NACCT 2023

  1. Our CFG poster

Thank you for visiting our web page.

Overall Goal

  1. Understand the sublethal toxicity of 2,4-dinitrophenol (DNP) using social media.
  2. Develop tools to automatically update this knowledge.
  3. Apply these tools to other substances.

Scientific Output

  1. Sublethal Patterns of 2,4-Dinitrophenol Use as Discussed on Social Media (manuscript in press) (preprint)
  2. Our poster on how rhetorical stance influences content when posting about DNP.
  3. Our poster on dose-effect associations related online for DNP.


Toxicology Previous research described lethal toxicity, reflecting the selection bias of case reports and calls to Poison Control Centers (PCC). DNP use appears to have increased. Of the 25 reports of fatalities to US Poison Control Centers, 12 occurred after 2000 (Grundlingh et al., 2011). This suggests that nonfatal use is increasing. But we don’t know how people use DNP safely, other than our expectation that they limit the dose.

Computational Linguistics An innovation for computational linguistics would be to extract health information from social media posting. Previous research has described extracting health information as named entities (Bhatia et al., 2019) and extracting relationships between those entities (Foufi et al., 2019), sometimes by inferring ontologies (Alobaidi et al., 2018; Painter, 2010).

  1. MedSpaCy
  2. SciSpaCy

Why Use Social Media for DNP

The FDA prohibited DNP for human consumption in 1938 (Yen & Ewald, 2012). Even if the FDA approved DNP for human consumption, it would still be unethical to conduct human trials. We know that some doses of DNP are lethal and are not certain there is any safe dose. We can, ethically, use information that people who consumed DNP of their own accord freely provide. Social media provide just such a data stream.


  1. Aggregate reports of DNP usage from social media posts
  2. Extract mentions substances and their effects from those posts using named entity recognition. Substances means DNP and coingestants.
  3. Cross-reference those mentions with a knowledge base that represents what toxicologists know about the topic using entity linking
  4. (Future) Extract dosage from those posts.
  5. (Future) Generate hypotheses as to how the stated coingestants give rise to the stated effects or could otherwise be beneficial.


  1. 1.Grundlingh, J., Dargan, P. I., El-Zanfaly, M. & Wood, D. M. 2, 4-dinitrophenol (DNP): a weight loss agent with significant acute toxicity and risk of death. Journal of medical toxicology 7, 205 (2011).
  2. 2.Bhatia, P., Celikkaya, B., Khalilia, M. & Senthivel, S. Comprehend medical: a named entity recognition and relationship extraction web service. in 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA) 1844–1851 (IEEE, 2019).
  3. 3.Foufi, V. et al. Mining of textual health information from Reddit: Analysis of chronic diseases with extracted entities and their relations. Journal of medical Internet research 21, e12876 (2019).
  4. 4.Alobaidi, M., Malik, K. M. & Hussain, M. Automated ontology generation framework powered by linked biomedical ontologies for disease-drug domain. Computer methods and programs in biomedicine 165, 117–128 (2018).
  5. 5.Painter, J. L. Toward automating an inference model on unstructured terminologies: Oxmis case study. in Advances in Computational Biology 645–651 (Springer, 2010).
  6. 6.Yen, M. & Ewald, M. B. Toxicity of weight loss agents. Journal of medical toxicology 8, 145–152 (2012).