CarD-T: Interpreting Carcinomic Lexicon via Transformers.

Area of Science:

• Computational toxicology
• Cancer epidemiology
• Bioinformatics

Background:

• Accurate carcinogen identification is crucial for cancer epidemiology.
• Existing methods face challenges with the increasing volume of biomedical literature.
• Manual vetting of scientific texts is time-consuming and prone to disparities.

Purpose of the Study:

• To introduce the Carcinogen Detection via Transformers (CarD-T) framework for automated carcinogen nomination.
• To improve the efficiency and accuracy of identifying potential carcinogens from scientific texts.
• To provide a scalable solution for toxicological investigations.

Main Methods:

• Developed the Carcinogen Detection via Transformers (CarD-T) framework, integrating transformer-based machine learning and probabilistic statistical analysis.
• Utilized Named Entity Recognition (NER) trained on PubMed abstracts and a context classifier.
• Analyzed 25 years of journal publication data indexed with carcinogenicity and carcinogenesis MeSH terms.
• Employed Bayesian temporal Probabilistic Carcinogenic Denomination (PCarD) for analyzing disputing evidence.

Main Results:

• CarD-T accurately identified all established IARC Group 1 and 2A carcinogens from the test data.
• Nominated approximately 1500 additional potential carcinogens with supporting publications.
• Achieved high recall (0.857) and F1 score (0.875) compared to GPT-4.
• Highlighted 554 entities with disputing evidence for carcinogenicity, further analyzed by PCarD.

Conclusions:

• The CarD-T framework is a robust, efficient, and scalable tool for identifying potential carcinogens in biomedical literature.
• This AI-driven approach enhances the agility of public health responses to carcinogen identification.
• CarD-T sets a new benchmark for automated toxicological investigations, even on consumer GPUs.