparative analyses revealed no statistically significant differences in AE ratios across length categories (Figure S2B), as confirmed by both parametric and non-parametric tests.

Collectively, these findings indicate that nucleotide length is not a major determinant of clinical toxicity in therapeutic ASOs. Importantly, this observation supports a structural model of ASO safety, whereby AE are primarily driven by chemical architecture, rather than sequence-specific properties. If toxicity were predominantly mediated by hybridization-dependent off-target effects, longer ASOs would be expected to demonstrate improved specificity and reduced AE burden – a trend that was not observed.

Sugar chemistry influences ASO safety profiles

To investigate the contribution of sugar chemistry to ASO safety, we analyzed AE profiles across major sugar modification classes.

Analysis of accumulated AE rates across clinical domains revealed substantial heterogeneity in overall AE burden between sugar chemistries, with clear clustering patterns according to modification type (Figure 5A). Among all classes, phosphorodiamidate morpholino oligomers (PMO) and 2′-MOE-modified ASOs exhibited the lowest overall AE rates, including when compared with unmodified DNA (N) ASOs. In contrast, higher AE burdens were observed for LNA and other sugar chemistries. These differences were particularly evident within general, infections, and neurologic AE categories, indicating domain-specific safety effects associated with sugar modifications.

Figure 5.

Adverse event profiles across sugar modifications

(A) Stacked bar plots showing accumulated adverse event (AE) rates (%) across clinical AE groups stratified by oligonucleotide sugar modification. (AI) Overall distribution across all treatments, (AII) Distribution stratified to gapmer designs only. Accumulated rates represent the summed percentage of observed AE events within each clinical group for a given sugar modification. Colors indicate clinical AE categories. (B) Weighted least squares (WLS) regression comparisons of AE rates between LNA and 2′-MOE modifications. (BI) Differences in overall AE rates, (BII) Differences in severe AE rates. Coefficients represent the difference in accumulated AE rates (LNA and 2′-MOE) with 95% confidence intervals. Positive values indicate higher AE rates associated with LNA (blue), whereas negative values indicate higher rates associated with 2′-MOE (red). Analyses are stratified by route of administration and by inclusion or exclusion of N-of-1 treatments. The number of included treatment cohorts (n) and total event weights are indicated for each comparison. ∗∗p <0.01, ∗∗∗ p < 0.001, ∗∗∗∗p <0.0001.

Stratification by administration route demonstrated that the favorable safety profile of 2′-MOE-modified ASOs was preserved across both i.t. and systemic (i.v./s.c.) contexts, whereas PMO-modified ASOs were represented exclusively within systemic administration (Figure S3A). Notably, the magnitude of the 2′-MOE safety advantage was more pronounced within IT-delivered treatments and remained detectable, albeit to a lesser extent, under systemic administration.

Because prior analyses identified gapmer architecture as a primary determinant of toxicity, we next restricted analyses to gapmer ASOs to assess whether sugar chemistry exerted independent effects (Figure 5AII). Within this subset, 2′-MOE-modified gapmers continued to demonstrate lower AE rates relative to other sugar chemistries, confirming that sugar modifications contribute independently to safety beyond structural configuration.

To directly compare the two most widely used sugar chemistries in gapmer ASOs, we performed exposure-weighted WLS regression analyses comparing LNA and 2′-MOE modifications (Figure 5B). Across all treatments, LNA-modified ASOs exhibited significantly higher overall AE rates compared with 2′-MOE-modified ASOs. This difference persisted after stratification by administration route and after inclusion or exclusion of N-of-1 studies.

When analyses were restricted to SAEs, the overall pattern was preserved, although effect sizes were attenuated. Consistent with the overall AE analysis, the strongest differences between LNA and 2′-MOE were observed within i.t. administration, while smaller but consistent differences remained detectable in systemic subsets (Figure 5BII).

Collectively, these findings identify sugar chemistry as an independent determinant of ASO safety, with 2′-MOE and PMO modifications demonstrating the most favorable tolerability profiles, and LNA associated with comparatively higher AE burden across multiple analytical approaches.

Backbone composition and AE patterns

Because PS linkages represent the dominant backbone chemistry across nearly all therapeutic ASOs, we focused our analyses on the principal source of backbone variation with sufficient data availability: the inclusion of phosphodiester linkages within PS backbones (PO/PS) compared with PS-only designs.

Across all treatments, PO/PS configurations were associated with higher overall AE rates compared with PS-only designs in the WLS analysis, including several route- and chemistry-defined subgroups (Figure 6A). This indicates an increased overall reported AE burden associated with PO/PS backbones in the current dataset.

Figure 6.

Comparison of adverse event rates between PS and mixed PO/PS backbone

(A) Coefficient plots showing weighted least squares (WLS) regression estimates comparing adverse event (AE) rates between PS and PO/PS backbone modifications within treatment subsets; (AI) Differences in overall AE rates, (AII) Differences in severe AE rates. Coefficients represent the difference in accumulated AE rates (PS − PO/PS) with 95% confidence intervals. Positive values indicate higher AE rates associated with PS backbones (blue), whereas negative values indicate higher rates associated with PO/PS backbones (red). The number of included treatments (n) and total event weights are indicated for each comparison. (B) Radar plot showing the distribution of AE categories comparing PS and PO/PS backbone treatments based on the proportion of treated patients with reported AEs. Values are displayed as log1p-transformed percentages. (C) Bar plot showing accumulated AE rates across clinical groups for 2′-MOE-modified treatments stratified by backbone type (PO/PS vs. PS), displayed on a log1p scale. ∗p < 0.05, ∗∗p < 0.01.

When analyses were restricted to SAEs, this association was less evident, with attenuated effect estimates and less consistent statistical significance across strata (Figure 6AII). Thus, the PO/PS-associated signal appears to primarily reflect overall reported AE frequency or non-severe tolerability burden rather than a consistent increase in severe toxicity.

To evaluate whether backbone chemistry influenced the qualitative distribution of AEs, we examined AE patterns across clinical categories. The overall distribution of AE categories was broadly comparable between PS and PO/PS ASOs (Figure 6B).

To reduce potential confounding by sugar chemistry, we next restricted analyses to 2′-MOE-modified gapmer ASOs. Within this subset, AE category patterns remained generally consistent between backbone configurations, with slightly higher representation of several AE categories observed in PS-only treatments. However, neurologic and procedure-related AE categories remained relatively more pronounced in PO/PS-modified ASOs (Figure 6C).

Taken together, these findings suggest that variation in phosphodiester inclusion within PS backbones is associated with relatively modest and context-dependent differences in AE patterns.

ASO chemistry and structure safety patterns persist after disease-indication adjustment

Because disease indication may confound comparisons across ASO chemistries and structural classes, we repeated the primary exposure-weighted WLS models with curated disease-indication category included as a covariate. Exact primary indication was not used as a covariate because it was nearly unique to individual ASOs, which would make the model comparable to adjusting for ASO identity and would provide limited interpretable within-category variation. After adjustment by broader disease category, the main chemistry- and structure-associated safety patterns were not attenuated, including the gapmer versus FM, LNA versus 2′-MOE, and PO/PS versus PS comparisons (Figure S4). These findings suggest that the observed associations are not explained solely by broad disease-indication differences and support chemistry and structural configuration as contributors to reported ASO safety patterns.

Safety profiles by route of administration

Intravenous administration is associated with higher severe and hematologic AE burden than s.c. administration

To assess whether systemic route of administration influences safety outcomes, we compared AE profiles between i.v. and s.c. delivery.

Across all treatments, i.v. and s.c. administration showed broadly similar overall AE rates (Figure 7AI). However, SAE rates were significantly higher among IV-administered treatments compared with SC-administered treatments (p = 0.04383), indicating a route-associated difference in severe reported AE burden.

Figure 7.

Route-dependent patterns of systemic and neurologic adverse events in ASO therapies

(A) Comparison of treatment-level AE rates between intravenous (IV) and subcutaneous (SC) administration. (AI) Overall AE rates and severe AE rates by route. (AII) Hematologic AE rates and severe hematologic AE rates by route. Boxplots show medians and interquartile ranges, with individual ASOs shown as points. Route differences were assessed using Welch’s t test. (B) Distribution of non-neurologic systemic AEs observed following intrathecal (IT) administration; (BI) overall AE prevalence by clinical AE group, (BII) severe AE prevalence within the same AE categories. (C) Distribution of neurologic AEs observed following systemic administration (i.v. and s.c. combined), shown as accumulated AE prevalence across specific neurologic event terms.

We next examined hematologic AEs specifically, given prior suggestions that systemic exposure may differentially affect hematologic parameters.²¹ Hematologic AE rates were higher among IV-administered treatments compared with SC, and this difference was significant when analyses were restricted to severe hematologic events (p = 0.01401) (Figure 7AII). Analysis of individual hematologic AE terms suggested that this pattern was partly driven by thrombocyte-related and coagulation-associated events (Figure 7AIII).

Collectively, these results indicate that i.v. and s.c. systemic administration routes exhibit broadly comparable overall safety profiles, with only modest, domain-specific variations in hematologic event patterns.

Cross-compartment AE patterns suggest limited peripheral toxicity after i.t. delivery and limited CNS toxicity after systemic delivery

To explore whether route-specific administration is associated with cross-compartment safety signals, we examined two complementary questions: (1) whether non-neurologic systemic AEs occur following i.t. delivery, potentially suggesting peripheral exposure and (2) whether neurologic AEs are observed following systemic administration, potentially indicating CNS exposure (Figures 7B and 7C).

Among IT-administered treatments, non-neurologic AEs were observed across multiple clinical categories (Figure 7BI). However, the majority of these events were procedure-related or nonspecific in nature (e.g., general or mortality categories), and many are plausibly attributable to the underlying disease context or the invasive administration procedure itself. When analyses were restricted to SAEs, the representation of non-neurologic systemic categories was further reduced (Figure 7B II), suggesting that clinically significant peripheral toxicity following i.t. delivery is uncommon within the present dataset.

Conversely, we assessed the prevalence of neurologic AE terms among systemically administered treatments (Figure 7C). Although neurologic events were reported, the most common terms included headache, dizziness, and fall, events that are relatively nonspecific and not necessarily indicative of direct CNS toxicity. Additional reported events, such as neuropathy, may reflect peripheral nervous system involvement rather than central exposure.

Taken together, these findings suggest limited cross-compartment safety signals between i.t. and systemic administration. The presence of some non-neurologic AEs following i.t. delivery may be compatible with low-level peripheral exposure. In contrast, neurologic AEs observed with systemic administration were predominantly nonspecific and do not suggest meaningful CNS exposure. While AE data cannot directly reflect pharmacokinetics, the overall patterns are consistent with minimal systemic exposure after IT delivery and likely negligible CNS exposure following systemic administration.

Clinical trials report higher AE frequencies than post-marketing sources

To evaluate whether AE patterns differ by data source, we compared treatment-level AE rates between clinical trial reports and post-marketing long-term data, including observational cohorts and FAERS records.

Mean AE rates were significantly higher in clinical trial datasets compared with post-marketing sources (Figure 8AI). A similar but less pronounced pattern was observed for SAEs (Figure 8AII). Exposure-weighted WLS analyses confirmed this difference across both systemic and i.t. administration routes (Figures S5A and S5B).

Figure 8.

Adverse-event landscapes across data sources, clinical indications, and GalNAc conjugation status

(A) Comparison of treatment-level mean adverse event (AE) percentages between real-world and clinical trial data sources.; I. Overall AE rates, (AII) severe AE rates. Each point represents a treatment; boxes indicate median and interquartile range. Each point represents one treatment; boxplots show medians and interquartile ranges. Group differences were assessed using Welch’s t test. ∗∗∗∗p <0.0001. (B) Heatmap showing the distribution of AE categories across treatment indications. Values represent the mean proportion of total observed AEs attributed to each AE category within each clinical indication. (C) Radar plot comparing the distribution of AE clinical groups between GalNAc-conjugated and unconjugated systemically administered ASO treatments.

Overall, these findings suggest more frequent AE reporting in clinical trial settings relative to post-marketing data, likely reflecting differences in monitoring intensity and structured reporting requirements rather than intrinsic differences in treatment safety.

AE distribution by clinical indication

To evaluate whether AE patterns aligned with underlying disease context, we examined the distribution of AE clinical groups across treatment indications (Figure 8B). For example, neurologic AEs were most prominent in neurodegenerative and neurodevelopmental indications, consistent with the underlying disease context.

These findings support the internal consistency of the curated dataset and indicate that AE patterns generally correspond to the treated disease category.

GalNAc conjugation and AE patterns

To evaluate whether liver-targeting conjugation influences safety profiles, we compared AE distributions between GalNAc-conjugated and non-conjugated systemic ASOs (Figure 8C).

Descriptive analyses demonstrated modest differences in AE patterns. GalNAc-conjugated ASOs showed relatively lower representation of hematologic, inflammatory, and neurologic AE categories compared with non-conjugated systemic treatments, while overall AE distributions across most other clinical domains were broadly similar.

Given that several GalNAc and non-GalNAc ASOs within the dataset share similar sequences and therapeutic indications, these observations suggest that conjugation status may be associated with subtle differences in systemic AE patterns. However, the magnitude of these differences was modest, and further controlled comparative analyses would be required to establish their clinical significance.

Discussion

In this paper, we present SafeSense, an ASO-centered safety database that aggregates clinical trials, post-marketing, prescribing information, and additional safety data to systematically analyze AE patterns in humans. By harmonizing heterogeneous sources and integrating ASO-level metadata, including structure, sugar chemistry, backbone composition, route of administration, and indication, we provide a framework for large-scale interrogation of ASO safety. This approach is particularly relevant for academic laboratories and non-profit groups developing individualized or N-of-1 ASOs, where prior clinical safety data for the exact sequence or indication are usually unavailable.^22,23

SafeSense includes data from more than 41,000 treated participants and 112 ASOs. The overall AE spectrum was consistent with the known safety profile of ASOs, including injection-site reactions, flu-like symptoms, neurologic events, and laboratory abnormalities. The predominance of non-severe AEs relative to severe events is consistent with the generally favorable tolerability profile of clinically advanced ASOs. In addition, the large number of ASO programs reaching phase 3 suggests that many candidates successfully pass early safety-centered development stages.

A key advantage of SafeSense is that harmonization across programs enables comparisons that are not feasible from individual trial reports alone. In addition to the curated dataset, SafeSense includes a web-based interface designed to improve transparency, usability, and future expandability.

AEs associated with ASO therapies may arise through several non-mutually exclusive mechanisms. These include hybridization-dependent effects such as on-target pharmacology, exaggerated target knockdown, and off-target RNA binding. Other contributors include RNase H1-mediated transcript degradation in gapmer designs, chemistry- and backbone-dependent protein interactions, innate immune activation, and route- or procedure-related effects. The underlying disease and treatment indication may also strongly influence the observed AE profile, particularly in severe neurological disorders, oncology indications, and N-of-1 settings. Accordingly, SafeSense analyses were structured around major ASO design and clinical features, including structure, sugar chemistry, backbone composition, route of administration, and indication. Importantly, disease-category-adjusted WLS analyses did not attenuate the main chemistry- and structure-associated safety patterns, supporting the interpretation that ASO design features contribute to reported safety profiles beyond broad indication differences. However, the database captures reported AE patterns rather than experimentally proven mechanisms, and mechanistic interpretation should therefore remain cautious.

Such targeted analysis showed a better safety profile for FM ASOs compared with RNAse-H1 recruiting gapmer. This finding is consistent with known mechanistic differences between the two ASO architectures. Gapmer ASOs induce RNase H1-mediated cleavage of target RNA, a catalytic and irreversible process that may amplify the consequences of unintended binding to partially complementary transcripts. In contrast, FM ASOs generally act through steric blocking, splice modulation, or interference with functional RNA motifs, so binding alone is less likely to produce transcript degradation unless it occurs at a functionally relevant site. Additional chemistry-dependent mechanisms may also contribute, as the central DNA gap and PS-rich backbone of many gapmers can mediate interactions with cellular proteins linked to sequence- and chemistry-dependent toxicities.^13,24 Although nusinersen contributes substantially to the FM 2′-MOE i.t. cohort, sensitivity analyses supported the primary findings most consistently for severe AEs. Thus, the gapmer-FM comparison should be interpreted as an exposure-weighted, clinically relevant safety association, with cautious mechanistic interpretation.

Our analysis also compared different sugar modifications, showing a preferable safety profile for 2′-MOE-modified ASOs, compared with LNA. 2′-MOE provides potency and stability, but in a moderate-affinity, whereas LNA higher affinity^25,26 increases residence time and raises probability for partial matches. In both cases, it is speculated that higher affinity can result in increased toxicity, creating a subtle balance between efficacy and safety.

Next, we evaluated the possible contribution of the ASO length to its safety; contrary to some expectations, analysis does not show a significant correlation of length with overall AE burden and severity. This suggests that length alone is unlikely to be a primary driver of clinical toxicity, unlike other factors, such as backbone and chemical modifications. Thus, sequence-specific off-target effects remain plausible toxicity-driving aspect, but are likely not explained by length alone. These findings are particularly relevant as reduction of gap length in RNAse-H1 ASO designs is recognized as specificity increasing, thus utilized often in cases of allele-specific targeting.^22,27,28

Backbone composition is another important determinant of ASO design. In our analysis, mixed PO/PS configurations were associated with higher overall reported AE rates than PS-only designs in several WLS comparisons. However, this association was less consistent for severe AEs, where effect estimates were attenuated, and statistical significance was less stable. This suggests that the PO/PS-associated signal may mainly reflect non-severe tolerability or reporting burden rather than a robust increase in severe toxicity. This finding should be interpreted cautiously because backbone composition is tightly linked to ASO chemistry, indication, route of administration, and individual treatment design.

Translational studies further support the relevance of backbone chemistry to ASO tolerability. Reduced-PS or mixed-backbone designs can retain activity while improving tolerability, as demonstrated for a mixed-backbone C9ORF72 ASO.²⁹ Mechanistically, PS linkages have been linked to nonspecific protein binding and chemistry-driven acute neurotoxicity, with recent data suggesting that lowering PS content can mitigate some of these effects.³⁰ However, these findings do not necessarily translate directly to aggregate clinical AE patterns, because backbone composition is confounded by chemistry, route, indication, and candidate selection. Moreover, PS-related toxicities may be eliminated before clinical testing, resulting in a clinical dataset enriched for better-tolerated ASOs.

Our next targeted analysis focused on different injection routes, as route-dependent exposure has a dramatic effect on the severity and nature of AE. Systemically administered ASOs (SC/IV) generally exhibit poor blood-brain-barrier (BBB) permeability,^31,32 yet, specific neurologic AE, mainly headache and neuropathy, were also reported for ASOs administered in these routes. Headache may reflect a nonspecific systemic tolerability signal, often a part of flu-like symptoms and injection/infusion reactions,^24,33 whereas neuropathy can be plausibly linked to peripheral sensory neuron (ganglion/DRG) exposure.³⁴ Notably, a distinct category of true CNS neurotoxicity after systemic dosing has been demonstrated for the s.c. ASO RGLS4326, where CNS signs were mechanistically traced to hybridization-independent AMPA receptor antagonism, a liability eliminated by single-nucleotide modification.³⁵ Importantly, this CNS-related toxicity was seen in chronic administration in NHP and mice, but not in humans.

Moreover, although i.t. ASO administration is intended to localize its exposure to the CNS, multiple clinical pharmacokinetic analyses and regulatory reviews demonstrate measurable transfer of IT-dosed ASOs from the CSF to systemic circulation,^36,37,38 with plasma exposure reflecting CSF clearance pathways rather than targeted peripheral distribution.³⁶ Consequently, non-CNS related AEs remain relevant, even though overall systemic toxicity risk is lower than with systemic (SC/IV) dosing.³⁹

In our i.v. versus s.c. analysis, i.v. dosing was associated with higher severe AE rates, particularly severe hematologic AEs. Another AE that appears in higher incidence following i.v. treatments in leukopenia; however, this could be confounded by the large proportion of cancer-targeting ASOs are administered intravenously. These patients are usually co-treated or past-treated with established treatment protocols, commonly associated with bone marrow suppression.

A key strength of SafeSense is the integration of clinical trial and post-marketing data. Real-world sources showed lower AE frequencies than clinical trials, particularly for non-severe events. This likely reflects differences in reporting intensity, follow-up, and surveillance practices rather than true differences in toxicity, consistent with the known reduction in AE reporting under passive compared with active surveillance.^40,41 Importantly, severity ratios were more concordant between data sources, supporting the use of SAE metrics as a more stable basis for cross-program comparisons and safety monitoring.

Beyond its value for retrospective analysis, SafeSense may be useful as a practical resource for prospective design of personalized N-of-1 ASO studies. By linking AE patterns to ASO-level features, SafeSense can help investigators define a priori which safety domains require focused monitoring, including clinical assessments, laboratory panels, neurologic evaluations, procedure-related follow-up, and visit frequency. It may also help distinguish expected procedure-associated events from potentially unexpected patient- or target-specific findings. Although SafeSense cannot replace standard toxicology or regulatory guidance, it can help expand monitoring panels and prioritize specific in vivo toxicology readouts.

Several limitations should be considered. Consolidation of heterogeneous sources inevitably reduces the resolution of some reports, particularly with respect to nominal values, timing, dose, and event-level detail. Dose level, dosing frequency, treatment duration, and cumulative exposure were not systematically captured across sources. Therefore, the current treatment-level analyses cannot distinguish acute administration-related reactions from AEs that emerge after repeated or prolonged dosing. Future versions should incorporate these variables when available.

Indication and target population also represent major inherent confounders, as AE profiles can differ substantially based on underlying disease, disease severity, comorbidities, and concomitant therapies. SafeSense also does not fully separate ASO-associated AEs from procedure-related events or manifestations of the underlying disease. However, in many ASO programs, especially severe neurologic indications, i.t. administration, and N-of-1 treatments, disease severity and route of administration are integral parts of the therapeutic setting. We therefore interpret these patterns as reflecting an indication – ASO interaction rather than isolated drug toxicity alone. Systematic placebo-based or natural-history correction was not feasible because many sources lacked appropriate comparator groups.

In summary, in this work, we present a chemistry-aware clinical safety database for ASO, integrating multiple resources, to enable systemic analysis of AE patterns. As ASO landscape expands toward increasingly diverse modalities, routes, and chemistries, the need for safety benchmarks increases. With SafeSense, we aim to provide the data and infrastructure for integrative analyses that will enable the discovery of such benchmarks and measurements, in order to both refine ASO design and optimize monitoring protocols.

Materials and methods

Data sources and collection

AE data were systematically collected from multiple heterogeneous sources to capture the safety landscape of ASO therapies. These included clinical trial registries and publications, regulatory prescription labels, post-marketing surveillance databases (including FAERS), scientific literature, and “gray” literature such as investor communications, press releases, conference abstracts, and non-peer-reviewed reports (Figure 2A). All records retained treatment identifiers and source provenance to ensure traceability.

Records were screened for eligibility based on relevance to ASO therapeutics and availability of evaluable human safety data. Records were excluded if they described non-human or non-ASO studies, lacked adverse-event information, represented efficacy-only or interim analyses without consistent safety reporting, corresponded to withdrawn trials for non-safety reasons, or contained uninterpretable duplicate or historical reporting. Because screening was performed iteratively across heterogeneous and partially overlapping sources, exclusion reasons were not prospectively tracked as mutually exclusive numerical categories. Therefore, the final flow diagram reports the total number of excluded records and lists the main exclusion reasons.

To facilitate comprehensive identification and extraction of safety data across distributed and partially structured sources, an AI-assisted retrieval framework was employed. Multiple large language model systems, including ChatGPT, Claude, Gemini, DeepSeek, Perplexity, and OpenEvidence, were used during the study period. These AI systems included a combination of publicly accessible platforms, subscription-based services, and internally developed private models. Additional in-house models were developed primarily to compare extracted data across AI systems and support consistency checks. Family model versions available between January 2025 and February 2026 were utilized to maximize retrieval coverage and cross-validation.

AI systems were used solely for information discovery, structured data extraction, cross-source comparison, and preliminary summarization, and were not used for analytical decision-making or interpretation. All candidate records underwent mandatory manual validation prior to database inclusion. Each entry (including AEs, treatment characteristics, chemical modifications, and references, etc.) was independently verified against primary source documents by domain-expert researchers.

Database curation and harmonization

The AE database was curated through a structured, multi-stage harmonization process designed to integrate heterogeneous safety data into a unified, analysis-ready resource. Raw AE records were consolidated across studies and data sources while preserving treatment identity, reporting context, and source provenance. The complete list of source references used for data extraction is provided in Table S1, and the canonical AE dictionary is provided in Table S2. Throughout curation, the guiding principle was to maximize information retention while ensuring internal consistency, reproducibility, and comparability across treatments.

All reported AEs were mapped to a canonical AE dictionary that defines standardized AE terms and assigns each to a higher-level AE group corresponding to physiological or clinical domains. This dictionary-driven approach resolved spelling variants, synonymous terms, and inconsistent reporting nomenclature across sources. Where multiple historical terms corresponded to a single canonical concept, records were harmonized accordingly to prevent artificial fragmentation of safety signals.

AE attribution to the drug or administration procedure was preserved as reported in the original sources and was not independently reassigned by the SafeSense authors. AEs were retained unless clearly reported as unrelated to the ASO or its administration procedure. When comparator data were available, AEs clearly more frequent in controls were not counted as ASO-associated unless the source authors indicated that ASO attribution could not be excluded.

Quantitative safety fields, including total treated subjects, observed event counts, and reported percentages, were systematically validated and standardized. Percentages were recalculated directly from underlying counts and denominators to ensure mathematical consistency and to avoid propagation of reporting artifacts.

Because AE severity is reported using heterogeneous scoring systems across clinical domains (e.g., CTCAE in oncology, cardiovascular composite endpoints such as MACE, and field-specific neurological severity scales), a unified binary severity classification was implemented. Severity status was determined by the reporting investigator based on the clinical context described in the source material and classified as either severe or non-severe to enable consistent cross-domain comparisons.

To address duplicate reporting across studies and data sources, AEs were aggregated using treatment- and source-aware rules. For each treatment within a given data source, exposed population sizes were estimated using a mode-based approach to identify the most representative denominator. Event counts were then aggregated at the treatment level while preserving stratification by AE term and severity. This strategy minimized double counting while retaining true recurrence of events across independent reports.

Derivation of treatment-level safety metrics

A treatment-level safety dataset (gold table) was generated by aggregating exposure and AE data across sources. For each ASO treatment, the number of treated subjects was collected from all reporting sources. Because exposure counts may differ between sources, for example, between planned enrollment in trial registries and actual treated populations reported in study results, these values were cross-validated and reconciled against primary source documentation. Verified treated populations were then summed across sources to determine the total exposed population for each treatment.

The number of patients observed with each reported AE was subsequently extracted and aggregated across sources while preserving AE term and severity classification. Using these reconciled exposure denominators and aggregated event counts, the prevalence of each AE was calculated for each treatment as the proportion of treated patients experiencing the event. Treatment metadata, including administration route, mechanism of action, chemical modifications, structural design features, conjugation status, nucleotide length, clinical indication, and N-of-1 status, were integrated to enable stratified safety analyses. Table S3 and the SafeSense web interface provide the complete ASO-level metadata used for all category-based comparisons.

Statistical analyses

All statistical analyses were performed at the treatment level using exposure-normalized AE prevalence metrics derived from the gold table. Descriptive statistics were used to summarize dataset composition, including counts of treatments, treated participants, AE terms, AE groups, and total observed events.

To account for substantial differences in treatment population sizes across studies, WLS regression models were used as the primary approach for estimating effect sizes. Each treatment was weighted according to the total number of treated subjects contributing safety data, ensuring that treatments with larger exposure populations exerted proportionally greater influence on model estimates. This exposure-weighting strategy was implemented to prevent small studies from disproportionately influencing safety comparisons and to more accurately reflect population-level safety patterns.

Comparative analyses of safety patterns across treatment characteristics, including administration route, chemical modifications, structural design features, mechanism of action, conjugation status, nucleotide length, clinical indication, and N-of-1 status, were conducted using nonparametric statistical tests. Mann-Whitney U tests were applied for two-group comparisons and Kruskal-Wallis tests for multi-group comparisons. Welch’s t tests were used where appropriate to evaluate differences in mean AE prevalence between groups with unequal variances.

Associations between continuous variables and safety outcomes were assessed using Pearson and Spearman correlation analyses. Statistical results are reported with corresponding effect size estimates, confidence intervals, and sample sizes where applicable.

Data and code availability

All data presented and curated are available at https://aso-adverse-events.streamlit.app.

Acknowledgments

The authors would like to thank additional N1C members for their valuable input and help during the curation of the resource and its analysis, especially Hélène Tran, Willeke van Roon-Mom, and Raúl Andrés Santamaría. N.M. is a schooler of the Nehemia Rubin Excellence in Biomedical Research, the TELEM Program at the Sheba Medical Center.

Author contributions

N.V., D.D., and N.M. conceptualized the idea and wrote the manuscript. N.V. performed extensive queries and curated the database. N.V., A.Z., T.C., S.A.-N., and N.M. manually validated and annotated the database. N.V., A.Z., and S.A.-N. analyzed the data with the help of NIC members. A.Z. developed the Strimlit web application.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used AI-based tools and services (including large language model systems such as ChatGPT, Claude, Gemini, DeepSeek, Perplexity, and OpenEvidence) to assist with data discovery, information retrieval, structured data extraction, and cross-source comparison. After using these tools, the authors manually reviewed, validated all content against primary sources, and edited it as needed. The authors take full responsibility for the published article’s content. For further details, see the materials and methods section (“data sources and collection”).