single-cellfate-analysis_skill

Overview

This skill implements CellFateGenie: an end-to-end workflow to discover pseudotime-associated and lineage-driving genes using adaptive threshold ridge regression and manifold density scoring. It combines Adaptive Threshold Regression (ATR) for compact gene selection with Mellon-based low-density detection to prioritize fate-driving genes along trajectories. Use it after you have computed pseudotime for single-cell or bulk trajectory analyses.

How this skill works

First, a ridge regression model is fit to predict pseudotime from gene expression, producing gene coefficients and baseline fit metrics. ATR then iteratively removes low-coefficient genes while monitoring R² to find the smallest gene set that preserves fit. Optionally, Mellon density estimation on a diffusion manifold detects low-density transition regions, and a lineage_score routine computes lineage-specific expression variability to flag fate-driving genes.

When to use it

• You have pseudotime values computed (Palantir, VIA, DPT, etc.) and want genes correlated with progression.
• You need a compact set of genes that explains pseudotime with minimal loss of model fit.
• You want to find genes specifically driving a particular lineage or branch of a trajectory.
• Your dataset is single-cell or bulk RNA-seq where log-normalized expression is available.
• You plan to combine trajectory results with downstream visualization or gene-set analysis.

Best practices

• Ensure pseudotime exists in adata.obs and contains no NaNs before initializing Fate.
• Run HVG selection and PCA (X_pca in adata.obsm) beforehand for more stable results.
• Install mellon (pip install mellon) before running low_density(); otherwise skip density steps.
• Tune ATR 'flux' and 'stop' to balance feature parsimony and R² preservation (lower flux keeps more genes).
• Use data augmentation when pseudotime estimates are noisy (jitter or expression noise options).
• If GPU memory is limited, force CPU or accept automatic fallback to sklearn Ridge.

Example use cases

• Identify top pseudotime-associated genes driving erythroid differentiation from Palantir outputs.
• Filter a genome-wide model to a parsimonious gene set that maintains high R² with ATR.
• Locate low-density transition regions to time branch commitment events using Mellon.
• Score and rank genes that vary specifically in a user-defined lineage (cluster list).
• Apply the same pipeline to scATAC peak counts via atac_init and get_related_peak.