Generative AI-assisted Bayesian-frequentist Hybrid Inference in Single-cell RNA Sequencing Analysis for Genes...

Area of Science:

• Genomics
• Statistical Inference
• Artificial Intelligence

Background:

• High-dimensional omics studies, like Alzheimer's disease genomics, require advanced statistical methods.
• Bayesian inference offers advantages in small-sample settings but is limited by the difficulty of specifying priors for numerous parameters.
• Current methods struggle with the scale and complexity of modern omics data.

Purpose of the Study:

• To develop an AI-assisted Bayesian-frequentist hybrid inference framework for large-scale omics data analysis.
• To overcome the challenge of prior elicitation in high-dimensional Bayesian analyses.
• To improve statistical power and control error rates in genomic studies.

Main Methods:

• Coupling large language model (LLM) based prior elicitation with hybrid inference theory.
• Utilizing ChatGPT-4o to assess gene-disease evidence and map responses to informative normal priors.
• Treating secondary covariates as frequentist parameters to maintain efficiency and avoid prior sensitivity.
• Deriving closed-form hybrid estimators and establishing their asymptotic properties.

Main Results:

• Simulations demonstrate high statistical power and accurate Type I error rate control with the hybrid inference approach.
• The framework successfully identified previously undetected biologically coherent pathways (e.g., gamma-secretase) in Alzheimer's disease single-cell RNA sequencing data.
• Closed-form hybrid estimators were derived and shown to be asymptotically equivalent to frequentist and full Bayes estimators.

Conclusions:

• The proposed AI-assisted framework provides a principled and computationally scalable method for genome-wide Bayesian analysis.
• This approach effectively addresses the challenges of prior elicitation in high-dimensional settings.
• The framework has broad applicability across various omics platforms and disease research areas.