CoChat

A large-scale concordance study of toxicity findings across preclinical species and humans for small molecules and biologics

Open access
Liu, Xin, Fan, Fan · 2026 · Frontiers in Toxicology

The DOI matches its Crossref record; 81 of the 81 references it cites were found in a registry or an index; 43 of the 89 sources its claims cite were checked for support.

Abstract

Translating preclinical safety findings into reliable insights for human risk assessment remains a fundamental challenge in drug development. Prior preclinical-clinical concordance studies have been constrained by limited drug coverage, reliance on identical-term matching for adverse events (AEs), and limited consideration of species, modality, exposure, and biological or mechanistic context. To address these gaps, we assembled a large cross-species concordance dataset, integrating standardized preclinical and clinical safety data for 7,565 marketed and investigational drugs from PharmaPendium and OFF-X. Our framework employs likelihood ratios to reduce prevalence bias and extends concordance assessment beyond identical-term matches to include semantically and mechanistically related AE pairs. Stratified analyses by species, modality, and exposure-matched subsets further refined translational relevance, while integration of on- and off-target annotations supports mechanistic interpretation and potential screening. Using this approach, we identified 850 significant identical-term AEs and 2,833 additional unique endpoints from cross-term associations. To promote reproducibility and transparency in animal research, we provide open access to the analytic code and statistical results via an interactive web application. An accompanying multi-agent AI system (ToxAgents) enables standardized querying and interpretation of concordance results. Together, these resources extend previous foundational efforts and establish a shared, data-driven platform to advance translational safety science, support evidence-based study design aligned with the 3Rs principles, and ultimately contribute to the development of safer medicines to improve human health.

Overview

CoChat Summary

TL;DR

This study assembled a harmonized cross-species safety dataset for 7,565 marketed and investigational drugs from PharmaPendium and OFF-X and used likelihood ratios, species/modality/exposure stratification, and cross-term mapping to assess preclinical-to-human toxicity concordance. It identified 850 significant identical-term adverse events and 2,833 additional unique endpoints, with generally low-to-moderate organ-system concordance but stronger phenotype-level signals, and released a web app plus the ToxAgents multi-agent AI system.

Key findings

  • Across all modalities, 850 significant identical-term phenotypes were identified, including 780 for small molecules, 237 for biologics, and 94 in the exposure-controlled subset; cross-term analysis added 2,833 unique endpoints (2,375 clinical, 854 preclinical, 396 shared).
  • At the SOC level, “infections and infestations” had the highest predictivity (LR+ >10) for both small molecules and biologics, followed by immune, renal/urinary, and musculoskeletal disorders with moderate LR+ <5; nervous system, psychiatric, investigations, general disorders, and pregnancy/perinatal conditions showed low LR+. Monkeys had higher median LR+ (SM 4.18, Bio 3.74) than other species, and most iLR– values were close to 1.
  • At the PT level, small-molecule examples included constipation (LR+ = 21.5), rash (20.0), neutropenia (16.1), and anemia (10.1); biologic examples included infection (116.0), rash (45.2), hyperglycemia (34.4), and neutropenia (21.8). In hepatobiliary/metabolic analyses, immune-mediated hepatitis (321.8), hyperinsulinemia (217.7), and hypertriglyceridemia (113.1) were highly concordant, whereas nonspecific terms such as hepatotoxicity (2.2), hepatic function abnormal (4.2), and liver disorder (3.2) were poor.
  • Compared with Giblin et al., 183 of 188 (97%) significant AEs were also significant in this dataset, with Pearson r = 0.61–0.88 across PPV, NPV, LR+, and iLR–; 70% of AEs with LR+ ≥10 in the prior study also met that threshold here. Cross-term semantic similarity quartiles showed median LR+ increasing from 5.6 to 6.7 (Kruskal–Wallis p = 1.8 × 10−100).

Methods

  • Data were compiled from PharmaPendium and OFF-X, with AEs coded in MedDRA; 13,949 unique AE terms were retrieved, then filtered to 7,565 drugs with at least one AE in both preclinical and clinical studies. Preclinical species were mouse, rat, rabbit, dog, and monkey; drugs were categorized as small molecule or biologic. Exposure-controlled comparisons required |Δlog10Cmax| or |Δlog10AUC| ≤1.
  • Concordance used 2 × 2 contingency tables, two-tailed Fisher’s Exact Test with Bonferroni correction within each SOC (adjusted p < 0.05), and likelihood ratios (LR+, iLR–) plus PPV and NPV. Identical-term and non-identical within-SOC AE pairs were assessed; OpenAI text-embedding-3 cosine similarity quantified semantic relatedness.
  • Mechanistic annotation used DrugBank and OFF-X on/off-targets, gprofiler with DisGeNET, STRING PPI networks, and MCODE. Results were made accessible via a Shiny web app and ToxAgents, a GPT-4.1/RAG multi-agent system using 26 ICH safety documents, with case studies and reproducibility/robustness checks.

Limitations

  • The paper states that source-data assignment of MedDRA terms, especially for preclinical behaviors or nuanced histopathology, may be subjective or inconsistent; missing AEs due to incomplete/selective reporting or extraction failure may mask translational signals, and OFF-X investigational-drug AEs may be neutral or lack severity because companies may not disclose more impactful findings.
  • The paper states that AEs are encoded as binary present/absent outcomes without frequency, severity, or explicit dose-response data, limiting use for clinical risk thresholds; biologics were not subtyped because current sources lack consistent subtype labels.
  • The paper states that target annotation can be confounded by complex pharmacology, multiple on- and off-targets, metabolites, or uncharacterized off-target interactions; it also notes that true negative counts are estimations because not all possible observations were measured or reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/ftox.2026.1731947

Cites

81 of 81 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1 Overview of data preparation and analysis workflow. AE data encoded in MedDRA terms are sourced from PharmaPendium and OFF-X API. Drugs are consolidated and categorized by modality, and pharmacokinetic (PK) data are retrieved for exposure-controlled analysis. Preclinical toxicity studies focus on mouse, rat, rabbit, dog, and monkey. Concordance is evaluated using Fisher’s Exact Test, with likelihood ratios (LR+, iLR−) and predictive values (PPV, NPV) assessing AE predictiveness. Drug targets/off-targets, pathways, and PPI networks are explored to provide mechanistic insights and to prioritize potential safety targets for further evaluation. Findings are made accessible through a web app and a multi-agent AI system for safety assessments.
FIGURE 1 Overview of data preparation and analysis workflow. AE data encoded in MedDRA terms are sourced from PharmaPendium and OFF-X API. Drugs are consolidated and categorized by modality, and pharmacokinetic (PK) data are retrieved for exposure-controlled analysis. Preclinical toxicity studies focus on mouse, rat, rabbit, dog, and monkey. Concordance is evaluated using Fisher’s Exact Test, with likelihood ratios (LR+, iLR−) and predictive values (PPV, NPV) assessing AE predictiveness. Drug targets/off-targets, pathways, and PPI networks are explored to provide mechanistic insights and to prioritize potential safety targets for further evaluation. Findings are made accessible through a web app and a multi-agent AI system for safety assessments.CC BY · PubMed Central
FIGURE 2 SOC-based AE frequency analysis across modalities and species. (A) Comparison of adverse event (AE) frequencies across SOCs in humans for biologics and small molecules. The bar graph indicates percentage distributions, and the heatmap visualizes absolute AE frequencies. (B) Corresponding AE distributions from preclinical studies. Bar graph and heatmap demonstrate relative and absolute AE occurrences, respectively, across modalities. (C) Difference in percentage between human and animal AE frequencies (calculated as human percentage minus animal percentage) stratified by SOC and drug modality, highlighting SOCs disproportionately represented in either human or animal studies. Positive values indicate SOCs more frequently reported in human studies; negative values indicate higher reporting in animal studies. (D) Species-specific and modality-specific detailed heatmap illustrating AE frequency percentages across SOCs. Darker shading reflects higher frequencies, providing detailed insights into interspecies and modality variability.
FIGURE 2 SOC-based AE frequency analysis across modalities and species. (A) Comparison of adverse event (AE) frequencies across SOCs in humans for biologics and small molecules. The bar graph indicates percentage distributions, and the heatmap visualizes absolute AE frequencies. (B) Corresponding AE distributions from preclinical studies. Bar graph and heatmap demonstrate relative and absolute AE occurrences, respectively, across modalities. (C) Difference in percentage between human and animal AE frequencies (calculated as human percentage minus animal percentage) stratified by SOC and drug modality, highlighting SOCs disproportionately represented in either human or animal studies. Positive values indicate SOCs more frequently reported in human studies; negative values indicate higher reporting in animal studies. (D) Species-specific and modality-specific detailed heatmap illustrating AE frequency percentages across SOCs. Darker shading reflects higher frequencies, providing detailed insights into interspecies and modality variability.CC BY · PubMed Central
FIGURE 3 Positive likelihood ratios at SOC level by species and modality. (A) LR+ values for combined species across SOCs. Bars depict LR+ for all modalities (green), small molecules (blue) and biologics (red). (B) Violin plots of LR+ distributions by species, comparing SM (blue) and Bio (red). Black dashed lines mark medians, with values displayed within each plot. Monkeys show notably higher LR+ compared with other species. (C) LR+ values for individual SOCs stratified by modality and species. Each point represents LR+ for a specific SOC–species pair in SM or Bio, with colors indicating species.
FIGURE 3 Positive likelihood ratios at SOC level by species and modality. (A) LR+ values for combined species across SOCs. Bars depict LR+ for all modalities (green), small molecules (blue) and biologics (red). (B) Violin plots of LR+ distributions by species, comparing SM (blue) and Bio (red). Black dashed lines mark medians, with values displayed within each plot. Monkeys show notably higher LR+ compared with other species. (C) LR+ values for individual SOCs stratified by modality and species. Each point represents LR+ for a specific SOC–species pair in SM or Bio, with colors indicating species.CC BY · PubMed Central
FIGURE 4 Concordance analysis at PT level between identical preclinical and clinical AEs. (A) UpSet plot illustrating the overlap of statistically significant AEs between animal and human. Bar heights represent the number of significant PT terms for each dataset: All modalities (blue), Small Molecule (orange), Biologics (green), along with the Exposure-Controlled subset (red). Connected black dots indicate intersections between datasets, with the bar above each set representing the count of significant PTs shared across those datasets. (B) Heatmap displaying the percentage of statistically significant PT-level AEs identified in each SOC, across species, and modalities. Brighter colors indicate higher counts of significantly concordant terms.
FIGURE 4 Concordance analysis at PT level between identical preclinical and clinical AEs. (A) UpSet plot illustrating the overlap of statistically significant AEs between animal and human. Bar heights represent the number of significant PT terms for each dataset: All modalities (blue), Small Molecule (orange), Biologics (green), along with the Exposure-Controlled subset (red). Connected black dots indicate intersections between datasets, with the bar above each set representing the count of significant PTs shared across those datasets. (B) Heatmap displaying the percentage of statistically significant PT-level AEs identified in each SOC, across species, and modalities. Brighter colors indicate higher counts of significantly concordant terms.CC BY · PubMed Central
FIGURE 5 Network visualization of significant AEs within the hepatobiliary system and metabolism/nutrition disorders, based on LR+. (A) Combined analysis of all drug modalities. Nodes represent individual AEs, and edge width corresponds to LR+ values. Red and blue nodes denotes the AEs with LR+ ≥10 and <10 respectively. (B) Stratified networks for small molecules (SM), biologics, and the exposure-controlled subset. Thicker edges denote higher LR+. (C) Species-level networks for small molecules, biologics, and the exposure-controlled subset, showing species-specific LR+ values across preclinical models. Edge width corresponds to LR+, and species icons indicate contributing data from mouse, rat, rabbit, dog, or monkey.
FIGURE 5 Network visualization of significant AEs within the hepatobiliary system and metabolism/nutrition disorders, based on LR+. (A) Combined analysis of all drug modalities. Nodes represent individual AEs, and edge width corresponds to LR+ values. Red and blue nodes denotes the AEs with LR+ ≥10 and <10 respectively. (B) Stratified networks for small molecules (SM), biologics, and the exposure-controlled subset. Thicker edges denote higher LR+. (C) Species-level networks for small molecules, biologics, and the exposure-controlled subset, showing species-specific LR+ values across preclinical models. Edge width corresponds to LR+, and species icons indicate contributing data from mouse, rat, rabbit, dog, or monkey.CC BY · PubMed Central
FIGURE 6 Cross-term AE concordance. (A) Distribution of LR+ across quartiles of semantic similarity for preclinical-clinical AE pairs. LR+ values and sample sizes are indicated above each box, with box colors scaled to the median value. (B) Sankey plot illustrating selected examples of significant, non-identical cross-species AE associations, with flow thickness proportional to LR+ value. (C) Bubble plots display enriched phenotypes among annotated targets and off-targets associated for drugs linked to vomiting (left) and seizure (right). Bubble size is proportional to the enrichment score.
FIGURE 6 Cross-term AE concordance. (A) Distribution of LR+ across quartiles of semantic similarity for preclinical-clinical AE pairs. LR+ values and sample sizes are indicated above each box, with box colors scaled to the median value. (B) Sankey plot illustrating selected examples of significant, non-identical cross-species AE associations, with flow thickness proportional to LR+ value. (C) Bubble plots display enriched phenotypes among annotated targets and off-targets associated for drugs linked to vomiting (left) and seizure (right). Bubble size is proportional to the enrichment score.CC BY · PubMed Central
FIGURE 7 AI-enabled framework for evaluating the translatability of preclinical AEs. The process integrates domain knowledge, including concordance logic and data schemas, mechanistic insights from on/off targets and Safetyome, and ICH safety guidelines. The AI system synthesized these inputs through reasoning, and observation-action loops to generate a comprehensive, context-aware, data-driven report with actional recommendations for follow-up.
FIGURE 7 AI-enabled framework for evaluating the translatability of preclinical AEs. The process integrates domain knowledge, including concordance logic and data schemas, mechanistic insights from on/off targets and Safetyome, and ICH safety guidelines. The AI system synthesized these inputs through reasoning, and observation-action loops to generate a comprehensive, context-aware, data-driven report with actional recommendations for follow-up.CC BY · PubMed Central

Verification

Claims the paper makes

30 Valid · 3 Not valid · 23 Not checked · 160 Own results and interpretations

103 more claims · 13 with their cited sources checked

Create a free account to see all 216 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 53
  • It may state two separate claims as one. 79

Create a free account to see these claims.

Citations

Cites

Show all 81 references

Research works printed but not linked: 1

  1. Shreffler, 2024, Diagnostic testing accuracy: sensitivity, specificity, predictive values and likelihood ratios

Cited by

See all 4

Metadata

Authors
Liu, X., & Fan, F.
Affiliations
Johnson & Johnson (United States) (US)
Corresponding author
Liu, Xin
Venue
Frontiers in Toxicology
Published
2026-04-02
Volume
8
Pages
1731947-1731947
ISSN
2673-3080 (ISSN-L)
OpenAlex
W7147433783
Cited by
4
References
81
Citations by year
2026: 4
FWCI
26.68
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Toxicology (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (submittedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Health Sciences › Veterinary › Small Animals › Animal testing and alternatives
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Toxicology › Pharmacovigilance and Adverse Drug Reactions
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
drug safety, small molecules, biologics, likelihood ratios, prevalence bias, 3Rs, human risk assessment

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: A large-scale concordance study of toxicity findings across preclinical species and humans for small molecules and biologics. https://app.cochat.ai/papers/10.3389/ftox.2026.1731947
The paper
Liu, X., & Fan, F. (2026). A large-scale concordance study of toxicity findings across preclinical species and humans for small molecules and biologics. Frontiers in Toxicology. https://doi.org/10.3389/ftox.2026.1731947

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.