Deep learning, deeper relief: pipeline toward tailored analgesia for experimental animal models

Luisa Barleben, M. Celeste Simon, Lukas Drees, Fabian B. Flohr, Christoph Jochum, Michela Di Virgilio, Frank Tacke, Sonja Bröer, Jana Wolf, Marina KolesnichenkoFrontiers in Immunology 2025 · DOI

TL;DR

This paper introduces a novel computational pipeline, PEDL+, that leverages deep learning to identify analgesics with minimal interference in experimental animal models, particularly for inflammatory conditions like colitis. By combining text mining with in vivo validation, they demonstrate that commonly used analgesics can significantly skew immune responses and epithelial cell differentiation, while their selected alternatives, like amantadine, offer a more precise approach to pain management in research.

What Problem Does This Paper Address?

Effective pain management in animal research is critical for ethical reasons and to ensure the scientific validity of experiments. However, many commonly used analgesics, such as opioids and NSAIDs, can interfere with immune system functions and signaling pathways, potentially confounding experimental results. This is particularly problematic in models of inflammatory diseases like Inflammatory Bowel Disease (IBD), where the immune response is central. The authors highlight that existing analgesics are often used without considering their specific off-target effects on the biological processes being studied, leading to a "one-size-fits-all" approach that can compromise reproducibility and accuracy. They aim to develop a method to identify analgesics that alleviate pain without disrupting the delicate biological systems under investigation.

Methods

The core of the paper is the development and application of a computational pipeline for analgesic selection. This pipeline involves several key steps:

  • Protein-Centered Association Extraction with Deep Language (PEDL+): This deep learning-based approach uses natural language processing to mine scientific literature (PubMed) for associations between drugs (analgesics) and biological targets, specifically focusing on NF-κB signaling. It quantifies the potential interference of analgesics by assigning a "score" based on the evidence found.
  • Database Cross-referencing (PubChem): For analgesics that showed no hits in PEDL+, the researchers cross-referenced them with PubChem to identify any known chemical-target interactions.
  • SwissTargetPrediction: This tool was used to predict potential targets for analgesics that passed the initial screening, based on structural similarity to known ligands.
  • In Vivo Validation: The most promising candidates identified computationally were then validated in an experimental setting using NF-κB reporter mice. This involved administering the analgesics to mice and assessing their impact on NF-κB activation, immune cell composition, and epithelial cell differentiation in the colon and ileum, both under basal conditions and during chemically induced colitis.

Datasets

  • PubMed: Used as the primary source for text mining to identify drug-gene interactions via the PEDL+ model.
  • PubChem: Utilized to cross-reference potential analgesic-target interactions for compounds that did not yield results in PEDL+.
  • SwissTargetPrediction: Employed to predict targets for analgesics based on structural similarity.
  • NF-κB Reporter Mouse Model (B6-Tg(k-EGFP)3Pt/Rsu): This model was used for in vivo validation, allowing for the visualization of NF-κB activity.
  • Mouse Colitis Model (DSS-induced): Used to simulate inflammatory bowel disease and test the efficacy and interference of analgesics in a disease context.

The paper does not introduce a new benchmark dataset but rather applies existing computational tools and biological models to address the problem.

Key Insights and Ideas

  • Common analgesics are not neutral: The study powerfully demonstrates that widely used analgesics like tramadol and acetaminophen significantly interfere with fundamental biological processes, including immune cell recruitment and NF-κB signaling, and even alter the differentiation of epithelial stem cells into goblet cells. This challenges the assumption that these drugs are simply "pain relievers" without broader biological consequences in research settings.
  • Deep learning can predict drug interference: The PEDL+ pipeline successfully predicted which analgesics were likely to interfere with NF-κB signaling, providing a valuable computational tool for pre-screening.
  • Tailored analgesia is crucial: The research advocates for a shift from a generic approach to analgesic use towards a "precision medicine" for animal models, where the choice of analgesic is tailored to the specific experimental context and biological system.
  • Beyond immune cells: The study highlights the often-overlooked impact of analgesics on epithelial cell differentiation, revealing a new dimension of potential interference.

Key Results

  • PEDL+ identified 13 analgesics with no predicted interference with NF-κB signaling.
  • Tramadol and acetaminophen showed significant interference with NF-κB signaling and immune cell responses in vivo. They also skewed epithelial stem cell differentiation towards goblet cells.
  • Amantadine and piritramide were identified as promising candidates with minimal interference with NF-κB signaling.
  • In vivo validation confirmed amantadine as the best-suited analgesic for DSS-induced colitis, showing no significant impact on colitis severity, NF-κB activation, or immune cell composition, although it did promote goblet cell accumulation.
  • Amantadine was predicted to impact IL6/JAK/STAT signaling and interferon gamma response pathways.

Limitations and Future Work

  • PEDL+ focuses on NF-κB: While powerful, the PEDL+ pipeline's current focus is on NF-κB signaling. The authors acknowledge that other signaling pathways could be affected by analgesics, and they identified some pathways impacted by amantadine (IL6/JAK/STAT, interferon gamma response). Future work could expand the scope of PEDL+ to include a broader range of biological targets.
  • De novo analgesic identification: The study focused on screening existing, approved analgesics. The authors suggest that future applications of deep learning could extend to identifying entirely new analgesic compounds.
  • Generalizability: While the pipeline is presented as a generalizable tool, its validation was primarily in a colitis model. Further studies are needed to confirm its applicability across a wider range of experimental models and disease contexts.
  • Goblet cell differentiation: Amantadine, while showing minimal NF-κB interference, did promote goblet cell accumulation. The implications of this specific effect on epithelial function and differentiation warrant further investigation.

Summary updated

Abstract

Effective pain management in animal models is crucial for maintaining ethical and scientific integrity. However, commonly used analgesics may affect immune responses and disturb signaling pathways, thereby potentially confounding the experimental outcomes. In mouse colitis models, opioids and non-steroidal anti-inflammatory drugs have been shown to interfere with the immune response and the activation of the central regulator of inflammation, the transcription factor nuclear factor kappa B (NF-κB). Here, we propose a tailored pipeline for the identification and the validation of analgesics with minimal off-target effects. This approach combines protein-centered relation extraction using deep language models and distant supervision via the Protein-Centered Association Extraction with Deep Language (PEDL<sup>+</sup>) together with an <i>in vivo</i> experimental validation with a NF-κB reporter mouse model that enables unambiguous visualization of direct NF-κB activity across different tissues. Our findings indicate that commonly used analgesics, such as tramadol and acetaminophen, not only interfere with immune cell recruitment and NF-κB activation but also skew the differentiation of epithelial stem cells into goblet cells, affecting epithelial functions even after short exposures. Conversely, the analgesics selected by our PEDL<sup>+</sup>-based workflow, such as piritramide, demonstrated no significant interference with NF-κB signaling. To validate our findings <i>in vivo</i>, we treated our NF-κB reporter mice with the analgesics selected by our computational pipeline. Amantadine demonstrated the least impact on the inflammatory responses and NF-κB activation. We then predicted and identified the signaling pathways that are impacted by amantadine treatment. In summary, our proposed pipeline facilitates a shift from one-size-fits-all analgesics to a precision medicine approach that considers the unique molecular interactions associated with each model.
BibTeX
@inproceedings{barleben2025deep,
  author = {Luisa Barleben and M. Celeste Simon and Lukas Drees and Fabian B. Flohr and Christoph Jochum and Michela Di Virgilio and Frank Tacke and Sonja Bröer and Jana Wolf and Marina Kolesnichenko},
  title = {{Deep learning, deeper relief: pipeline toward tailored analgesia for experimental animal models}},
  booktitle = {Frontiers in Immunology},
  year = {2025},
  doi = {10.3389/fimmu.2025.1639881},
}
Figure