Conformal inference for reliable single cell RNA-seq annotation.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Supervised learning models for cell type assignment often lack rigorous uncertainty quantification.
• Existing methods for handling uncertain annotations rely on arbitrary assumptions and lack statistical guarantees.

Purpose of the Study:

• To propose a methodology using conformal prediction for reliable single-cell annotations.
• To address challenges in single-cell RNA sequencing (scRNA-seq) annotations, including detecting out-of-distribution cell types and quantifying annotation uncertainty.

Main Methods:

• Leveraging the conformal prediction framework for statistical guarantees on predictions.
• Developing an anomaly detector and an uncertainty-aware annotator for scRNA-seq data.
• Evaluating the methodology across various tissues, annotation taxonomies, and non-conformity measures.

Main Results:

• The anomaly detector successfully identified previously unseen cell types.
• The uncertainty-aware annotator produced well-calibrated prediction sets, maintaining coverage probabilities.
• Conformal prediction outputs enhanced downstream analyses.

Conclusions:

• Conformal prediction offers a robust framework for reliable single-cell annotation with statistical guarantees.
• The proposed method effectively handles uncertainty and detects novel cell types in scRNA-seq data.
• This approach improves the accuracy and trustworthiness of automated cell type identification.