A new research paper published on arXiv details biases present in protein-protein interaction (PPI) datasets that can mislead machine learning models. The study, led by Judith Bernett, analyzed several major PPI databases including HIPPIE, IntAct, and STRING, as well as data from the Protein Data Bank. Researchers found that common data splitting methods and the construction of negative datasets can introduce "shortcuts" that models exploit instead of learning genuine biological signals. To address this, the paper introduces an open-source Nextflow pipeline designed to mitigate these biases through optimized dataset splitting and negative sampling techniques. AI
IMPACT Identifies critical data quality issues that could lead to unreliable AI models in biological research, necessitating careful data auditing and preprocessing.
RANK_REASON Research paper published on arXiv detailing biases in biological datasets. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →